Colorectal cancer (CRC) is one of the top leading causes of cancer deaths around the world. Although colonoscopy is an effective tool for CRC screening, two major difficulties faced by endoscopists are missed polyps and misclassified polyps. This proposal aims to develop a real-time computer-aided system (CAS) for polyp detection and polyp type classification during colonoscopy. Existing CAS only perform either polyp detection (i.e. with or without polyps) or polyp type classification (neoplastic or non-neoplastic). In this study, a system that can perform both tasks simultaneously will be studied, i.e. by formulating the problem as a three-class image classification task (non-polyp, non-neoplastic polyp and neoplastic polyp). A set of unstructured endoscopic images, with clean labels, will be collected. Advanced machine learning technologies, including deep learning and transfer learning techniques with a novel architecture, will be implemented and evaluated on the set of endoscopic images. It is anticipated that the proposed CAS will display the diagnostic results of each frame in real time during colonoscopy to aid decision making during CRC screening. 雖然腸鏡檢查是篩查大腸癌的有效方法，但內窺鏡醫生面臨兩大難題，即對篩查瘜肉的遺漏和對瘜肉病理的誤判。本項提案的目的是開發一套能在腸鏡檢查期間實時對瘜肉進行篩查和對病理類型分類的計算機輔助系統(CAS) 。現有的計算機輔助系統只能單一地執行瘜肉篩查(即有或沒有瘜肉)或瘜肉病理類型分類(贅生性或非贅生性)的任務。在本項研究中，開發的系統將通過執行一個三分類圖像分析任務( 非瘜肉，非腫瘤性瘜肉和腫瘤瘜肉)來同時完成篩查瘜肉和對其進行病理分析。本研究將收集一組有清晰病理信息標註的內窺鏡圖像。以新穎的結構結合的深度學習和遷移學習搭建一套計算機輔助系統并在收集的內窺鏡圖像中進行測試和評估。

2018

TSSSU Company