HAAF-RT-DETR: hybrid attention and high-resolution feature fusion based RT-DETR for kidney stone detection

Accurate localization and detection of kidney stones in computed tomography (CT) images are important for computer-aided medical image analysis. However, accurate detection remains challenging due to the small size of stones, weak boundaries, and complex renal anatomical backgrounds. To address these issues, this work proposes HAAF-RT-DETR (Hybrid Attention and High-Resolution Feature Fusion Based RT-DETR) for kidney stone detection. HAAF-RT-DETR integrates hybrid attention mechanisms and high-resolution feature fusion strategies into the RT-DETR framework to enhance feature representation and improve small-object detection performance. Within this framework, the HGBlock-AssemFormer module is introduced to jointly model local convolutional feature aggregation and Transformer-based global dependencies, thereby strengthening high-level semantic representations. In addition, a Convolution–Attention Fusion Module (CAFM) is employed to further refine high-level feature representations and reduce background interference. Furthermore, a high-resolution P2 feature level is incorporated, extending the original three-scale pyramid (P3–P5) to a four-scale representation (P2–P5), enabling fine-grained spatial details to participate in the DETR-style unified decoding process. Experiments on stone-positive CT slices from TEZ_ROI_AUG and the Kidney Stone Dataset demonstrate improvements in mAP, Recall, and F1-score over the evaluated YOLO-based detectors and RT-DETR baseline.