SMOD: An Accurate and Efficient Segmentation-Based Medical Object Detector

Medical object detection simultaneously locates and recognizes the objects/lesions in medical images, and thus it is of high clinical relevance. There have been many kinds of detectors used in medical object detection tasks, such as transformer-based detectors. However, to the best of our knowledge, segmentation-based detectors for medical images have not been exploited yet. In this paper, a novel Segmentation-based Medical Object Detector (SMOD) with our newly designed
Seg2Det loss function
and
Non-connected Region Merging
(NRM) method is proposed, which provides: i) fast inference speed of detection due to the efficient segmentation-based detection framework without processing on anchors; and ii) high detection accuracy as our novel Seg2Det loss can effectively replace detection by segmentation of bounding boxes and our NRM method can enclose segmentation predictions with proper boxes. Our SMOD has been evaluated over two common medical object detection tasks and compared with both anchor-based and anchor-free state-of-the-art (SOTA) detectors. In experiments, our SMOD outperforms all compared detectors and achieves the fastest inference speed. Specifically, our SMOD increases the overall mean Average Precision (mAP) score by at least 2.38% and 3.55% for two pneumonia detection datasets, respectively; for the detection of gastrointestinal polyps, our SMOD also achieves at least a 2.59% increase on the mAP score; in terms of inference speed, our SMOD can process more than 80 images per second on all compared datasets; and in visualization results, our SMOD achieves satisfactory performance in detecting lesions from images with low-contrast lesions, blurred object boundaries, small lesions, and multiple lesions.