在多GPU架構下以深度學習方法篩檢肺腫瘤及確認其良惡性

Abstract

本論文使用機器學習的方法來定位肺結節的位置，以及判斷該肺結節是否已發生病變。本文提出了一種U-net的變型方法，用於肺結節的分割和檢測。此外，本文通過使用肺結節分割的預測提供額外的信息，探索了肺結節檢測的二元分類任務。最後，對於分段任務，變型u-net的dice coefficient可達到0.8182；對於二進制分類任務，其錯誤率僅為4.22%。

In this thesis, lung nodule segmentation and lung nodule classification have been performed using database Image Database Consortium and Image Database Resource Initiative (LIDC). A variant U-net is proposed in this article for lung nodule segmentation and detection. Also, the binary classification task of lung nodule detection has been by using the predict of lung nodule segmentation as extra information. Finally, a variant u-net a dice coefficient of 0.8182 for segmentation task and an error rate of 4.22% for binary classification task have been accomplished.