TWI809488B - Evaluating method of microvascular invasion in hepatocellular carcinoma and evaluating system thereof - Google Patents
- ️Fri Jul 21 2023
Info
-
Publication number
- TWI809488B TWI809488B TW110132316A TW110132316A TWI809488B TW I809488 B TWI809488 B TW I809488B TW 110132316 A TW110132316 A TW 110132316A TW 110132316 A TW110132316 A TW 110132316A TW I809488 B TWI809488 B TW I809488B Authority
- TW
- Taiwan Prior art keywords
- interest
- region
- image
- hepatocellular carcinoma
- computed tomography Prior art date
- 2021-08-31
Links
- 230000009545 invasion Effects 0.000 title claims abstract description 80
- 206010073071 hepatocellular carcinoma Diseases 0.000 title claims abstract description 69
- 231100000844 hepatocellular carcinoma Toxicity 0.000 title claims abstract description 69
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000002591 computed tomography Methods 0.000 claims abstract description 86
- 206010028980 Neoplasm Diseases 0.000 claims abstract description 30
- 238000012549 training Methods 0.000 claims description 46
- 238000013527 convolutional neural network Methods 0.000 claims description 41
- 238000013135 deep learning Methods 0.000 claims description 30
- 238000011156 evaluation Methods 0.000 claims description 29
- 238000004364 calculation method Methods 0.000 claims description 24
- 239000000427 antigen Substances 0.000 claims description 18
- 102000036639 antigens Human genes 0.000 claims description 18
- 108091007433 antigens Proteins 0.000 claims description 18
- 238000003325 tomography Methods 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 16
- 230000008595 infiltration Effects 0.000 claims description 15
- 238000001764 infiltration Methods 0.000 claims description 15
- 208000005176 Hepatitis C Diseases 0.000 claims description 9
- 102000013529 alpha-Fetoproteins Human genes 0.000 claims description 9
- 108010026331 alpha-Fetoproteins Proteins 0.000 claims description 9
- 230000000694 effects Effects 0.000 claims description 9
- 208000002672 hepatitis B Diseases 0.000 claims description 9
- 210000001367 artery Anatomy 0.000 claims description 8
- 238000012937 correction Methods 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 7
- 238000009966 trimming Methods 0.000 claims description 5
- 238000001514 detection method Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims 1
- 238000002372 labelling Methods 0.000 abstract 2
- 230000000052 comparative effect Effects 0.000 description 27
- 238000010586 diagram Methods 0.000 description 23
- 238000012795 verification Methods 0.000 description 13
- 208000014018 liver neoplasm Diseases 0.000 description 6
- 239000002872 contrast media Substances 0.000 description 5
- 238000007405 data analysis Methods 0.000 description 5
- 230000035945 sensitivity Effects 0.000 description 5
- 206010019695 Hepatic neoplasm Diseases 0.000 description 3
- 201000007270 liver cancer Diseases 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000010200 validation analysis Methods 0.000 description 3
- 230000000903 blocking effect Effects 0.000 description 2
- 201000011510 cancer Diseases 0.000 description 2
- 230000003111 delayed effect Effects 0.000 description 2
- 238000003702 image correction Methods 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 230000002980 postoperative effect Effects 0.000 description 2
- 230000004083 survival effect Effects 0.000 description 2
- 210000001631 vena cava inferior Anatomy 0.000 description 2
- 239000008280 blood Substances 0.000 description 1
- 210000004369 blood Anatomy 0.000 description 1
- 210000004204 blood vessel Anatomy 0.000 description 1
- 210000000988 bone and bone Anatomy 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 230000005182 global health Effects 0.000 description 1
- 238000002347 injection Methods 0.000 description 1
- 239000007924 injection Substances 0.000 description 1
- 210000003734 kidney Anatomy 0.000 description 1
- 210000004185 liver Anatomy 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 210000004088 microvessel Anatomy 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000005855 radiation Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000002271 resection Methods 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 238000001356 surgical procedure Methods 0.000 description 1
- 230000009790 vascular invasion Effects 0.000 description 1
Images
Landscapes
- Investigating Or Analysing Biological Materials (AREA)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
An evaluating method of microvascular invasion in hepatocellular carcinoma and an evaluating system thereof are provided. The evaluating method of microvascular invasion in hepatocellular carcinoma includes steps as follows. A computed tomography image of a subject is provided, a region of interest labeling step is performed, an edge cutting step is performed, a predicting step is performed, and a determining and classifying step is performed. The computed tomography image includes a tumor area. The region of interest labeling step is to label a region of interest on the computed tomography image. The edge cutting step is to cut off the edge of the region of interest so as to obtain a cut region of interest. Each of a side length of the cut region of interest is 0.5 times to 1 times of each of a side length of the region of interest. Therefore, the goal of rapidly predicting and precisely determining of microvascular invasion according to computed tomography images from different sources can be achieved.
Description
本發明是有關於一種肝細胞癌微血管浸潤評估方法及其評估系統，特別是有關於一種以卷積神經網路模型進行預測與判斷，且可對不同拍攝來源的電腦斷層掃描圖像進行評估之肝細胞癌微血管浸潤評估方法及其評估系統。The present invention relates to a method for evaluating microvascular invasion of hepatocellular carcinoma and its evaluation system, in particular to a method for predicting and judging with a convolutional neural network model, and for evaluating computerized tomography images from different shooting sources Evaluation method and evaluation system of microvascular invasion in hepatocellular carcinoma.
肝癌是一個全球性的健康威脅的疾病，其發病率在全球範圍內不斷增加。其中肝細胞癌(Hepatocellular carcinoma, HCC)為最常見的原發性肝腫瘤，約佔所有肝癌病例的90%。Liver cancer is a global health threat and its incidence is increasing worldwide. Among them, hepatocellular carcinoma (HCC) is the most common primary liver tumor, accounting for about 90% of all liver cancer cases.
於眾多的肝癌治療方法中，其中肝切除手術仍是許多患者的首選治療方法，但是其五年內的癌症復發率可高達七成。血管浸潤是導致HCC早期復發的危險因素。文獻研究報告指出微血管浸潤(Microvascular invasion, MVI)與患者術後的復發率和存活率有重要的相關性，因此，對於正確的預測肝細胞癌是否存在微血管浸潤可以作為病人術後的癌症復發率和長期的存活率的判斷依據。Among the many treatment methods for liver cancer, liver resection is still the first choice for many patients, but the cancer recurrence rate within five years can be as high as 70%. Vascular invasion is a risk factor leading to early recurrence of HCC. Literature research reports point out that microvascular invasion (MVI) has an important correlation with postoperative recurrence rate and survival rate of patients. Therefore, for the correct prediction of whether there is microvascular invasion in hepatocellular carcinoma, it can be used as the postoperative cancer recurrence rate of patients. and long-term survival rates.
然而，微血管浸潤一般在進行手術前無法得到明確診斷，故開發一套具有高判斷準確度的肝細胞癌微血管浸潤評估方法及其評估系統，對於輔助醫生對病人治療方式的決策上具有相當的重要性。However, microvascular invasion cannot be clearly diagnosed before surgery. Therefore, it is very important to develop a high-accuracy evaluation method and evaluation system for microvascular invasion in hepatocellular carcinoma to assist doctors in making decisions about patient treatment. sex.
本發明之一目的在於提供一種肝細胞癌微血管浸潤評估方法及其評估系統，藉由使用動脈相電腦斷層掃描圖像進行感興趣區域的圈選，並對感興趣區域進行剪裁，得到各邊長為感興趣區域之各邊長的0.5倍至1倍的剪裁後感興趣區域後，再以卷積神經網路模型對剪裁後感興趣區域進行數據預測並進行微血管浸潤的判斷，以達到快速且可對不同拍攝來源的電腦斷層掃描圖像進行評估的效果。One object of the present invention is to provide a method for assessing microvascular invasion of hepatocellular carcinoma and its assessment system, by using arterial phase computed tomography images to circle the region of interest, and clipping the region of interest to obtain the length of each side After clipping the region of interest that is 0.5 to 1 times the length of each side of the region of interest, the convolutional neural network model is used to predict the data of the clipped region of interest and determine the microvascular infiltration to achieve fast and The effect can be evaluated on computed tomography images from different sources.
本發明之一態樣提供一種肝細胞癌微血管浸潤評估方法，包含提供一受試者之一電腦斷層掃描圖像、進行一感興趣區域圈選步驟、進行一裁邊步驟、進行一數據預測步驟以及進行一判斷分類步驟。電腦斷層掃描圖像包含一腫瘤區域。感興趣區域圈選步驟係於電腦斷層掃描圖像圈選一感興趣區域，其中感興趣區域包含腫瘤區域。裁邊步驟係將感興趣區域之邊緣進行剪裁，以得一剪裁後感興趣區域，其中剪裁後感興趣區域之各邊長為感興趣區域之各邊長的0.5倍至1倍。數據預測步驟係利用一卷積神經網路模型對剪裁後感興趣區域進行數據預測，以得一影像特徵值。判斷分類步驟係利用卷積神經網路模型對影像特徵值進行判斷，若影像特徵值大於或等於一閾值，受試者將被判定為發生微血管浸潤，若影像特徵值小於閾值，受試者將被判定為未發生微血管浸潤。An aspect of the present invention provides a method for assessing microvascular invasion of hepatocellular carcinoma, comprising providing a computed tomography image of a subject, performing a region of interest circle selection step, performing a trimming step, and performing a data prediction step And a step of judging and classifying is performed. Computed tomography image contains an area of tumor. The region-of-interest selection step is to select a region-of-interest in the computed tomography image, wherein the region-of-interest includes a tumor region. The step of trimming the edges is to trim the edges of the ROI to obtain a trimmed ROI, wherein the length of each side of the trimmed ROI is 0.5 to 1 times the length of each side of the ROI. The data prediction step uses a convolutional neural network model to perform data prediction on the clipped region of interest to obtain an image feature value. The judgment and classification step is to use the convolutional neural network model to judge the feature value of the image. If the feature value of the image is greater than or equal to a threshold, the subject will be judged as having microvascular infiltration. If the feature value of the image is less than the threshold, the subject will be judged as It was judged that microvascular invasion did not occur.
依據前述之肝細胞癌微血管浸潤評估方法，更佳地，剪裁後感興趣區域之各邊長可為感興趣區域之各邊長的0.8倍至0.9倍。According to the aforementioned method for assessing microvascular invasion of hepatocellular carcinoma, preferably, the length of each side of the trimmed region of interest may be 0.8 to 0.9 times the length of each side of the region of interest.
