Correcting misalignment of automatic 3D detection by classification: ileo-cecal valve false positive reduction in CT colonography

• Authors: Le Lu;Matthias Wolf;Jinbo Bi;Marcos Salganicoff
• Venue: MCV'10 Proceedings of the 2010 international MICCAI conference on Medical computer vision: recognition techniques and applications in medical imaging
• Year: 2010

Ileo-Cecal Valve (ICV) is an important small soft organ which appears in human abdomen CT scans and connects colon and small intestine. Automated detection of ICV is of great clinical value for removing false positive (FP) findings in computer aided diagnosis (CAD) of colon cancers using CT colongraphy (CTC) [1,2,3]. However full 3D object detection, especially for small objects with large shape and pose variations as ICV, is very challenging. The final spatial detection accuracy often trades for robustness to find instances under variable conditions [4]. In this paper, we describe two significant post-parsing processes after the normal procedure of object (e.g., ICV) detection [4], to probabilistically interpret multiple hypotheses detections. It achieves nearly 300% performance improvement on (polyp detection) FP removal rate of [4], with about 1% extra computional overhead. First, a new multiple detection spatial-fusion method utilizes the initial single detection as an anchor identity and iteratively integrates other "trustful" detections by maximizing their spatial gains (if included) in a linkage. The ICV detection output is thus a set of N spatially connected boxes instead of a single box as top candidate, which allows to correct 3D detection misalignment inaccuracy. Next, we infer the spatial relationship between CAD generated polyp candidates and the detected ICV bounding boxes in 3D volume, and convert as a set of continuous valued, ICV-association features per candidate which allows further statistical analysis and classification for more rigorous false positive deduction in colon CAD. Based on our annotated 116 training cases, the spatial coverage ratio between the new N-box ICV detection and annotation is improved by 13.0% (N=2) and 19.6% (N=3) respectively. An evaluation on large scale datasets of total ∼ 1400 CTC volumes, with different tagging preparations, reports average 5.1 FP candidates are removed at Candidate-Generation stage per scan; and the final CAD system mean FP rate drops from 2.2 to 1.82 per volume, without affecting the sensitivity.