الوصف:Bone tumor diagnosis from X-ray images remains challenging due to subtle radiographic
differences between benign and malignant lesions, limited access to specialized radiologists,
and severe class imbalance in medical datasets. This research developed and validated a
comprehensive three-stage deep learning system for automated bone tumor diagnosis using
the publicly available BTXRD dataset (3,746 images).
Stage 1 employed EfficientNet-B5 integrated with 16 clinical features (age, gender,
anatomical location, projection) for initial triage into normal, benign, or malignant categories.
Stage 2 utilized YOLOv8x object detection at 800×800 resolution to localize tumor regions.
Stage 3 applied EfficientNetV2-M on YOLO-cropped regions with 23 clinical features to
classify tumors into six specific types. We addressed severe class imbalance (137:1 ratio)
through intelligent grouping of rare classes, focal loss with class weights, balanced
oversampling, and label smoothing (ε=0.05).
Results: Stage 1 achieved 94.84% accuracy and AUC 0.989 on 562 test images, with per-
class F1-scores of 0.98 (normal), 0.94 (benign), and 0.79 (malignant). Critically, 74.51%
malignant recall with zero cases misclassified as normal ensured no cancers were
completely missed. Stage 2 optimization improved detection from 51.14%
(YOLOv8m@640) to 71.88%
[email protected] (YOLOv8x@800), with 79.79% malignant
detection. Bounding box detection substantially outperformed segmentation (71.88% vs
11.78%), validating our detection-only approach. Stage 3 achieved 77.94% accuracy and
0.732 macro F1-score on 281 test cases, with YOLO enhancement improving accuracy by
+4.2 percentage points over full-image processing. Per-class F1-scores ranged from 0.875
(multiple osteochondromas) to 0.500 (giant cell tumor). Complete processing required 1.2
seconds per case, enabling real-time clinical deployment.
This work represents the first comprehensive three-stage system integrating screening,
localization, and classification for bone tumors, demonstrating that YOLO-based localization
significantly improves classification, detection outperforms segmentation for downstream
tasks, and multi-modal fusion enhances diagnostic performance. The system achieved
clinically relevant accuracy while handling extreme class imbalance on the largest public
bone tumor dataset available.