依據前述之肝細胞癌微血管浸潤評估方法，其中電腦斷層掃描圖像可為一動脈相電腦斷層掃描圖像。According to the aforementioned method for evaluating microvascular invasion of hepatocellular carcinoma, the computed tomography image may be an arterial phase computed tomography image.
依據前述之肝細胞癌微血管浸潤評估方法，其中動脈相電腦斷層掃描圖像可於動脈之阻射率為120亨氏單位時拍攝而得。According to the aforementioned evaluation method for microvascular invasion of hepatocellular carcinoma, the arterial phase computed tomography image can be obtained when the arterial blocking rate is 120 Hounsfield units.
依據前述之肝細胞癌微血管浸潤評估方法，其中判斷分類步驟可更包含以一受試者臨床資料進行判斷，以提高判斷準確度，所述受試者臨床資料包含受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果。According to the aforementioned method for assessing microvascular invasion of hepatocellular carcinoma, the step of judging and classifying may further include judging based on a subject's clinical data to improve the accuracy of judgment. The said subject's clinical data includes the subject's age, gender, Maximum tumor diameter, α-fetoprotein value, Child-Pugh score, hepatitis B surface antigen test results and hepatitis C surface antigen test results.
依據前述之肝細胞癌微血管浸潤評估方法，可更包含進行一建模步驟，所述建模步驟包含提供一參照電腦斷層掃描圖像資料庫、進行一參照感興趣區域圈選步驟、進行一訓練前影像處理步驟以及進行一訓練步驟。參照電腦斷層掃描圖像資料庫包含複數個參照電腦斷層掃描圖像，各參照電腦斷層掃描圖像包含一參照腫瘤區域。參照感興趣區域圈選步驟係於各參照電腦斷層掃描圖像圈選一參照感興趣區域，以得複數個參照感興趣區域，其中各參照感興趣區域包含各參照腫瘤區域。訓練前影像處理步驟係對各參照感興趣區域進行剪裁，以得複數個剪裁後參照感興趣區域，其中各剪裁後參照感興趣區域之各邊長為各參照感興趣區域之各邊長的0.5倍至1倍。訓練步驟係以剪裁後參照感興趣區域對一深度學習演算模組進行訓練至收斂，以得卷積神經網路模型。According to the aforementioned method for assessing microvascular invasion of hepatocellular carcinoma, it may further include a modeling step, the modeling step includes providing a reference computed tomography image database, performing a reference region of interest circle selection step, and performing a training A pre-image processing step and a training step are performed. The reference computed tomography image database includes a plurality of reference computed tomography images, and each reference computed tomography image includes a reference tumor area. The reference ROI selection step is to circle a reference ROI in each reference computed tomography image to obtain a plurality of reference ROIs, wherein each reference ROI includes each reference tumor area. The pre-training image processing step is to clip each reference region of interest to obtain a plurality of clipped reference regions of interest, wherein the length of each side of each clipped reference region of interest is 0.5 of the length of each side of each reference region of interest times to 1 times. The training step is to train a deep learning algorithm module until convergence by referring to the region of interest after clipping, so as to obtain a convolutional neural network model.
依據前述之肝細胞癌微血管浸潤評估方法，更佳地，各剪裁後參照感興趣區域之各邊長可為各參照感興趣區域之各邊長的0.8倍至0.9倍。According to the aforementioned method for assessing microvascular invasion of hepatocellular carcinoma, preferably, the length of each side of each trimmed reference region of interest may be 0.8 to 0.9 times the length of each side of each reference region of interest.
依據前述之肝細胞癌微血管浸潤評估方法，其中各參照電腦斷層掃描圖像可為一參照動脈相電腦斷層掃描圖像。According to the aforementioned method for assessing microvascular invasion of hepatocellular carcinoma, each reference computed tomography image may be a reference arterial phase computed tomography image.
依據前述之肝細胞癌微血管浸潤評估方法，其中各參照動脈相電腦斷層掃描圖像可於動脈之阻射率為120亨氏單位時拍攝而得。According to the aforementioned evaluation method for microvascular invasion of hepatocellular carcinoma, the computed tomography images of each reference arterial phase can be obtained when the radioblocking rate of the artery is 120 Hounsfield units.
依據前述之肝細胞癌微血管浸潤評估方法，其中訓練前影像處理步驟可更包含對各剪裁後參照感興趣區域進行旋轉、裁剪或水平翻轉。According to the aforementioned method for assessing microvascular invasion of hepatocellular carcinoma, the pre-training image processing step may further include rotating, cropping or horizontally flipping each trimmed reference region of interest.
依據前述之肝細胞癌微血管浸潤評估方法，其中訓練步驟可更包含以複數個參照受試者臨床資料進行訓練，以提高訓練效果，所述參照受試者臨床資料包含各參照受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果。According to the aforementioned method for assessing microvascular invasion of hepatocellular carcinoma, the training step may further include training with the clinical data of a plurality of reference subjects to improve the training effect, and the clinical data of the reference subjects include the age of each reference subject , gender, largest tumor diameter, α-fetoprotein value, Child-Pugh score, hepatitis B surface antigen test results and hepatitis C surface antigen test results.
依據前述之肝細胞癌微血管浸潤評估方法，其中深度學習演算模組可為一ResNet-18演算模組。According to the aforementioned evaluation method for microvascular invasion of hepatocellular carcinoma, the deep learning algorithm module can be a ResNet-18 algorithm module.
本發明之另一態樣提供一種肝細胞癌微血管浸潤評估系統，包含一電腦斷層掃描圖像擷取裝置以及一處理器。電腦斷層掃描圖像擷取裝置係用以擷取一電腦斷層掃描圖像。處理器訊號連接電腦斷層掃描圖像擷取裝置，其中處理器包含一影像校正程式及一卷積神經網路模型。影像校正程式包含一圖形校正模組，圖形校正模組係將電腦斷層掃描圖像之一感興趣區域進行圖形轉換、剪裁、旋轉或水平翻轉，以得一剪裁後感興趣區域，其中剪裁後感興趣區域之各邊長為感興趣區域之各邊長的0.5倍至1倍。卷積神經網路模型包含一影像特徵值計算模組及一判斷分類模組。影像特徵值計算模組係利用剪裁後感興趣區域進行微血管浸潤評估，以得一影像特徵值。判斷分類模組係以影像特徵值相對於一閾值之大小關係，輸出一微血管浸潤評估結果，若影像特徵值大於或等於閾值，受試者將被判定為發生微血管浸潤，若影像特徵值小於閾值，受試者將被判定為未發生微血管浸潤。Another aspect of the present invention provides a system for assessing microvascular invasion of hepatocellular carcinoma, comprising a computed tomography image capture device and a processor. The computer tomography image capturing device is used for capturing a computer tomography image. The signal of the processor is connected to the computer tomography image acquisition device, wherein the processor includes an image correction program and a convolutional neural network model. The image correction program includes a graphic correction module. The graphic correction module converts, crops, rotates or horizontally flips a region of interest in a computed tomography image to obtain a region of interest after clipping. The length of each side of the ROI is 0.5 to 1 time the length of each side of the ROI. The convolutional neural network model includes an image feature value calculation module and a judgment classification module. The image feature value calculation module utilizes the clipped region of interest to evaluate microvascular infiltration to obtain an image feature value. The judgment and classification module outputs a microvascular infiltration evaluation result based on the relationship between the image feature value and a threshold value. If the image feature value is greater than or equal to the threshold value, the subject will be judged to have microvascular infiltration. If the image feature value is less than the threshold value , subjects will be judged as having no microvascular invasion.
依據前述之肝細胞癌微血管浸潤評估系統，其中卷積神經網路模型可由一深度學習演算模組訓練至收斂而得。According to the aforementioned evaluation system for microvascular invasion of hepatocellular carcinoma, the convolutional neural network model can be obtained by training a deep learning algorithm module until convergence.
依據前述之肝細胞癌微血管浸潤評估系統，其中深度學習演算模組可為一ResNet-18演算模組。According to the aforementioned evaluation system for microvascular invasion of hepatocellular carcinoma, the deep learning algorithm module can be a ResNet-18 algorithm module.
依據前述之肝細胞癌微血管浸潤評估系統，其中卷積神經網路模型可更包含一資料庫，所述資料庫包含一受試者臨床資料，卷積神經網路模型之判斷分類模組搭配受試者臨床資料作為參數，與影像特徵值一起進行微血管浸潤評估，其中受試者臨床資料包含受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果。According to the aforementioned evaluation system for microvascular invasion of hepatocellular carcinoma, the convolutional neural network model may further include a database, the database includes a subject's clinical data, and the judgment and classification module of the convolutional neural network model is matched with the subject The clinical data of the subjects were used as parameters to evaluate the microvascular invasion together with the image characteristic values. The clinical data of the subjects included the subject's age, gender, largest tumor diameter, α-fetoprotein value, Child-Pugh score, hepatitis B Surface antigen test results and hepatitis C surface antigen test results.
藉此，本發明之肝細胞癌微血管浸潤評估方法及其評估系統可達到快速且可對不同拍攝來源的電腦斷層掃描圖像進行評估的效果。Thereby, the method for evaluating the microvascular invasion of hepatocellular carcinoma and the evaluation system thereof of the present invention can achieve the effect of being fast and capable of evaluating computerized tomography images from different shooting sources.
以下將參照圖式說明本發明之複數個實施例。為明確說明起見，許多實務上的細節將在以下敘述中一併說明。然而，應瞭解到，這些實務上的細節不應用以限制本發明。也就是說，在本發明部分實施例中，這些實務上的細節是非必要的。此外，為簡化圖式起見，一些習知慣用的結構與元件在圖式中將以簡單示意的方式繪示之，並且重複之元件將可能使用相同的編號表示之。Several embodiments of the present invention will be described below with reference to the drawings. For the sake of clarity, many practical details are included in the following narrative. It should be understood, however, that these practical details should not be used to limit the invention. That is, in some embodiments of the present invention, these practical details are unnecessary. In addition, for the sake of simplifying the drawings, some well-known and commonly used structures and elements will be shown in a simple and schematic manner in the drawings, and repeated elements may be denoted by the same reference numerals.
請參照第1圖，第1圖係繪示本發明一態樣之一實施方式的肝細胞癌微血管浸潤評估方法100之步驟流程圖。肝細胞癌微血管浸潤評估方法100包含步驟110、步驟120、步驟130、步驟140及步驟150。Please refer to FIG. 1 . FIG. 1 is a flowchart showing the steps of a method 100 for evaluating microvascular invasion of hepatocellular carcinoma according to an embodiment of the present invention. The method 100 for assessing microvascular invasion of hepatocellular carcinoma includes step 110 , step 120 , step 130 , step 140 and step 150 .
