使用深度卷積類神經網路於乳房超音波癌症轉移預測

Abstract

乳癌是最常被診斷出來的癌症，同時也是女性的第二大死因（繼肺癌之後）。儘管乳癌的死亡率很高，患有乳癌的患者卻可以透過早期診斷與適當的治療達到康復。然而，沒有接受治療的話乳癌細胞則可能透過淋巴腺或血管擴散到並侵入其他的器官，成為轉移性乳癌。因此，本文提出以深度卷積類神經網路為基礎，並藉由透過腫瘤切割決定腫瘤區域以及利用影像消光法所取得不同厚度的腫瘤周圍組織影像，開發電腦輔助系統進行轉移性乳癌的預測。本研究取了6種不同像素(5、10、15、20、25、到30)提取腫瘤周圍組織的影像進行系統訓練，其結果發現由腫瘤以及厚度15像素的周圍組織取得特徵進行訓練與測試系統，可以達到最佳的表現，其準確率為84.8%、靈敏度為88.8%、專一度為81.3%、以及ROC的曲線面積為0.926。此外，厚度為10像素的結果也不遜色，準確度達84.1%、靈敏度為83.2%、專一度為84.8%、曲線下面積為0.906，因此將來也可納入參考。

Breast cancer is the most commonly diagnosed cancer and is the second leading cause of death among women after lung cancer. Patients who suffer from breast cancer can still recuperate with early diagnosis and proper treatment, despite the fact that breast cancer possesses a high mortality rate. If untreated, breast cancer can surreptitiously spread and invade other organs by transporting its cells via nearby hematogeneous or lymphatic routes. This type of breast cancer is classified as metastatic breast cancer. Whereas, the non-metastatic cancer does not possess this ability. This thesis presents a breast cancer classification between the metastasis and non-metastasis using the densely connected convolutional neural network (DenseNet). Several studies have also successfully revealed the presence of suspicious tissue surrounding the tumor region (peritumor). Inspired by a previous study that utilized image matting to obtain the peritumor from the trimap, the peritumor can further be extracted at different pixel thicknesses by adjusting the unknown region thickness of the trimap. Thus, this study trained the neural network using peritumor images (instead of the tumor only) at different pixel thicknesses: 5, 10, 15, 20, 25, and 30 pixels. This study finds that the peritumor 15 pixels achieved the best performance with an accuracy of 84.8%, a sensitivity of 88.8%, a specificity of 81.3%, and an area under the curve (AUC) of receiver operating characteristic (ROC) of 0.926. In addition, based on the results, the peritumor 10 pixels may not be too farfetched from being considered for future studies as it scored an accuracy of 84.1%, a sensitivity of 83.2%, a specificity of 84.8%, and an AUC score of 0.906.