Detection of Colorectal Polyps with Deep Neural Networks

10.29428/9789860544169.201801.0081

Wei-Ting Xiao；Wei-Min Liu

息肉偵測 ； 大腸鏡影像 ； 深度學習 ； Polyp detection ； Colonoscopy images ； Deep learning

NCS 2017 全國計算機會議

2017（2018 / 01 / 01）

425 - 428

繁體中文

近年來有許多在大腸鏡影像下偵測息肉的研究，但是相對較少能同時兼顧到息肉偵測要能即時輔助醫師的重要性，使得大部分成果僅應用在離線影像分析。在此議題上本研究使用了現在非常熱門的深度學習技術，其中架構為現有的SegNet。影像資料集使用MICCAI Sub-Challenge 2015針對Automatic polyp detection in colonoscopy videos 議題所提供的內視鏡影像CVC-ClinicDB。經過300張影像訓練，312張影像測試後發現Accuracy 有93.05%，Sensitivity 及Specificity 各得到61.58%及95.64%，平均一張處理時間為0.045秒，因此若應用在臨床大腸鏡檢查時，可達到即時輔助醫師顯示息肉位置的效果。

Recently, there are many studies focused on detection of polyps in colonoscopy images, but not all of them addressed the issue about providing the detection results for diagnosis in real time. In this work, we utilized the existing Segnet, a deep learning framework to deal with the problem. The database is from MICCAI sub-challenge 2015, and under the issue of automatic polyp detection in colonoscopy videos of colonoscopy image, called CVC-ClinicDB. After training with 300 images and testing with 312 images, we got a good performance: 93.05% accuracy, 61.58% sensitivity, and 95.64% specificity. The average process time is 0.045 second per image. Once the model is applied in clinical colonoscopy exams, the system should be able to provide effective second opinions to aid the diagnosis in real time.

1. 陳宣輯(2016)。視訊辨識技術應用於智慧型監控系統之研究。國立臺灣大學資訊工程學系學位論文。2016。1-163。