步驟110為提供一受試者之一電腦斷層掃描(Computed tomography, CT)圖像，其中所述電腦斷層掃描圖像包含一腫瘤區域。詳細來說，所述電腦斷層掃描圖像係以一拍攝設備對肝臟腫瘤區域進行拍攝而得。更仔細地說，所述電腦斷層掃描圖像可為一動脈相電腦斷層掃描圖像，其係於動脈相(Arterial phase, AP)時所拍攝的電腦斷層掃描圖像，前述動脈相係指受試者於注射顯影劑後，顯影劑隨血液流動至動脈內並顯影的時期。Step 110 is to provide a computed tomography (CT) image of a subject, wherein the computed tomography image includes a tumor region. Specifically, the computed tomography image is obtained by photographing the liver tumor area with a photographing device. More specifically, the computerized tomography image can be an arterial phase computer tomography image, which is a computerized tomography image taken in the arterial phase (Arterial phase, AP), the aforementioned arterial phase refers to the affected After the subject is injected with the contrast agent, the contrast agent flows into the artery with the blood and develops the period.
實際來說，顯影劑經注射後通常會經動脈相、門靜脈相(Portal venous phase, PVP)及延遲相(Delay phase, DP)等時期，然各醫療機構對於門靜脈相及延遲相的定義不盡相同，故於本發明之實施方式中，所使用的電腦斷層掃描圖像均可為於動脈相時掃描而得的電腦斷層掃描圖像，且於前述之動脈相時，顯影劑於動脈進行顯影使動脈之阻射率可達120亨氏單位。故本發明之肝細胞癌微血管浸潤評估方法100在評估來自不同醫療機構的電腦斷層掃描圖像時，依然可具有良好的評估效能。In practice, the contrast agent usually passes through the arterial phase, portal venous phase (PVP) and delayed phase (Delay phase, DP) after injection. However, the definitions of portal venous phase and delayed phase vary in different medical institutions Similarly, in the embodiments of the present invention, the computerized tomography images used can all be computed tomography images scanned during the arterial phase, and during the aforementioned arterial phase, the contrast agent develops in the artery The radiation blocking rate of the artery can reach 120 Heinz units. Therefore, the method 100 for evaluating microvascular invasion of hepatocellular carcinoma of the present invention can still have good evaluation performance when evaluating CT images from different medical institutions.
步驟120為進行一感興趣區域圈選步驟，其係於電腦斷層掃描圖像圈選一感興趣區域，其中所述感興趣區域包含腫瘤區域。詳細來說，感興趣區域係由放射科醫師進行圈選，腫瘤區域位於所述感興趣區域之正中心，且感興趣區域之範圍會略大於腫瘤區域的邊界，以保證腫瘤區域完全包含於所圈選之感興趣區域中。進一步來說，放射科醫師對於感興趣區域之圈選範圍大小是基於電腦斷層掃描圖像的像素(pixel)所代表的物理距離決定，在本發明之實施例中，1像素所對應的物理距離約為0.8 mm。因於進行感興趣區域之圈選時，會先將電腦斷層掃描圖像進行放大以利圈選，而不同電腦斷層掃描圖像中的腫瘤大小有所不同，故每張電腦斷層掃描圖像的放大倍率及像素所代表的物理距離亦有所差異，因此實際上放射科醫師對於感興趣區域之圈選範圍大小係依據電腦斷層掃描圖像放大後，實際距離1 cm所對應的像素數(約為12-13個像素)進行圈選。於本發明之實施例中，感興趣區域之範圍約大於腫瘤區域的邊界1 cm，但本發明並不以此為限。完成初步圈選後的感興趣區域為圓形，後經圖形轉換為矩形，以供後續判斷使用。Step 120 is to perform a region of interest selection step, which is to select a region of interest in the computed tomography image, wherein the region of interest includes a tumor region. Specifically, the region of interest is circled by a radiologist, the tumor region is located in the center of the region of interest, and the range of the region of interest will be slightly larger than the border of the tumor region to ensure that the tumor region is completely contained within the region of interest. In the circled region of interest. Further, the radiologist's circle size of the region of interest is determined based on the physical distance represented by the pixel (pixel) of the computed tomography image. In the embodiment of the present invention, the physical distance corresponding to 1 pixel About 0.8mm. When performing circle selection of the region of interest, the computerized tomography image will be enlarged first to facilitate the circle selection, and the tumor size in different computer tomography images is different, so the size of each computer tomography image is different. The magnification and the physical distance represented by the pixels are also different. Therefore, in fact, the size of the radiologist’s circle selection for the region of interest is based on the number of pixels corresponding to the actual distance of 1 cm after the computerized tomography image is enlarged (approximately for 12-13 pixels) to circle. In the embodiment of the present invention, the range of the region of interest is about 1 cm larger than the boundary of the tumor region, but the present invention is not limited thereto. The region of interest after preliminary circle selection is circular, and then converted into a rectangle by graphics for subsequent judgment.
步驟130為進行一裁邊步驟，其係將感興趣區域之邊緣進行剪裁，以得一剪裁後感興趣區域，其中剪裁後感興趣區域之各邊長為感興趣區域之各邊長的0.5倍至1倍。更佳地，剪裁後感興趣區域之各邊長可為感興趣區域之各邊長的0.8倍至0.9倍。請一併參照第2圖，第2圖係繪示感興趣區域121與剪裁後感興趣區域131之示意圖。如第2圖所示，由於感興趣區域121之範圍約大於腫瘤區域的邊界1 cm，故感興趣區域121可能會包含會影響判斷結果的不必要資訊，如空氣、骨骼組織、腎臟、粗大血管(Great vessels)或下腔靜脈(Inferior vena cava, IVC)等，故對感興趣區域之邊緣進行剪裁可有效減少上述不必要資訊對判斷結果的影響，並提高判斷效果。於第2圖中，感興趣區域121為正方形且其邊長為d，經剪裁後得到之剪裁後感興趣區域131，且剪裁後感興趣區域131之邊長為0.8d，但本發明並不以此為限。Step 130 is to perform an edge trimming step, which is to trim the edges of the region of interest to obtain a trimmed region of interest, wherein the length of each side of the region of interest after trimming is 0.5 times the length of each side of the region of interest to 1x. More preferably, the length of each side of the region of interest after clipping may be 0.8 to 0.9 times the length of each side of the region of interest. Please also refer to FIG. 2 . FIG. 2 is a schematic diagram showing the ROI 121 and the cropped ROI 131 . As shown in FIG. 2, since the region of interest 121 is about 1 cm larger than the boundary of the tumor region, the region of interest 121 may contain unnecessary information that will affect the judgment result, such as air, bone tissue, kidney, and large blood vessels. (Great vessels) or inferior vena cava (IVC), etc. Therefore, clipping the edge of the region of interest can effectively reduce the influence of the above unnecessary information on the judgment result and improve the judgment effect. In Fig. 2, the region of interest 121 is a square and its side length is d, and the region of interest 131 after clipping is obtained after clipping, and the side length of the region of interest 131 after clipping is 0.8d, but the present invention does not This is the limit.
步驟140為進行一數據預測步驟，其係利用一卷積神經網路模型對剪裁後感興趣區域進行數據預測，以得一影像特徵值。所述卷積神經網路模型的訓練細節會於後段說明，在此不另贅述。Step 140 is to perform a data prediction step, which uses a convolutional neural network model to perform data prediction on the cropped ROI to obtain an image feature value. The details of the training of the convolutional neural network model will be described later, and will not be repeated here.
步驟150為進行一判斷分類步驟，其係利用卷積神經網路模型對影像特徵值進行判斷，若影像特徵值大於或等於閾值，受試者將被判定為發生微血管浸潤，若影像特徵值小於閾值，受試者將被判定為未發生微血管浸潤。其中每當卷積神經網路模型進行訓練後，所述閾值將重新選取而改變。詳細來說，所述閾值可以透過繪製接收者操作特徵曲線(Receiver operating characteristic curve, ROC曲線)上的座標點，得到一系列靈敏度(Sensitivity)的值與特異度(Specificity)的值，並進行約登指數(Youden’s index)的計算，其中計算得到的約登指數之最大值為最佳閾值。約登指數係表示本發明之肝細胞癌微血管浸潤評估方法及其評估系統發現真陽性與偽陽性的能力，靈敏度是指本發明之肝細胞癌微血管浸潤評估方法及其評估系統能將實際有微血管浸潤的人正確地判定為有微血管浸潤的比例，特異度是指本發明之肝細胞癌微血管浸潤評估方法及其評估系統能將實際無微血管浸潤的人正確地判定為無微血管浸潤的比例，其計算方式如下式(I)所示： 約登指數=靈敏度+特異度-1……………………… 式(I)。 藉此以得到約登指數之最大值作為閾值，其他閾值選取方式如以固定一靈敏度的原則或以固定一特異度的原則等均可應用於本發明之肝細胞癌微血管浸潤評估方法及其評估系統中，可依據實際狀況來進行確定與調整，本發明並不以此為限。 Step 150 is to perform a judgment and classification step, which uses the convolutional neural network model to judge the image feature value. If the image feature value is greater than or equal to the threshold value, the subject will be judged to have microvascular infiltration. If the image feature value is less than Threshold, subjects will be judged as no microvascular invasion. Wherein each time after the convolutional neural network model is trained, the threshold will be reselected and changed. In detail, the threshold can be obtained by plotting the coordinate points on the Receiver operating characteristic curve (ROC curve) to obtain a series of sensitivity (Sensitivity) values and specificity (Specificity) values, and make an approximation Calculation of Youden's index, where the maximum value of the calculated Youden's index is the optimal threshold. Youden index represents the ability of the method for evaluating microvascular invasion of hepatocellular carcinoma of the present invention and its evaluation system to find true positives and false positives, and the sensitivity refers to the ability of the method for evaluating microvascular invasion of hepatocellular carcinoma of the present invention and its evaluation system to detect the actual microvascular invasion. The proportion of infiltrated people who are correctly judged to have microvascular invasion, and the specificity refers to the proportion of people who actually have no microvascular infiltration can be correctly judged as having no microvascular infiltration by the method for assessing hepatocellular carcinoma microvascular invasion and its evaluation system of the present invention. The calculation method is shown in the following formula (I): Youden Index = Sensitivity + Specificity - 1…………………… Formula (I). In this way, the maximum value of the Youden index can be obtained as the threshold, and other threshold selection methods such as the principle of fixed-sensitivity or the principle of fixed-specificity can be applied to the method for evaluating microvascular invasion of hepatocellular carcinoma of the present invention and its evaluation In the system, it can be determined and adjusted according to the actual situation, and the present invention is not limited thereto.
