YOLOv12DETRADY: A Transformer-Enhanced Hybrid Detector for Thoracic Disease Localization in Chest X-rays

Abstract

We present YOLOv12DETRADY, a high-performance hybrid object detection architecture that sets a new state of the art for multi-abnormality localization in chest radiographs. By synergistically integrating the multi-scale efficiency of YOLOv12 with the global relational reasoning of a transformer-based detection head, our model achieves an mAP@0.5 of 0.801 and mAP@[0.5:0.95] of 0.521 on the challenging VinBigData benchmark—surpassing all published detector variants. Learned with a highly regulated training regime that includes radiologist-grounded annotations, DICOM-aware preprocessing, intensity-balanced histogram equalization, and domain-specific extensions, the architecture convinces to perform well on 14 clinically significant thoracic diseases. Based on AdamW optimization, cosine-annaled learning rates, and mixed-precision training distributed, YOLOv12DETRADY provides pixel-precise localization as well as real-time inference without non-maximum suppression and handcrafted anchors.