Improving CT-image universal lesion detection with comprehensive data and feature enhancements

As a crucial task in Computer Vision, object detection has substantially improved in recent years, with the aid of deep learning and increasingly abundant datasets. However, compared with natural image detection, medical CT images require more precision due to the obvious clinical implications. Detecting multiple lesions or clusters with relatively few training samples and indistinctive feature representation is extremely problematic. In this paper, we propose comprehensive improvements to the original YOLOv3, such as data augmentation, feature attention enhancement and feature complementarity enhancement to increase general lesion area detection performance. Ablation studies use the open DeepLesion dataset to validate these improvements and confirm the effectiveness of each amendment. Comparisons between state-of-the-art counterparts demonstrated that the proposed lesion object detector has enhanced salient accuracy (under two commonly used metrics) and an exceptional speed-accuracy trade-off. The proposed model achieved 57.5% mAP and 85.07% sensitivity at 4 false positives (FPs) per image, while running at reliable 35.6 frames per second (FPS). These findings indicate that the proposed detector is more practicable than other currently available computer aided diagnostics (CAD)