再者，所述判斷分類步驟更可包含以受試者臨床資料(Clinical factors, CF)進行判斷，以提高判斷準確度。受試者臨床資料包含受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果，但本發明並不以此為限。Furthermore, the step of judging and categorizing may further include judging based on the clinical factors (CF) of the subjects, so as to improve the accuracy of judging. The clinical data of the subject includes the subject's age, gender, largest tumor diameter, α-fetoprotein value, Child-Pugh score, hepatitis B surface antigen test results and hepatitis C surface antigen test results, but the present invention does not This is the limit.
請同時參照第1圖及第3圖，第3圖係繪示本發明一態樣之另一實施方式的肝細胞癌微血管浸潤評估方法100a之步驟流程圖。肝細胞癌微血管浸潤評估方法100a的步驟110a、步驟120a、步驟130a、步驟140a及步驟150a與第1圖之肝細胞癌微血管浸潤評估方法100的步驟110、步驟120、步驟130、步驟140及步驟150相同，在此不另贅述。以下將配合第1圖與第3圖說明本發明之卷積神經網路模型的建構細節。步驟160為進行建模步驟，其包含步驟161、步驟162、步驟163及步驟164。Please refer to FIG. 1 and FIG. 3 at the same time. FIG. 3 is a flow chart showing the steps of a method 100 a for evaluating microvascular invasion of hepatocellular carcinoma according to another embodiment of an aspect of the present invention. Step 110a, step 120a, step 130a, step 140a, and step 150a of the method 100a for evaluating microvascular invasion of hepatocellular carcinoma and step 110, step 120, step 130, step 140, and steps of the method 100 for evaluating microvascular invasion of hepatocellular carcinoma in FIG. 1 150 are the same, and will not be repeated here. The details of the construction of the convolutional neural network model of the present invention will be described below with reference to FIG. 1 and FIG. 3 . Step 160 is a modeling step, which includes step 161 , step 162 , step 163 and step 164 .
步驟161為提供一參照電腦斷層掃描圖像資料庫，其中參照電腦斷層掃描圖像資料庫包含複數個參照電腦斷層掃描圖像，各參照電腦斷層掃描圖像包含一參照腫瘤區域。詳細來說，所述參照電腦斷層掃描圖像係以一拍攝設備對肝臟腫瘤區域進行拍攝而得。更仔細地說，所述參照電腦斷層掃描圖像可為一參照動脈相電腦斷層掃描圖像，其係於動脈相(Arterial phase, AP)時所拍攝的參照電腦斷層掃描圖像，且於前述之動脈相時，顯影劑於動脈進行顯影使動脈之阻射率可達120亨氏單位。Step 161 is to provide a reference CT scan image database, wherein the reference CT scan image database includes a plurality of reference CT scan images, and each reference CT scan image includes a reference tumor region. In detail, the reference computerized tomography image is obtained by photographing the liver tumor area with a photographing device. More specifically, the reference computed tomography image may be a reference computed tomography image of the arterial phase, which is a reference computed tomography image taken in the arterial phase (Arterial phase, AP), and in the aforementioned In the arterial phase, the contrast agent is developed in the artery so that the radioblocking rate of the artery can reach 120 Hounsfield units.
步驟162為進行一參照感興趣區域圈選步驟，其係於各參照電腦斷層掃描圖像圈選一參照感興趣區域，以得複數個參照感興趣區域，其中各參照感興趣區域包含各參照腫瘤區域。於步驟162中，圈選參照感興趣區域的方式與第1圖之肝細胞癌微血管浸潤評估方法100的步驟120中圈選感興趣區域之方式相同，在此不另贅述。Step 162 is to perform a reference region of interest circle selection step, which is to circle a reference region of interest in each reference computed tomography image to obtain a plurality of reference regions of interest, wherein each reference region of interest includes each reference tumor area. In step 162 , the method of framing the reference region of interest is the same as the method of circling the region of interest in step 120 of the method 100 for assessing hepatocellular carcinoma microvascular invasion in FIG. 1 , and will not be repeated here.
步驟163為進行一訓練前影像處理步驟，其係對各參照感興趣區域進行剪裁，以得複數個剪裁後參照感興趣區域，其中各剪裁後參照感興趣區域之各邊長可為各參照感興趣區域之各邊長的0.5倍至1倍。更佳地，各剪裁後參照感興趣區域之各邊長可為各參照感興趣區域之各邊長的0.8倍至0.9倍。Step 163 is to perform a pre-training image processing step, which is to clip each reference ROI to obtain a plurality of clipped reference ROIs, wherein the lengths of each side of each clipped reference ROI can be the reference sense 0.5 times to 1 times the length of each side of the region of interest. More preferably, each side length of each clipped reference ROI may be 0.8 to 0.9 times that of each side length of each reference ROI.
請額外參照第4A圖、第4B圖、第4C圖、第4D圖、第4E圖、第4F圖、第5A圖、第5B圖、第5C圖及第5D圖，第4A圖為一參照電腦斷層掃描圖像之剪裁後參照感興趣區域之影像。第4B圖為將第4A圖中的剪裁後參照感興趣區域旋轉-10 o之影像。第4C圖為將第4A圖中的剪裁後參照感興趣區域旋轉-5 o之影像。第4D圖為將第4A圖中的剪裁後參照感興趣區域旋轉5 o之影像。第4E圖為將第4A圖中的剪裁後參照感興趣區域旋轉10 o之影像。第4F圖為將第4A圖中的剪裁後參照感興趣區域水平翻轉之影像。第5A圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之示意圖。第5B圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之另一示意圖。第5C圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之再一示意圖。第5D圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之又一示意圖。 Please also refer to Figure 4A, Figure 4B, Figure 4C, Figure 4D, Figure 4E, Figure 4F, Figure 5A, Figure 5B, Figure 5C and Figure 5D. Figure 4A is a reference computer The cropped tomographic image is referenced to the image of the region of interest. Figure 4B is the cropped image in Figure 4A rotated by -10o with reference to the region of interest. Figure 4C is the cropped image in Figure 4A rotated by -5o with reference to the region of interest. Figure 4D is the cropped image in Figure 4A rotated 5o with reference to the region of interest. Figure 4E is the cropped image in Figure 4A rotated 10o with reference to the region of interest. Fig. 4F is an image horizontally flipped with reference to the region of interest after cropping in Fig. 4A. FIG. 5A is a schematic diagram of clipping in FIG. 4A with reference to the region of interest. FIG. 5B is another schematic diagram of clipping in FIG. 4A with reference to the region of interest. FIG. 5C is another schematic diagram of clipping in FIG. 4A with reference to the region of interest. FIG. 5D is another schematic diagram of clipping in FIG. 4A with reference to the region of interest.
詳細來說，訓練前影像處理步驟可更包含對各剪裁後參照感興趣區域進行旋轉、裁剪或水平翻轉，並將經旋轉、裁剪或水平翻轉的各剪裁後參照感興趣區域分別建檔，藉此以增加訓練之樣本數量。於此實施方式中，對各剪裁後參照感興趣區域進行旋轉的角度為-10 o至10 o。另一方面，對各剪裁後參照感興趣區域進行裁剪之尺寸比例可為原圖的0.8倍至1倍，且長寬比可為0.95至1.05，如第5A圖、第5B圖、第5C圖及第5D圖中黑框處所示，其尺寸比例及長寬比數值列於下表一，然本發明並不以此為限。 表一 第5A圖 第5B圖 第5C圖 第5D圖 尺寸比例 0.96 0.95 0.91 0.91 長寬比 1.04 1.02 0.94 0.98 In detail, the pre-training image processing step may further include rotating, cropping or horizontally flipping each trimmed reference ROI, and building files for each trimmed reference ROI after rotation, cropping or horizontal flipping, thereby This increases the number of training samples. In this embodiment, the angle of rotation for each clipped reference ROI is -10 ° to 10 ° . On the other hand, the size ratio of cropping with reference to the region of interest after clipping can be 0.8 times to 1 times that of the original image, and the aspect ratio can be 0.95 to 1.05, as shown in Figure 5A, Figure 5B, and Figure 5C And shown in the black box in Figure 5D, its size ratio and aspect ratio are listed in Table 1 below, but the present invention is not limited thereto. Table I Figure 5A Figure 5B Figure 5C Figure 5D Size ratio 0.96 0.95 0.91 0.91 aspect ratio 1.04 1.02 0.94 0.98
步驟164為進行一訓練步驟，其係以剪裁後參照感興趣區域對一深度學習演算模組進行訓練至收斂，以得卷積神經網路模型。詳細來說，於本發明之實施例中所使用之深度學習演算模組為ResNet-18演算模組，但本發明並不以此為限。Step 164 is to perform a training step, which is to train a deep learning algorithm module with reference to the region of interest after clipping until convergence, so as to obtain a convolutional neural network model. Specifically, the deep learning calculation module used in the embodiment of the present invention is a ResNet-18 calculation module, but the present invention is not limited thereto.
再者，訓練步驟可更包含以複數個參照受試者臨床資料進行訓練，以提高訓練效果，參照受試者臨床資料包含各參照受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果，但本發明並不以此為限。Moreover, the training step may further include training with the clinical data of a plurality of reference subjects to improve the training effect. The clinical data of the reference subjects include age, sex, maximum tumor diameter, α-fetoprotein value, Child-Pugh score, hepatitis B surface antigen detection results and hepatitis C surface antigen detection results, but the present invention is not limited thereto.
