Transformer-powered precision: A DETR-based approach for robust detection in medical ultrasound with cholelithiasis as a case study.

Background and objective: Transformers have demonstrated strong capabilities in capturing long-range dependencies in visual data, but their application to noisy, low-contrast medical imaging such as ultrasound remains limited. Cholelithiasis detection, often hindered by the inherent limitations of ultrasound imaging, requires more precise and robust computational approaches. This study introduces a detection architecture that integrates convolutional feature extraction with a customized Detection Transformer (DETR) to improve localization and detection in challenging ultrasound conditions. Methods: The proposed method combines convolutional inductive biases with transformer-based self-attention to capture both local and global spatial relationships. It was applied to the task of detecting gallstones and gallbladder regions in ultrasound images of cholelithiasis. The approach was benchmarked against state-of-the-art object detection models, including RT-DETR, YOLOv8, and YOLO-NAS. Radiologists validated the bounding boxes generated by the model to assess clinical reliability. Results: The custom DETR achieved confidence score up to 99 % in detecting gallstones and gallbladder regions, outperforming RT-DETR, YOLOv8, and YOLO-NAS. The method demonstrated mean average precision improvements of 13 % and 14 % compared to YOLOv8 and YOLO-NAS, respectively. Radiologist validation confirmed the clinical accuracy and robustness of the proposed detection framework. Conclusions: By effectively addressing the challenges of low-quality ultrasound imaging, the proposed DETR-based framework provides a reliable and generalizable approach for automated cholelithiasis detection. The findings highlight its potential for integration into real-world diagnostic workflows and its applicability to broader intelligent diagnostic systems at the intersection of computational science, medical informatics, and vision transformers.