Abstract
Automatic lesion detection from computed tomography (CT) scans is animportant task in medical imaging analysis. It is still very challenging due tosimilar appearances (e.g. intensity and texture) between lesions and othertissues, making it especially difficult to develop a universal lesion detector.Instead of developing a specific-type lesion detector, this work builds aUniversal Lesion Detector (ULDor) based on Mask R-CNN, which is able to detectall different kinds of lesions from whole body parts. As a state-of-the-artobject detector, Mask R-CNN adds a branch for predicting segmentation masks oneach Region of Interest (RoI) to improve the detection performance. However, itis almost impossible to manually annotate a large-scale dataset withpixel-level lesion masks to train the Mask R-CNN for lesion detection. Toaddress this problem, this work constructs a pseudo mask for each lesion regionthat can be considered as a surrogate of the real mask, based on which the MaskR-CNN is employed for lesion detection. On the other hand, this work proposes ahard negative example mining strategy to reduce the false positives forimproving the detection performance. Experimental results on the NIH DeepLesiondataset demonstrate that the ULDor is enhanced using pseudo masks and theproposed hard negative example mining strategy and achieves a sensitivity of86.21% with five false positives per image.