請參照第6圖，第6圖係繪示本發明另一態樣之一實施方式的肝細胞癌微血管浸潤評估系統200之元件示意圖。肝細胞癌微血管浸潤評估系統200包含電腦斷層掃描圖像擷取裝置210及處理器220。電腦斷層掃描圖像擷取裝置210係用以擷取電腦斷層掃描圖像。處理器220訊號連接電腦斷層掃描圖像擷取裝置210，且處理器220包含影像校正程式221及卷積神經網路模型222。影像校正程式221包含一圖形校正模組223，圖形校正模組223係將電腦斷層掃描圖像之感興趣區域進行圖形轉換、剪裁、旋轉或水平翻轉，以得剪裁後感興趣區域，其中剪裁後感興趣區域之各邊長為感興趣區域之各邊長的0.5倍至1倍。卷積神經網路模型222包含一影像特徵值計算模組224及一判斷分類模組225。影像特徵值計算模組224係利用剪裁後感興趣區域進行微血管浸潤評估，以得一影像特徵值。判斷分類模組225係以影像特徵值相對於一閾值之大小關係，輸出一微血管浸潤評估結果，若影像特徵值大於或等於閾值，受試者將被判定為發生微血管浸潤，若影像特徵值小於閾值，受試者將被判定為未發生微血管浸潤。Please refer to FIG. 6 . FIG. 6 is a schematic diagram of components of a system 200 for assessing microvascular invasion of hepatocellular carcinoma according to an embodiment of another aspect of the present invention. The system 200 for assessing microvascular invasion of hepatocellular carcinoma includes a computed tomography image capture device 210 and a processor 220 . The computer tomography image capturing device 210 is used for capturing computer tomography images. The processor 220 is signally connected to the computerized tomography image capture device 210 , and the processor 220 includes an image correction program 221 and a convolutional neural network model 222 . The image correction program 221 includes a pattern correction module 223. The pattern correction module 223 converts, crops, rotates or horizontally flips the region of interest of the computed tomography image to obtain the region of interest after clipping, wherein the region of interest after clipping The length of each side of the region of interest is 0.5 to 1 time the length of each side of the region of interest. The convolutional neural network model 222 includes an image feature calculation module 224 and a judgment classification module 225 . The image feature value calculation module 224 uses the cropped ROI to evaluate microvascular infiltration to obtain an image feature value. The judging and classifying module 225 outputs a microvascular infiltration evaluation result based on the relationship between the image feature value and a threshold value. If the image feature value is greater than or equal to the threshold value, the subject will be judged to have microvascular infiltration. If the image feature value is less than Threshold, subjects will be judged as no microvascular invasion.
詳細來說，卷積神經網路模型222可由一深度學習演算模組訓練至收斂而得，所使用之深度學習演算模組可為ResNet-18演算模組。Specifically, the convolutional neural network model 222 can be obtained by training a deep learning calculation module until convergence, and the used deep learning calculation module can be a ResNet-18 calculation module.
請參照第7圖，第7圖係繪示本發明另一態樣之另一實施方式的肝細胞癌微血管浸潤評估系統200a之元件示意圖。肝細胞癌微血管浸潤評估系統200a中的電腦斷層掃描圖像擷取裝置210a、處理器220a、影像校正程式221a、卷積神經網路模型222a、圖形校正模組223a、影像特徵值計算模組224a及判斷分類模組225a與第6圖之肝細胞癌微血管浸潤評估系統200中的電腦斷層掃描圖像擷取裝置210、處理器220、影像校正程式221、卷積神經網路模型222、圖形校正模組223、影像特徵值計算模組224及判斷分類模組225之技術細節相同，在此不另贅述。Please refer to FIG. 7 . FIG. 7 is a schematic diagram of elements of a hepatocellular carcinoma microvascular invasion assessment system 200a according to another embodiment of another aspect of the present invention. Computed tomography image capture device 210a, processor 220a, image correction program 221a, convolutional neural network model 222a, image correction module 223a, image feature value calculation module 224a in the hepatocellular carcinoma microvascular invasion assessment system 200a And the judgment classification module 225a and the computed tomography image acquisition device 210, processor 220, image correction program 221, convolutional neural network model 222, and pattern correction in the hepatocellular carcinoma microvascular invasion assessment system 200 of FIG. 6 The technical details of the module 223, the image feature value calculation module 224, and the judgment and classification module 225 are the same, and will not be repeated here.
相較於肝細胞癌微血管浸潤評估系統200，肝細胞癌微血管浸潤評估系統200a更包含一資料庫226，資料庫226包含一受試者臨床資料，卷積神經網路模型222a之判斷分類模組225a搭配受試者臨床資料作為參數，與影像特徵值一起進行微血管浸潤評估，其中受試者臨床資料包含受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果。Compared with the evaluation system 200 for microvascular invasion of hepatocellular carcinoma, the evaluation system for microvascular invasion of hepatocellular carcinoma 200a further includes a database 226, the database 226 includes a subject's clinical data, and the judgment and classification module of the convolutional neural network model 222a 225a The subject’s clinical data is used as a parameter, and the microvascular invasion is evaluated together with the image feature value. The subject’s clinical data includes the subject’s age, gender, maximum tumor diameter, α-fetoprotein value, Child-Pugh score, Hepatitis B surface antigen test results and hepatitis C surface antigen test results.
以下將以本發明之實施例及另一實施例與複數個比較例進行比較，進而佐證本發明之肝細胞癌微血管浸潤評估方法具有良好的評估效果與判斷準確度。特別說明的是，用於訓練(Training)之參照電腦斷層掃描圖像及參照受試者臨床資料(訓練組)與用於驗證(Validation)之電腦斷層掃描圖像及受試者臨床資料(驗證組)均是於中國醫藥大學附設醫院進行拍攝與檢驗而得。以下所述之外部資料組(External)均為於其他醫療機構所得之電腦斷層掃描圖像及受試者臨床資料。The following will compare the embodiment of the present invention and another embodiment with a plurality of comparative examples, and further prove that the method for evaluating microvascular invasion of hepatocellular carcinoma of the present invention has good evaluation effect and judgment accuracy. In particular, the reference computed tomography images used for training (Training) and the clinical data of reference subjects (training group) and the computed tomography images used for validation (Validation) and clinical data of subjects (validation) Group) were obtained by photographing and testing in the Affiliated Hospital of China Medical University. The external data sets (External) mentioned below are computed tomography images obtained in other medical institutions and clinical data of subjects.
用於訓練組之參照電腦斷層掃描圖像經圈選出參照感興趣區域後，再經剪裁以得剪裁後參照感興趣區域，隨後剪裁後參照感興趣區域會再經旋轉、裁剪或水平翻轉增加以訓練樣本數後，用於對深度學習演算模組進行訓練。The reference computed tomography image used for the training group is circled to select the reference region of interest, and then clipped to obtain the cropped reference region of interest, and then the cropped reference region of interest is then rotated, cropped or horizontally flipped to increase After the number of training samples, it is used to train the deep learning algorithm module.
用於驗證組之電腦斷層掃描圖像經圈選出感興趣區域後，再經剪裁以得剪裁後感興趣區域，並可用於驗證由深度學習演算模組訓練而得之卷積神經網路模型的判斷準確度。The computed tomography images used in the verification group were selected by circles to select the region of interest, and then clipped to obtain the cropped region of interest, which can be used to verify the convolutional neural network model trained by the deep learning algorithm module Judgment accuracy.
用於外部資料組之電腦斷層掃描圖像可用於經圈選出感興趣區域後，再經剪裁以得剪裁後感興趣區域，並可用於驗證由深度學習演算模組訓練而得之卷積神經網路模型於判斷經不同拍攝設備而得之電腦斷層掃描圖像或來自不同醫療機構之電腦斷層掃描圖像的判斷準確度。Computed tomography images used for external data sets can be used to select regions of interest and then clipped to obtain cropped regions of interest, and can be used to verify the convolutional neural network trained by the deep learning algorithm module The road model is used to judge the accuracy of computer tomography images obtained by different shooting equipment or computer tomography images from different medical institutions.
請參照第8圖，第8圖係繪示訓練組、驗證組及外部資料組的CT圖像拍攝設備及數量關係圖。其中空心柱體部分為於中國醫藥大學附設醫院進行拍攝所得之CT圖像，斜線柱體部分為於其他醫療機構進行拍攝所得之CT圖像，且所有CT圖像均為於動脈相時拍攝而得，即，用於訓練組及驗證組之CT圖像係以中國醫藥大學附設醫院之Aquilion ONE、BrightSpeed、LightSpeed VCT、LightSpeed16及Optima CT660等設備拍攝而得，而用於外部驗證組之CT圖像係以非中國醫藥大學附設醫院之醫療機構的第8圖所陳列之所有設備分別拍攝而得，藉此以驗證本發明之肝細胞癌微血管浸潤評估方法及其評估系統對於各種不同拍攝設備而得之CT圖像的判斷準確度。Please refer to Figure 8, which shows the relationship between the CT image capturing equipment and quantities of the training group, the verification group and the external data group. Among them, the hollow cylinder part is the CT image obtained in the Affiliated Hospital of China Medical University, and the oblique cylinder part is the CT image obtained in other medical institutions, and all the CT images are taken in the arterial phase That is, the CT images used in the training group and the verification group were taken by Aquilion ONE, BrightSpeed, LightSpeed VCT, LightSpeed16 and Optima CT660 equipment in the Affiliated Hospital of China Medical University, and the CT images used in the external verification group The images were taken separately with all the equipment displayed in Figure 8 of a medical institution that is not a hospital affiliated to China Medical University, so as to verify that the method for evaluating microvascular invasion of hepatocellular carcinoma of the present invention and its evaluation system are suitable for various shooting equipment The judgment accuracy of the obtained CT images.
請參照第9圖，第9圖係繪示不同剪裁後參照感興趣區域對參照感興趣區域之各邊長比(於下段及於第8圖中簡稱邊長比)的剪裁後參照感興趣區域對ResNet-18深度學習演算模組訓練結果。如第6圖所示，將不同邊長比的剪裁後參照感興趣區域對ResNet-18深度學習演算模組邊長比進行5折(5-fold)交叉驗證訓練後，可發現以邊長比為0.5倍至1倍之剪裁後參照感興趣區域訓練而得的卷積神經網路模型，其ROC曲線之曲線下面積(Area under curve, AUC)平均值均大於0.75，而其中以邊長比為0.8倍至0.9倍之剪裁後參照感興趣區域訓練而得的卷積神經網路模型具有更高的AUC平均值(Mean AUC)及更高的AUC中位數值，顯示剪裁後感興趣區域之各邊長為感興趣區域之各邊長的0.8倍至0.9倍對於ResNet-18深度學習演算模組的訓練具有更好的效果，也使後續肝細胞癌微血管浸潤的評估具有更佳的判斷準確度。關於ROC曲線之相關細節會於後段說明，在此不另贅述。Please refer to Figure 9. Figure 9 shows the ratio of each side length of different clipped reference ROIs to the reference ROI (referred to as the side length ratio in the lower paragraph and in Figure 8) after clipping the reference ROI The training results of the ResNet-18 deep learning algorithm module. As shown in Figure 6, after clipping different side length ratios and referring to the region of interest, after performing 5-fold cross-validation training on the side length ratio of the ResNet-18 deep learning algorithm module, it can be found that the side length ratio For the convolutional neural network model trained with reference to the region of interest after clipping from 0.5 times to 1 times, the average area under the curve (AUC) of the ROC curve is greater than 0.75, and the side length ratio The convolutional neural network model trained with reference to the region of interest after clipping from 0.8 to 0.9 times has a higher mean AUC (Mean AUC) and a higher median value of AUC, showing that the region of interest after clipping The length of each side is 0.8 times to 0.9 times the length of each side of the region of interest, which has a better effect on the training of the ResNet-18 deep learning algorithm module, and also makes the subsequent assessment of hepatocellular carcinoma microvascular invasion better and more accurate Spend. Relevant details about the ROC curve will be described later, and will not be repeated here.
