Bone Anomaly Identification and Localization from Musculoskeletal X-ray images using Hybrid DenseNet-Vision Transformers.

DIVYA MANOHARLAL BHATIA, Aparna Kanakatte Gurumurthy, RUPSHA MUKHERJEE, Avik Ghose, Murali Poduval

Keywords: Musculoskeletal anomalies, Fracture, Arthritis, Visual transformers

Abstract: Musculoskeletal anomalies present a global health challenge, often requiring expert review for diagnosis. Automated detection systems can alleviate this burden and facilitate timely intervention. We introduce a hybrid framework combining DenseNet and Vision Transformers (ViT) for detecting anomalies in the MURA dataset. This model uses CNNs for local feature extraction and ViT for global context modeling, effectively capturing the subtle features of bone pathologies. Our binary model achieved 92% accuracy across seven bone types, demonstrating strong generalizability through cross-anatomy testing. To enhance fine-grained annotations, we collaborated with an orthopedic expert to label 1,379 finger radiographs into three sub-anomaly categories: arthritis, fracture, and implant, with bounding boxes. This framework reached 96% classification accuracy and 91% localization efficiency (AUC), marking the first instance of sub-anomaly classification on MURA finger data with bounding box predictions providing visual validation for clinicians. Additionally, we also analyze the importance of positional encodings in ViT, showing their necessity for localization but not for binary classification. This work offers a new annotated dataset and a robust framework, advancing automated orthopedic anomaly detection.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Application: Other

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 227