Robust Large Scale Prone-Supine Polyp Matching Using Local Features: A Metric Learning Approach

Abstract

Abstract. Computer aided detection (CAD) systems have emerged as noninvasive and effective tools, using 3D CT Colonography (CTC) for early detection of colonic polyps. In this paper, we propose a robust and automatic polyp prone-supine view matching method, to facilitate the regular CTC workflow where radiologists need to manually match the CAD findings in prone and supine CT scans for validation. Apart from previous colon registration approaches based on global geometric information [1][2][3][4], this paper presents a feature selection and metric distance learning approach to build a pairwise matching function (where true pairs of polyp detections have smaller distances than false pairs), learned using local polyp classification features [5][6][7]. Thus our process can seamlessly handle collapsed colon segments or other severe structural artifacts which often exist in CTC, since only local features are used, whereas other global geometry dependent methods may become invalid for collapsed segmentation cases. Our automatic approach is extensively evaluated using a large multi-site dataset of 195 patient cases in training and 223 cases for testing. No external examination on the correctness of colon segmentation topology [2] is needed. The results show that we achieve significantly superior matching accuracy than previous methods [1][2][3][4], on at least one order-of-magnitude larger CTC datasets.

Toward embedded detection of polyps in WCE images for early diagnosis of colorectal cancer

“…Several previous references have considered the detection of intestinal polyps in videocolonoscopy images in the last few years ( [8][9][10][11][12] among recent ones). They are mainly divided into two categories: those based on geometric features of the polyps (size and shape) and those based on textural features.…”

Regression and classification based distance metric learning for medical image retrieval

“…The triplet-based learning methods [6,7] are probably the most popular now, with their explicit modeling of the similar-dissimilar triplet relationships between data samples. Variations of such methods have also been proposed for medical images recently [8,9,10]. A common problem with these triplet-learning approaches is, however, the large number of training samples that is proportional to the possible combinations of all triplets, and the resultant long training time.…”