2017/10/30 by Faisal Mahmood, Mahmood, Faisal, Nicholas J. Durr +1 · 3 citations
Computer Science · Medicine · #Advanced Image and Video Retrieval Techniques #Colorectal Cancer Screening and Detection #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Image Retrieval and Classification Techniques
paper · pdf · doi:10.48550/arxiv.1710.11216
openalex publication_date 2017/10/30 · openalex created_date 2022/10/06 · openalex updated_date 2026/07/28
Colorectal cancer is the fourth leading cause of cancer deaths worldwide and\nthe second leading cause in the United States. The risk of colorectal cancer\ncan be mitigated by the identification and removal of premalignant lesions\nthrough optical colonoscopy. Unfortunately, conventional colonoscopy misses\nmore than 20% of the polyps that should be removed, due in part to poor\ncontrast of lesion topography. Imaging tissue topography during a colonoscopy\nis difficult because of the size constraints of the endoscope and the deforming\nmucosa. Most existing methods make geometric assumptions or incorporate a\npriori information, which limits accuracy and sensitivity. In this paper, we\npresent a method that avoids these restrictions, using a joint deep\nconvolutional neural network-conditional random field (CNN-CRF) framework.\nEstimated depth is used to reconstruct the topography of the surface of the\ncolon from a single image. We train the unary and pairwise potential functions\nof a CRF in a CNN on synthetic data, generated by developing an endoscope\ncamera model and rendering over 100,000 images of an anatomically-realistic\ncolon. We validate our approach with real endoscopy images from a porcine\ncolon, transferred to a synthetic-like domain, with ground truth from\nregistered computed tomography measurements. The CNN-CRF approach estimates\ndepths with a relative error of 0.152 for synthetic endoscopy images and 0.242\nfor real endoscopy images. We show that the estimated depth maps can be used\nfor reconstructing the topography of the mucosa from conventional colonoscopy\nimages. This approach can easily be integrated into existing endoscopy systems\nand provides a foundation for improving computer-aided detection algorithms for\ndetection, segmentation and classification of lesions.\n