請參照第10圖，第10圖係繪示不同深度學習演算模組之訓練結果。如第9圖所示，以ResNet-18深度學習演算模組進行訓練所得之卷積神經網路模型具有更高的AUC平均值，故以ResNet-18深度學習演算模組進行訓練所得之卷積神經網路模型具有最佳之判斷準確度，但本發明並不以ResNet-18深度學習演算模組進行訓練所得之卷積神經網路模型為限，所有可經訓練而得到相同功效之卷積神經網路模型的深度學習演算模組均可用於本發明之肝細胞癌微血管浸潤評估方法中。Please refer to Figure 10, which shows the training results of different deep learning algorithm modules. As shown in Figure 9, the convolutional neural network model trained with the ResNet-18 deep learning algorithm module has a higher average AUC, so the convolutional neural network model trained with the ResNet-18 deep learning algorithm module The neural network model has the best judgment accuracy, but the present invention is not limited to the convolutional neural network model obtained by training the ResNet-18 deep learning algorithm module. All convolutional neural network models that can be trained to obtain the same effect The deep learning calculation module of the neural network model can be used in the method for evaluating the microvascular invasion of hepatocellular carcinoma of the present invention.
請參照第11A圖、第11B圖及第12圖，第11A圖係繪示實施例1及實施例2對不同輸入資料的數據預測結果之ROC曲線。第11B圖係繪示比較例1與比較例2對不同輸入資料的數據預測結果之ROC曲線。第12圖係繪示實施例1、實施例2、比較例1與比較例2對不同輸入資料的數據預測結果之ROC曲線的AUC示意圖。其中實施例1、實施例2、比較例1與比較例2之訓練參數如下表二所示，實施例1及實施例2均係以第3圖所示之肝細胞癌微血管浸潤評估方法100a訓練而得。 表二 實施例1 實施例2 比較例1 比較例2 使用之演算模組 ResNet-18深度學習演算模組 ResNet-18深度學習演算模組 SVM機器學習演算模組 SVM機器學習演算模組 訓練方式 僅使用動脈相CT圖像 使用動脈相CT圖像及CF 僅使用CF 使用動脈相CT圖像及CF Please refer to FIG. 11A, FIG. 11B and FIG. 12. FIG. 11A shows the ROC curves of the data prediction results for different input data in Embodiment 1 and Embodiment 2. FIG. 11B shows the ROC curves of the data prediction results of Comparative Example 1 and Comparative Example 2 for different input data. FIG. 12 is a schematic diagram showing the AUC curves of the ROC curves of the data prediction results of Example 1, Example 2, Comparative Example 1 and Comparative Example 2 for different input data. Wherein the training parameters of Embodiment 1, Embodiment 2, Comparative Example 1 and Comparative Example 2 are shown in Table 2 below, and Embodiment 1 and Embodiment 2 are all trained with the hepatocellular carcinoma microvascular invasion evaluation method 100a shown in Fig. 3 and get. Table II Example 1 Example 2 Comparative example 1 Comparative example 2 Calculation module used ResNet-18 deep learning algorithm module ResNet-18 deep learning algorithm module SVM Machine Learning Calculation Module SVM Machine Learning Calculation Module training method Use only arterial phase CT images Using arterial phase CT images and CF Only use CF Using arterial phase CT images and CF
詳細來說，ROC空間可被從(0,0)到(1,1)連線而成之對角線劃分為左上區域及右下區域，於左上區域的ROC曲線代表具有較佳的分類結果，即真陽性率高於偽陽性率；於右下區域的ROC曲線代表具有較差的分類結果，即真陽性率低於偽陽性率。於第11A圖可見，無論是驗證組或外部資料組，實施例2相較於實施例1均具有較佳的判斷準確性，此外亦可由第11A圖之結果，再依據前述約登指數之定義分別對實施例1與實施例2進行計算，並將計算而得的約登指數之最大值作為閾值。於實施例1中驗證組的閾值為0.5694，實施例2之驗證組的閾值為0.8082，但本發明並不以此為限。於第11B圖可見，無論是驗證組或外部資料組，比較例1相較於比較例2均具有較差的判斷準確性，顯示SVM機器學習演算模組並不適合以圖像進行訓練。且將第11A圖與第11B圖相互比較，可發現第11A圖之曲線更偏向ROC空間的左上方。此外於第12圖中顯示，除訓練組外，實施例2在驗證組及外部資料組均具有較高的AUC平均值，以上結果均顯示使用動脈相電腦斷層掃描圖像及受試者臨床資料對ResNet-18深度學習演算模組所得之卷積神經網路模型具有更高的判斷準確度。In detail, the ROC space can be divided into the upper left area and the lower right area by the diagonal line connecting from (0,0) to (1,1). The ROC curve in the upper left area represents better classification results , that is, the true positive rate is higher than the false positive rate; the ROC curve in the lower right area represents a poor classification result, that is, the true positive rate is lower than the false positive rate. It can be seen in Figure 11A that whether it is the verification group or the external data group, Example 2 has better judgment accuracy than Example 1. In addition, the results in Figure 11A can also be used according to the definition of the Youden index above The calculations were carried out for Example 1 and Example 2 respectively, and the maximum value of the calculated Youden index was used as the threshold. In embodiment 1, the threshold value of the verification group is 0.5694, and in embodiment 2, the threshold value of the verification group is 0.8082, but the present invention is not limited thereto. It can be seen in Fig. 11B that whether it is the verification group or the external data group, Comparative Example 1 has poorer judgment accuracy than Comparative Example 2, indicating that the SVM machine learning algorithm module is not suitable for training with images. And comparing Figure 11A with Figure 11B, it can be found that the curve in Figure 11A is more inclined to the upper left of the ROC space. In addition, it is shown in Figure 12 that, except for the training group, Example 2 has a higher AUC average value in the verification group and the external data group, and the above results all show that the arterial phase computed tomography image and the clinical data of the test subject were used The convolutional neural network model obtained by the ResNet-18 deep learning algorithm module has higher judgment accuracy.
請參照第13A圖、第13B圖及第13C圖，第13A圖係繪示實施例1、實施例2、比較例1與比較例2對訓練組產生之預測結果的數據分析圖。第13B圖係繪示實施例1、實施例2、比較例1與比較例2對驗證組產生之預測結果的數據分析圖。第13C圖實施例1、實施例2、比較例1與比較例2對外部資料組產生之預測結果的數據分析圖。由第13A圖、第13B圖及第13C圖可見，實施例1及實施例2在訓練組、驗證組及外部資料組均具有良好的準確度(Accuracy)、F1-score、精密度(Precision)、靈敏度及特異度，並無過度擬合(overfitting)的跡象。Please refer to FIG. 13A, FIG. 13B and FIG. 13C. FIG. 13A is a data analysis diagram showing the prediction results of the training group produced by Example 1, Example 2, Comparative Example 1 and Comparative Example 2. FIG. 13B is a data analysis diagram showing the prediction results of Example 1, Example 2, Comparative Example 1 and Comparative Example 2 for the verification group. FIG. 13C is a data analysis chart of the prediction results generated by the external data set in Example 1, Example 2, Comparative Example 1 and Comparative Example 2. It can be seen from Figure 13A, Figure 13B and Figure 13C that Example 1 and Example 2 have good accuracy (Accuracy), F1-score, and precision (Precision) in the training group, verification group and external data group , sensitivity and specificity, and there was no sign of overfitting.
綜上所述，本發明之肝細胞癌微血管浸潤評估方法及其評估系統可藉由對電腦斷層掃描圖像進行感興趣區域的圈選，並將感興趣區域剪裁至原邊長的0.5倍至1倍，以得到剪裁後感興趣區域後，再將剪裁後感興趣區域進行深度學習演算模組的訓練，並以訓練所得之卷積神經網路模型用於對不同拍攝來源的進行快速且準確的肝細胞癌微血管浸潤評估，以有效達到輔助醫生對病人治療方式的決策的目標。In summary, the method for evaluating microvascular invasion of hepatocellular carcinoma and its evaluation system of the present invention can be used to circle the region of interest on the computerized tomography image, and cut the region of interest to 0.5 times the original side length to 1 times, to get the trimmed region of interest, then train the trimmed region of interest with the deep learning algorithm module, and use the trained convolutional neural network model to quickly and accurately capture different shooting sources The evaluation of microvascular invasion of hepatocellular carcinoma can effectively achieve the goal of assisting doctors in making decisions about patients' treatment methods.
雖然本發明已以實施方式揭露如上，然其並非用以限定本發明，任何熟習此技藝者，在不脫離本發明之精神和範圍內，當可作各種之更動與潤飾，因此本發明之保護範圍當視後附之申請專利範圍所界定者為準。Although the present invention has been disclosed above in terms of implementation, it is not intended to limit the present invention. Anyone skilled in this art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection of the present invention The scope shall be defined by the appended patent application scope.
