利用深度卷積神經網路建立乳房X光影像分類暨病兆偵測模型

Classification and Lesion Detection models of Mammography Using Deep Convolutional Neural Network

傅筱芸(Hsiao-Yun Fu)

指導教授 ： 丁慧枝

共同指導教授 ： 許士彥(Shih-Yen Hsu)

義守大學/智慧科技學院/資訊工程學系/碩士(2022年)

摘要

乳房攝影(Mammography)為早期診斷乳癌的篩檢工具之一，臨床上針對乳房攝影影像判讀全仰賴臨床醫師之經驗進行診斷。無相對應之輔助診斷之工具，不僅費時且長時間閱片所產生之疲勞使得造成疾病的誤判。因此本研究運用人工智慧(Artificial Intelligence)卷積神經網路(Convolutional neural network, CNN)技術，開發無異常、良性及惡性之多類別乳房攝影影像分類以及病兆偵測模型。本研究目的為輔助放射科醫師減少判讀大量乳房攝影影像所需花費的時間及降低誤判率。
研究收集乳房攝影影像及依據診斷結果進行分類，影像類型包含頭腳向及斜位向，共計有1554張。建立模型前，首先將影像進行切割、對比強化、影像擴增等前處理，接著透過預訓練模型VGG19及RESNET50搭配轉移學習技術，建立無異常、良性及惡性之影像分類模型，同時建立良性及惡性類別之病兆位置偵測模型進行自動標記。最後透過多個統計指標評估模型效能。
研究結果於影像分類中，透過RESNET-50模型可獲得Accuracy、Recall、Precision、F-score分別為87%、78%、89%、83%之分類模型；在病兆位置自動偵測中，經由YOLO模型，分別完成良性鈣化點、惡性鈣化點、以及惡性腫瘤之AI模型開發，其偵測結果Average precision分別為76%、71%、以及63%。
本研究應用AI深度卷積神經網路技術於乳房X光影像進行兩階段分析，可提供臨床進行影像期別分類與病兆位置偵測等輔助，未來可協助臨床作業、減少影像判讀時間及人為之誤判。

關鍵字

人工智慧 ； 卷積神經網路 ； 乳房攝影

並列摘要

Mammography is one of the imaging tools for the early diagnosis of breast cancer. Clinically, the interpretation of mammography images relies on the experience of clinicians for diagnosis. There is no corresponding accessorial diagnosis tool, which is not only time-consuming but also caused by long-term reading of images. Therefore, this study uses artificial intelligence (AI) convolutional neural network (CNN) technology to develop normal, benign and malignant multi-category mammography image classification and detection models. To assist radiologists to reduce the time spent interpreting large numbers of mammography images and reduce the false positives rate.
This study collected mammography images and classified them according to the diagnostic results. The types of images included head-to-foot and oblique directions, with a total of 1554 images. Before building the model, all the images were pre-processed such as cutting, contrast enhancement, and image augmentation. Then, the pre-trained models VGG19 and RESNET50 were combined with transfer learning technology to establish a non-abnormal, benign and malignant image classification model, as well as benign and malignant images. The category's symptom location detection model was automatically marked. Finally, the model performance is evaluated through multiple statistical indicators.
In the research results, the classification model was obtained the accuracy, recall, precision, and F-score of 87%, 78%, 89%, and 83% through the RESNET-50 model; The YOLO model has completed the development of AI models for benign calcifications, malignant calcifications, and malignant tumors. The average precision of the detection results is 76%, 71%, and 63%, respectively.
This study applies AI deep convolutional neural network technology to perform two-stage analysis of mammography images, which can provide clinical assistance in image classification and symptom location detection. In the future, it can assist clinical operations, reduce image interpretation time and human misjudgment.

並列關鍵字

artificial intelligence ； Convolutional Neural Networks ； Mammography

參考文獻

1. 衛生福利部國民健康署, 癌症篩檢介紹(大腸癌、口腔癌、子宮頸癌、乳癌). 2010.

2. 衛生福利部統計處, 乳房攝影檢查人數. 2021.

3. Gøtzsche, P.C. and K.J. Jørgensen, Screening for breast cancer with mammography. Cochrane Database of Systematic Reviews, 2013(6).

4. Vyborny, C.J. and M.L. Giger, Computer vision and artificial intelligence in mammography. American Journal of Roentgenology, 1994. 162(3): p. 699-708.

5. Gøtzsche, P.C., Breast Cancer Screening. International Agency for Research on Cancer (IARC) Handbooks of Cancer Prevention. Vol. 7.: Harri Vainio and Franca Bianchini (eds). Lyon: IARC Press, 2002, pp. 229, Euro 40.00. ISBN: 92-832-3007-8. International Journal of Epidemiology, 2003. 32(3): p. 472-472.