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

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