100,100a:肝細胞癌微血管浸潤評估方法 110,110a,120,120a,130,130a,140,140a,150,150a,160,161,162,163,164:步驟 121:感興趣區域 131:剪裁後感興趣區域 200,200a:肝細胞癌微血管浸潤評估系統 210,210a:電腦斷層掃描圖像擷取裝置 220,220a:處理器 221,221a:影像校正程式 222,222a:卷積神經網路模型 223,223a:圖形校正模組 224,224a:影像特徵值計算模組 225,225a:判斷分類模組 226:資料庫 d:邊長 100,100a: Methods for the assessment of microvascular invasion in hepatocellular carcinoma 110,110a,120,120a,130,130a,140,140a,150,150a,160,161,162,163,164: steps 121: Region of interest 131: Region of interest after clipping 200,200a: Microvascular invasion assessment system for hepatocellular carcinoma 210, 210a: Computed tomography image capture device 220, 220a: Processor 221,221a: Image correction programs 222, 222a: Convolutional Neural Network Models 223,223a: Graphics Correction Module 224,224a: Image feature value calculation module 225,225a: Judgment classification module 226: Database d: side length
為讓本發明之上述和其他目的、特徵、優點與實施例能更明顯易懂，所附圖式之說明如下： 第1圖係繪示本發明一態樣之一實施方式的肝細胞癌微血管浸潤評估方法之步驟流程圖； 第2圖係繪示感興趣區域與剪裁後感興趣區域之示意圖； 第3圖係繪示本發明一態樣之另一實施方式的肝細胞癌微血管浸潤評估方法之步驟流程圖； 第4A圖為一參照電腦斷層掃描圖像之剪裁後參照感興趣區域之影像； 第4B圖為將第4A圖中的剪裁後參照感興趣區域旋轉-10 o之影像； 第4C圖為將第4A圖中的剪裁後參照感興趣區域旋轉-5 o之影像； 第4D圖為將第4A圖中的剪裁後參照感興趣區域旋轉5 o之影像； 第4E圖為將第4A圖中的剪裁後參照感興趣區域旋轉10 o之影像； 第4F圖為將第4A圖中的剪裁後參照感興趣區域水平翻轉之影像； 第5A圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之示意圖； 第5B圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之另一示意圖； 第5C圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之再一示意圖； 第5D圖係繪示將第4A圖中的剪裁後參照感興趣區域進行裁剪之又一示意圖； 第6圖係繪示本發明另一態樣之一實施方式的肝細胞癌微血管浸潤評估系統之元件示意圖； 第7圖係繪示本發明另一態樣之另一實施方式的肝細胞癌微血管浸潤評估系統之元件示意圖； 第8圖係繪示訓練組、驗證組及外部資料組的電腦斷層掃描圖像拍攝設備及數量關係圖； 第9圖係繪示不同剪裁後參照感興趣區域對參照感興趣區域之各邊長比的剪裁後參照感興趣區域對ResNet-18深度學習演算模組訓練結果； 第10圖係繪示不同深度學習演算模組之訓練結果； 第11A圖係繪示實施例1及實施例2對不同輸入資料的數據預測結果之接收者操作特徵曲線； 第11B圖係繪示比較例1與比較例2對不同輸入資料的數據預測結果之接收者操作特徵曲線； 第12圖係繪示實施例1、實施例2、比較例1與比較例2對不同輸入資料的數據預測結果之接收者操作特徵曲線的曲線下面積示意圖； 第13A圖係繪示實施例1、實施例2、比較例1與比較例2對訓練組產生之預測結果的數據分析圖； 第13B圖係繪示實施例1、實施例2、比較例1與比較例2對驗證組產生之預測結果的數據分析圖；以及 第13C圖實施例1、實施例2、比較例1與比較例2對外部資料組產生之預測結果的數據分析圖。 In order to make the above and other objects, features, advantages and embodiments of the present invention more comprehensible, the accompanying drawings are described as follows: Fig. 1 shows the hepatocellular carcinoma microvessel of an embodiment of an aspect of the present invention A flow chart of the steps of the invasion assessment method; Figure 2 is a schematic diagram showing the region of interest and the clipped region of interest; Figure 3 is a schematic diagram of the method for assessing microvascular invasion of hepatocellular carcinoma according to another embodiment of an aspect of the present invention Figure 4A is an image of a reference region of interest after cropping of a reference computerized tomography image; Figure 4B is an image of the cropped reference region of interest rotated by -10o in Figure 4A; Figure 4C is the cropped image in Figure 4A and rotated by -5o with reference to the region of interest; Figure 4D is the image rotated by 5o with reference to the region of interest after cropping in Figure 4A; Figure 4E is the image of The cropped image in Figure 4A is rotated 10o with reference to the region of interest; Figure 4F is the horizontally flipped image after cropping in Figure 4A with reference to the ROI; Figure 5A shows the cropped image in Figure 4A A schematic diagram of clipping with reference to the region of interest; Figure 5B is another schematic diagram of clipping in Figure 4A with reference to the region of interest; Figure 5C is a schematic diagram of clipping in Figure 4A with reference to Another schematic diagram of clipping the region of interest; FIG. 5D is another schematic diagram of clipping the clipping in FIG. 4A with reference to the region of interest; FIG. 6 is an implementation of another aspect of the present invention The schematic diagram of the components of the hepatocellular carcinoma microvascular invasion assessment system; Figure 7 is a schematic diagram of the components of the hepatocellular carcinoma microvascular invasion assessment system according to another embodiment of the present invention; Figure 8 is a training set , verification group and external data group computer tomography image capture equipment and quantitative relationship diagram; Fig. 9 shows the reference region of interest after clipping and the ratio of each side length of the reference region of interest after clipping The training results of the ResNet-18 deep learning algorithm module; Fig. 10 shows the training results of different deep learning calculus modules; Fig. 11A shows the difference between the data prediction results of Example 1 and Example 2 for different input data Receiver operating characteristic curve; Figure 11B shows the receiver operating characteristic curve of comparative example 1 and comparative example 2 on the data prediction results of different input data; Figure 12 shows embodiment 1, embodiment 2, comparative example 1 and Comparative Example 2 are schematic diagrams of the area under the curve of the receiver operating characteristic curve of the data prediction results of different input data; Figure 13A is a diagram showing the results of Example 1, Example 2, Comparative Example 1 and Comparative Example 2 for the training group Figure 13B is a data analysis chart showing the prediction results of Example 1, Example 2, Comparative Example 1 and Comparative Example 2 for the verification group; and Figure 13C Example 1, Implementation Example 2, the data analysis diagram of the prediction results generated by comparative example 1 and comparative example 2 on the external data set.
100:肝細胞癌微血管浸潤評估方法 100: Methods for the assessment of microvascular invasion in hepatocellular carcinoma
110,120,130,140,150:步驟 110,120,130,140,150: steps
Claims (9)
一種肝細胞癌微血管浸潤評估方法，包含：提供一受試者之一電腦斷層掃描圖像，該電腦斷層掃描圖像由一電腦斷層掃描圖像擷取裝置擷取而得，其中該電腦斷層掃描圖像為一動脈相電腦斷層掃描圖像，且該電腦斷層掃描圖像包含一腫瘤區域；進行一感興趣區域圈選步驟，其係於該電腦斷層掃描圖像圈選一感興趣區域，其中該感興趣區域係由一放射科醫師進行人工圈選，且該感興趣區域包含該腫瘤區域；進行一裁邊步驟，其係以一影像校正程式將該感興趣區域之邊緣進行剪裁，以得一剪裁後感興趣區域，其中該剪裁後感興趣區域之各邊長為該感興趣區域之各邊長的0.8倍至0.9倍；進行一數據預測步驟，其係利用一卷積神經網路模型對該剪裁後感興趣區域進行數據預測，以得一影像特徵值，其中該卷積神經網路模型是由一深度學習演算模組訓練而得，且該深度學習演算模組為一ResNet-18演算模組；以及進行一判斷分類步驟，其係利用該卷積神經網路模型對該影像特徵值進行判斷，若該影像特徵值大於或等於一閾值，該受試者將被判定為發生微血管浸潤，若該影像特徵值小於該閾值，該受試者將被判定為未發生微血管浸潤。 A method for assessing microvascular invasion of hepatocellular carcinoma, comprising: providing a computerized tomography image of a subject, the computerized tomography image being captured by a computerized tomography image capture device, wherein the computerized tomography The image is an arterial phase computed tomography image, and the computed tomography image includes a tumor region; performing a region of interest circle selection step, which is to circle a region of interest in the computer tomography image, wherein The region of interest is manually selected by a radiologist, and the region of interest includes the tumor region; a trimming step is performed, which is to clip the edge of the region of interest with an image correction program to obtain A clipped region of interest, wherein the length of each side of the clipped region of interest is 0.8 to 0.9 times the length of each side of the region of interest; a data prediction step is performed using a convolutional neural network model Perform data prediction on the cropped region of interest to obtain an image feature value, wherein the convolutional neural network model is trained by a deep learning algorithm module, and the deep learning algorithm module is a ResNet-18 Calculation module; and performing a judgment and classification step, which is to use the convolutional neural network model to judge the image feature value, if the image feature value is greater than or equal to a threshold, the subject will be judged as microvascular Infiltration, if the image feature value is less than the threshold, the subject will be judged as no microvascular invasion. 如請求項1所述之肝細胞癌微血管浸潤評估 方法，其中該動脈相電腦斷層掃描圖像係於一動脈之阻射率為120亨氏單位時拍攝而得。 Assessment of microvascular invasion of hepatocellular carcinoma as described in claim 1 The method, wherein the arterial phase computed tomography image is taken when the radioblocking rate of an artery is 120 Hounsfield units. 如請求項1所述之肝細胞癌微血管浸潤評估方法，其中該判斷分類步驟更包含以一受試者臨床資料進行判斷，以提高判斷準確度，該受試者臨床資料包含該受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果。 The method for assessing microvascular invasion of hepatocellular carcinoma as described in claim 1, wherein the step of judging and classifying further includes making a judgment based on the clinical data of a subject to improve the accuracy of judgment, and the clinical data of the subject includes the subject’s clinical data Age, gender, largest tumor diameter, α-fetoprotein value, Child-Pugh score, hepatitis B surface antigen test results and hepatitis C surface antigen test results. 如請求項1所述之肝細胞癌微血管浸潤評估方法，更包含：進行一建模步驟，包含下述步驟：提供一參照電腦斷層掃描圖像資料庫，其中該參照電腦斷層掃描圖像資料庫包含複數個參照電腦斷層掃描圖像，各該參照電腦斷層掃描圖像為一參照動脈相電腦斷層掃描圖像，且各該參照電腦斷層掃描圖像包含一參照腫瘤區域；進行一參照感興趣區域圈選步驟，其係於各該參照電腦斷層掃描圖像圈選一參照感興趣區域，以得複數個參照感興趣區域，其中各該參照感興趣區域包含各該參照腫瘤區域；進行一訓練前影像處理步驟，其係對各該參照感興趣區域進行剪裁，以得複數個剪裁後參照感興趣區域，其 中各該剪裁後參照感興趣區域之各邊長為各該參照感興趣區域之各邊長的0.8倍至0.9倍；以及進行一訓練步驟，其係以該些剪裁後參照感興趣區域對該深度學習演算模組進行訓練至收斂，以得該卷積神經網路模型。 The method for assessing microvascular invasion of hepatocellular carcinoma according to claim 1, further comprising: performing a modeling step, comprising the step of: providing a reference computerized tomography image database, wherein the reference computer tomography image database comprising a plurality of reference computed tomography images, each of the reference computed tomography images being a reference arterial phase computed tomography image, and each of the reference computed tomography images including a reference tumor region; performing a reference region of interest The circle selection step is to circle a reference region of interest in each of the reference computerized tomography images to obtain a plurality of reference regions of interest, wherein each of the reference regions of interest includes each of the reference tumor regions; before performing a training The image processing step is to clip each reference region of interest to obtain a plurality of clipped reference regions of interest. Each side length of the reference region of interest after the clipping is 0.8 times to 0.9 times of each side length of the reference region of interest; The deep learning algorithm module is trained until convergence to obtain the convolutional neural network model. 