運用深度學習於乳房X光照腫塊檢測

Abstract

乳癌是女性最常見的癌症之一，在每年國民健康署公布的十大癌症排名中，乳癌的名次幾乎保持在第一名，並且每十萬人的發生率逐年上升。而乳癌是一個經由篩檢可以提早發現與早期治療、五年內存活率大於九成的癌症，國民健康署自民國93年起大力推動乳癌篩檢，呼籲民眾面對乳癌要及早預防、及早發現、及早治療等，至103年超過80萬女性接受乳房X光攝影檢查。 目前最有效檢測乳癌的方式是透過乳房X光攝影檢查乳房中的微鈣化或腫塊，在本論文中設計出多個卷積神經網路，讓卷積神經網路自動學習影像中微鈣化與腫塊的特徵，省去傳統影像辨識需人工設計影像特徵的步驟，並使用INbreast資料庫提供之微鈣化與腫塊位置等資訊，並提取適當大小的子影像進行模型訓練，再使用訓練好的模型對全域影像進行可疑區域偵測，實驗結果顯示腫塊與微鈣化子影像分類器ROC曲線面積分別達到0.98與0.89，可見卷積神經網路確實具有在子影像中分辨腫塊、微鈣化與正常組織的能力。最後使用訓練完成的分類器在全域影像中進行偵測，在選擇適當的網路參數時，腫塊偵測的sensitivity達到80%時平均每張影像會出現0.80個false positive，具有相當程度的辨識能力，在微鈣化偵測對於單獨的微鈣化辨識能力受限於影像中雜訊過多而較差，對於群聚微鈣化則仍有一定的偵測能力。

Breast cancer is one of the most common cancers that affect women lives. According to the list of the top ten common cancer types published by the Health Promotion Administration, MOHW, Breast cancer remains the first place most of the time. Moreover, the incidence rate of the breast cancer per 100 thousand people is increasing every year. Since breast cancer can be detected and treated earlier through cancer screening and the survival rate is more than ninety percent within five years, MOHW started promoting breast cancer screening and appealing people to prevent, detect and treat it as soon as possible since 2004. Till 2014, more than 800,000 women have done mammography screening. Currently, the most efficient way to detect breast cancer is using mammography screening. In this thesis, we designed several convolution neural networks to learn the signatures of microcalcification and mass from images, respectively. By using this method, the scientists do not have to design image features manually from scratch, which is much different from traditional approaches. This thesis used the information of microcalcification and mass locations provided by the INbreast database and selected proper sizes of sub-images to proceed model training. After integrating models of using sub-images, the final predictor can be used on full images to find suspicious areas. The result showed that the area under the ROC curve (AUC) reached 0.98 and 0.89 for mass and microcalcification, respectively, revealing that the convolution neural network is capable of recognizing the differences between mass, microcalcification and normal function tissues when given sub-images directly. Ultimately, we performed screening by applying the well-trained model on the full images. After choosing proper hyper-parameters, the sensitivity of mass detection reached 80%, under a satisfied false positives per image on average, 0.80. However, the performance of microcalcification screening is not good due to too much noises when attempting to distinguish single microcalcifications from other materials. On the other hand, this thesis still demonstrated that convolution neural networks might be able to detect microcalcifications when they are clustered. We believe that it is worth more image training, in order to improve the ability to determine the suspicious microcalcification areas with the assistance of computers in the near future.