Conditional Training with Bounding Map for Universal Lesion Detection

Universal Lesion Detection (ULD) in computed tomography plays an essential role computer-aided diagnosis. Promising ULD results have been reported by coarse-to-fine two-stage detection approaches, but such methods still suffer from issues like imbalance of positive v.s. negative anchors during object proposal and insufficient supervision problem localization regression classification the region interest (RoI) proposals. While leveraging pseudo segmentation masks as bounding map (BM) can reduce above to some degree, it is open effectively handle diverse lesion shapes sizes ULD. In this paper we propose a BM-based conditional training for ULD, which (i) vs. anchor via conditioning (BMC) mechanism sampling instead traditional IoU-based rule; (ii) adaptively compute size-adaptive BM (ABM) bounding-box, used improving accuracy ABM-supervised segmentation. Experiments with four state-of-the-art show that proposed approach bring almost free improvement without requiring expensive mask annotations.