如請求項4所述之肝細胞癌微血管浸潤評估方法，其中各該參照動脈相電腦斷層掃描圖像係於一動脈之阻射率為120亨氏單位時拍攝而得。 The method for assessing microvascular invasion of hepatocellular carcinoma according to Claim 4, wherein each of the reference arterial phase computed tomography images is taken when the radioblocking rate of an artery is 120 Hounsfield units. 如請求項4所述之肝細胞癌微血管浸潤評估方法，其中該訓練前影像處理步驟更包含對各該剪裁後參照感興趣區域進行旋轉、裁剪或水平翻轉。 The method for assessing microvascular invasion of hepatocellular carcinoma according to claim 4, wherein the pre-training image processing step further includes rotating, cropping or horizontally flipping each trimmed reference region of interest. 如請求項4所述之肝細胞癌微血管浸潤評估方法，其中該訓練步驟更包含以複數個參照受試者臨床資料進行訓練，以提高訓練效果，該些參照受試者臨床資料包含各該參照受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果。 The method for assessing microvascular invasion of hepatocellular carcinoma as described in Claim 4, wherein the training step further includes training with clinical data of a plurality of reference subjects to improve the training effect, and the clinical data of these reference subjects include each of the reference subjects Subject's age, gender, largest tumor diameter, α-fetoprotein value, Child-Pugh score, hepatitis B surface antigen test results and hepatitis C surface antigen test results. 一種肝細胞癌微血管浸潤評估系統，包含：一電腦斷層掃描圖像擷取裝置，其係用以擷取一電腦斷層掃描圖像，其中該電腦斷層掃描圖像為一動脈相電腦斷 層掃描圖像；以及一處理器，其訊號連接該電腦斷層掃描圖像擷取裝置，其中該處理器包含：一影像校正程式，該影像校正程式包含一圖形校正模組，該圖形校正模組係將該電腦斷層掃描圖像之一感興趣區域進行圖形轉換、剪裁、旋轉或水平翻轉，以得一剪裁後感興趣區域，其中該剪裁後感興趣區域之各邊長為該感興趣區域之各邊長的0.8倍至0.9倍；及一卷積神經網路模型，包含：一影像特徵值計算模組，其係利用該剪裁後感興趣區域進行微血管浸潤評估，以得一影像特徵值；及一判斷分類模組，其係以該影像特徵值相對於一閾值之大小關係，輸出一微血管浸潤評估結果，若該影像特徵值大於或等於該閾值，該受試者將被判定為發生微血管浸潤，若該影像特徵值小於該閾值，該受試者將被判定為未發生微血管浸潤。 其中，該卷積神經網路模型係由一深度學習演算模組訓練至收斂而得，且該深度學習演算模組為一ResNet-18演算模組。 A system for assessing microvascular invasion of hepatocellular carcinoma, comprising: a computed tomography image capture device, which is used to capture a computed tomography image, wherein the computed tomography image is an arterial phase computed tomography layer scan image; and a processor, the signal of which is connected to the computer tomography image capture device, wherein the processor includes: an image correction program, the image correction program includes a graphic correction module, the graphic correction module It is to perform graphic conversion, clipping, rotation or horizontal flipping on a region of interest of the computerized tomography image to obtain a clipped region of interest, wherein the length of each side of the clipped region of interest is the length of the region of interest 0.8 times to 0.9 times the length of each side; and a convolutional neural network model, including: an image feature value calculation module, which uses the cropped region of interest to evaluate microvascular infiltration to obtain an image feature value; and a judgment classification module, which outputs a microvascular infiltration evaluation result based on the size relationship between the image feature value and a threshold value, and if the image feature value is greater than or equal to the threshold value, the subject will be judged as microvascular Infiltration, if the image feature value is less than the threshold, the subject will be judged as no microvascular invasion. Wherein, the convolutional neural network model is obtained by training a deep learning calculation module until convergence, and the deep learning calculation module is a ResNet-18 calculation module. 如請求項8所述之肝細胞癌微血管浸潤評估系統，其中該卷積神經網路模型更包含：一資料庫，其包含一受試者臨床資料，該卷積神經網路模型之該判斷分類模組搭配該受試者臨床資料作為參數， 與該影像特徵值一起進行該微血管浸潤評估，其中該受試者臨床資料包含該受試者之年齡、性別、最大腫瘤直徑、α-胎兒蛋白數值、Child-Pugh評分、B型肝炎表面抗原檢測結果及C型肝炎表面抗原檢測結果。 The evaluation system for microvascular invasion of hepatocellular carcinoma as described in Claim 8, wherein the convolutional neural network model further includes: a database containing clinical data of a subject, the judgment classification of the convolutional neural network model The module is matched with the subject's clinical data as parameters, The evaluation of microvascular invasion is carried out together with the image characteristic value, wherein the subject's clinical data includes the subject's age, sex, largest tumor diameter, α-fetoprotein value, Child-Pugh score, and hepatitis B surface antigen detection Results and hepatitis C surface antigen detection results.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW110132316A TWI809488B (en) | 2021-08-31 | 2021-08-31 | Evaluating method of microvascular invasion in hepatocellular carcinoma and evaluating system thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW110132316A TWI809488B (en) | 2021-08-31 | 2021-08-31 | Evaluating method of microvascular invasion in hepatocellular carcinoma and evaluating system thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| TW202312183A TW202312183A (en) | 2023-03-16 |
| TWI809488B true TWI809488B (en) | 2023-07-21 |
Family
ID=86690617
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| TW110132316A TWI809488B (en) | 2021-08-31 | 2021-08-31 | Evaluating method of microvascular invasion in hepatocellular carcinoma and evaluating system thereof |
Country Status (1)
| Country | Link |
|---|---|
| TW (1) | TWI809488B (en) |
Citations (4)
* Cited by examiner, † Cited by third party| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107368690A (en) * | 2017-08-09 | 2017-11-21 | 贵阳朗玛信息技术股份有限公司 | The preprocess method and device of medical image picture |
| CN110852350A (en) * | 2019-10-21 | 2020-02-28 | 北京航空航天大学 | Pulmonary nodule benign and malignant classification method and system based on multi-scale migration learning |
| US20200182877A1 (en) * | 2017-09-01 | 2020-06-11 | The Regents Of The University Of California | Phenotypic profiling of hepatocellular carcinoma circulating tumor cells for treatment selection |
| CN112651507A (en) * | 2020-12-22 | 2021-04-13 | 福建医科大学附属第一医院 | Method for constructing microvascular invasion prediction model of hepatocellular carcinoma and probability prediction method |
-
2021
- 2021-08-31 TW TW110132316A patent/TWI809488B/en active
Patent Citations (4)
* Cited by examiner, † Cited by third party| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107368690A (en) * | 2017-08-09 | 2017-11-21 | 贵阳朗玛信息技术股份有限公司 | The preprocess method and device of medical image picture |
| US20200182877A1 (en) * | 2017-09-01 | 2020-06-11 | The Regents Of The University Of California | Phenotypic profiling of hepatocellular carcinoma circulating tumor cells for treatment selection |
| CN110852350A (en) * | 2019-10-21 | 2020-02-28 | 北京航空航天大学 | Pulmonary nodule benign and malignant classification method and system based on multi-scale migration learning |
| CN112651507A (en) * | 2020-12-22 | 2021-04-13 | 福建医科大学附属第一医院 | Method for constructing microvascular invasion prediction model of hepatocellular carcinoma and probability prediction method |
Non-Patent Citations (1)
* Cited by examiner, † Cited by third party| Title |
|---|
| 期刊 Song, D., Wang, Y., Wang, W. et al. Using deep learning to predict microvascular invasion in hepatocellular carcinoma based on dynamic contrast-enhanced MRI combined with clinical parameters. J Cancer Res Clin Oncol 147 2021/4/10 3757–3767 * |
Also Published As
| Publication number | Publication date |
|---|---|
| TW202312183A (en) | 2023-03-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112561908B (en) | 2021-11-23 | Mammary gland image focus matching method, device and storage medium |
| Tolkachev et al. | 2020 | Deep learning for diagnosis and segmentation of pneumothorax: The results on the Kaggle competition and validation against radiologists |
| US10004448B2 (en) | 2018-06-26 | Osteoporosis diagnostic support apparatus |
| JP2007307358A (en) | 2007-11-29 | Method, apparatus and program for image treatment |
| CN110728239B (en) | 2023-10-13 | An automatic recognition system for gastric cancer enhanced CT images using deep learning |
| CN105488781A (en) | 2016-04-13 | Dividing method based on CT image liver tumor focus |
| CN113712594A (en) | 2021-11-30 | Medical image processing apparatus and medical imaging apparatus |
| TWI839758B (en) | 2024-04-21 | Processing method of medical image and computing apparatus for processing medical image |
| Wang et al. | 2022 | Development and validation of a deep learning model for survival prognosis of transcatheter arterial chemoembolization in patients with intermediate-stage hepatocellular carcinoma |
| CN113469942B (en) | 2022-02-22 | CT image lesion detection method |
| Ma et al. | 2020 | Automated identification of optimal portal venous phase timing with convolutional neural networks |
| Feng et al. | 2022 | Multi-scale information with attention integration for classification of liver fibrosis in B-mode US image |
| Hoogi et al. | 2017 | A fully-automated pipeline for detection and segmentation of liver lesions and pathological lymph nodes |
| Zhu et al. | 2022 | A multi-task two-path deep learning system for predicting the invasiveness of craniopharyngioma |
| Xiao et al. | 2021 | A cascade and heterogeneous neural network for CT pulmonary nodule detection and its evaluation on both phantom and patient data |
| CN105193436B (en) | 2018-04-20 | Method for evaluating exposure dose of digital mammary X-ray imaging system |
| Franco et al. | 2021 | Glaucoma patient screening from online retinal fundus images via Artificial Intelligence |
| TWI809488B (en) | 2023-07-21 | Evaluating method of microvascular invasion in hepatocellular carcinoma and evaluating system thereof |
| CN112330640B (en) | 2024-07-16 | Method, device and equipment for segmenting nodule region in medical image |
| US9014440B2 (en) | 2015-04-21 | Dental cysts detector |
| JP2003299646A (en) | 2003-10-21 | Image analyzer |
| CN117524487A (en) | 2024-02-06 | Artificial intelligence-based method and system for evaluating risk of arteriosclerotic plaque |
| Xiong et al. | 2019 | Deep ensemble learning network for kidney lesion detection |
| KR102690591B1 (en) | 2024-07-31 | Automatic detection and segmentation method of urinary stones based on deep learning |
| TWI796684B (en) | 2023-03-21 | Non-invasive method of evaluating hemoglobin concentration and evaluating system thereof |