Lesion Detection in CT Scans: A Comparative Study of Traditional, One-Two Stages, and Transformer-Based Models

Lesion detection from computed tomography (CT) scans is a crucial task for early diagnosis and appropriate treatment planning. The task is typically obstructed by the presence of incredibly small lesions, low contrast, and involved anatomical backgrounds. For this issue to be solved, some of the latest breakthroughs in deep learning have emerged with some object detection methods, including single-stage models, two-stage models, and transformer models. This paper presents a side-by-side evaluation of some of the best-performing detection models like YOLOv12n, YOLOv12l, RetinaNet, Faster R-CNN, and RT-DETR based on a well-curated subset of the NIH DeepLesion dataset. All methods were critically evaluated with respect to detection accuracy, time, and computational expense in controlled test settings. It was found that transformer-based RT-DETR performed best with an mAP@50 score of 0.891 while maintaining real-time performance. YOLOv12-Large also turned out to be a balanced trade-off between speed and accuracy and could practically be applied in clinical contexts, particularly with limited resources.