Comparative evaluation of explainable vision models for bone tumour detection on radiographs

Abstract

Accurate detection of bone tumours on radiographs can be challenging because lesion appearance varies and interpretation requires specialist expertise. This study evaluates deep learning (DL) models for bone tumour localisation, segmentation, classification, and explainability using the publicly available Bone Tumour X-Ray Radiograph Dataset (BTXRD), comprising 3,746 musculoskeletal radiographs. A YOLOv8s-seg model was evaluated alongside MobileNetV3-Large, EfficientNet-B3, and ConvNeXt-Tiny classifiers trained under a common-resolution protocol. Grad-CAM, Grad-CAM++, and Layer-CAM activation maps were compared with expert tumour masks using mean Intersection over Union and a thresholded lesion-overlap measure. For the retrospective agreement analysis, EfficientNet-B3 was paired with YOLOv8s-seg. An exploratory external series of 11 tumour-positive radiographs was also examined. ConvNeXt-Tiny achieved the highest binary classification accuracy (0.879), while YOLOv8s-seg achieved an overall box mAP50 of 0.676 and mask mAP50 of 0.668. The components produced concordant binary tumour-presence predictions for 605 of 750 test radiographs, corresponding to 80.7% coverage (95% confidence interval [CI]: 77.7%–83.3%). Binary accuracy within the agreed subset was 93.6% (566/605; 95% CI: 91.3%–95.2%), while the remaining 145 cases were discordant and received no consensus output. The XAI analyses quantify spatial correspondence rather than explanation faithfulness or clinical reliability. Because the exploratory external series contained seven benign and four malignant tumour-positive radiographs and no normal radiographs, it cannot assess specificity or false-positive behaviour and is not presented as representative external validation.