Answers are grounded in this paper and cite their sources.
Background Automatic localization and delineation of bone tumors on radiographs are challenging because lesion appearance, contrast, and anatomical background vary substantially. Methods We developed ADC-YOLO by integrating an AOD-Net-derived learnable enhancement block, DySample dynamic upsampling, and CBAM channel-spatial attention into YOLOv8m. The public BTXRD dataset contained 3, 746 radiographs. Images from sources 1 and 3 were divided into training (n = 2, 557) and validation (n = 640) sets, while source 2 was reserved as a source-held-out test set (n = 549). ADC-YOLO was compared with Mask R-CNN, YOLOv8m, YOLOv9c, and ASF-YOLO using precision, recall, mAP50, and mAP50:95 for bounding boxes and masks. Results On the validation set, ADC-YOLO achieved bounding-box mAP50/mAP50:95 of 0.772/0.482 and mask mAP50/mAP50:95 of 0.797/0.490. On the source-held-out test set, the corresponding values were 0.849/0.551 and 0.869/0.545. Compared with YOLOv8m, test-set bounding-box mAP50 and mAP50:95 increased by 6.6 and 6.3 percentage points, while mask mAP50 and mAP50:95 increased by 6.7 and 4.9 percentage points. Component effects were task dependent, but the complete three-module configuration achieved the highest values for all reported mAP measures. Conclusion ADC-YOLO achieved the best overall performance among the evaluated models for bone tumor detection and instance segmentation on radiographs, demonstrating its potential as an accurate unified framework for automated lesion localization and delineation.
No takes yet. Share an insight, caveat, or question.
Yao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: