Fracture Morphology Classification: Local Multiclass Modeling for Multilabel Complexity

Between $15,%$ and $45,%$ of children experience a fracture during their growth years, making accurate diagnosis essential. Fracture morphology, alongside location and fragment angle, is a key diagnostic feature. In this work, we propose a method to extract fracture morphology by assigning automatically global AO codes to corresponding fracture bounding boxes. This approach enables the use of public datasets and reformulates the global multilabel task into a local multiclass one, improving the average F1 score by $7.89,%$. However, performance declines when using imperfect fracture detectors, highlighting challenges for real-world deployment. Our code is available on GitHub.

💡 Research Summary

This paper presents a novel deep learning methodology for automating the classification of fracture morphology, a critical diagnostic feature in pediatric wrist trauma. Recognizing that 15-45% of children experience fractures, the authors address the challenge of using the standardized but complex AO/OTA classification system within a computational framework.

The core innovation lies in reformulating the problem from a “global multilabel” classification task—where a single whole X-ray image is analyzed for the potential presence of multiple fracture types—into a “local multiclass” task. The authors achieve this by developing a sophisticated pipeline that maps globally provided AO codes from the public GRAZPEDWRI-DX dataset to individual fracture bounding boxes. This pipeline utilizes bone segmentation masks (for radius, ulna, and epiphyses) to assign each bounding box to a specific bone. By simultaneously mapping the global AO codes to bone classes, a one-to-one match is established, allowing the extraction of a single morphology label (e.g., Transverse, Torus, Avulsion) for each localized fracture patch. Five morphology classes with sufficient samples were used for experiments.

The study compares three main approaches: 1) A lower baseline using a ResNet18 for multilabel classification on the entire image. 2) An upper baseline (“multi-class on GT BBox”) where the proposed pipeline provides perfectly localized fracture patches with their morphology labels, training a multiclass classifier on these patches. 3) A practical scenario (“YOLO BBox”) where a YOLOv10x detector first identifies fracture regions, and the same multiclass classifier is applied to these detected patches. The impact of different YOLO confidence thresholds and a false-positive reduction technique (adding a “Healthy” class) was evaluated.

Key results demonstrate the promise of the local reformulation. When using ground-truth bounding boxes, the patch-based multiclass approach achieved the highest F1-score (0.7630), representing a 7.89% average improvement over the whole-image multilabel baseline, with the most significant gain for the “Avulsion” class. This confirms the conceptual advantage of simplifying the task. However, performance drastically declined when relying on the YOLO detector. While lowering the detection confidence threshold increased recall and slightly improved F1-scores, metrics remained substantially lower (best F1 of 0.2786 at confidence 0.01). The false-positive reduction strategy failed to improve performance, likely because it added complexity to an already difficult classification task.

The discussion concludes that the local multiclass modeling strategy holds significant potential, as evidenced by the strong results with perfect localization. However, its real-world applicability is currently severely limited by the performance of automated fracture detection systems. The YOLO model’s frequent misses (false negatives) heavily penalized evaluation metrics. Therefore, future work must prioritize the development of more robust and accurate fracture detection algorithms as a prerequisite for a viable clinical application. Additionally, expanding the methodology to a more diverse dataset encompassing a wider range of AO codes and fracture morphologies is suggested for better generalization.