Radiographic Classification of Bone Tumours Using Recommender Systems

Overview

This study presents a deep learning-based recommender system (RS) that classifies multiple bone tumour pathologies by recommending similar patient cases based on radiographic features. Using a curated dataset of 809 patients with 1792 radiographs, the model leverages historical clinical data to enhance early and specific diagnosis of bone tumours.

Background

Bone tumours are rare and diverse, with most primary tumours being benign and malignant tumours accounting for a small fraction of adult malignancies. Early diagnosis is critical for prognosis and treatment, but delays often occur due to nonspecific symptoms and limited experience among non-specialists. Radiographs remain the initial imaging modality recommended for screening, with further imaging and histopathology required for definitive diagnosis. Advances in artificial intelligence, particularly deep learning, offer promising tools to improve diagnostic accuracy in musculoskeletal oncology.

Data Highlights

Key Findings

• The RS model clusters radiographic features to recommend similar patient cases, aiding classification of ten frequent bone tumour entities.
• Malignant lesions were confirmed by histopathology; benign/intermediate lesions were verified by histopathology or tumour board consensus.
• The model achieved classification accuracy, precision, and recall metrics calculated on an external test subset comprising 10% of data from other institutions.
• Use of historical clinical and imaging data from hospital information systems enhances diagnostic support in musculoskeletal tumour centres.
• The RS approach bypasses extensive algorithm training by leveraging similarity-based recommendations, potentially improving early tumour identification.

Clinical Implications

Integrating recommender systems into clinical workflows can support non-specialist physicians by providing access to similar historical cases, potentially reducing diagnostic delays. Early and accurate radiographic classification of bone tumours facilitates timely referral to specialised centres and initiation of appropriate treatment. This approach may enhance precision medicine in musculoskeletal oncology by harnessing existing clinical data repositories.

Conclusion

The study demonstrates that a deep learning-based recommender system utilizing radiographic features and historical patient data can effectively classify multiple bone tumour types. This novel method holds promise for improving early diagnosis and clinical decision-making in musculoskeletal tumour care.

References

1. Musculoskeletal Tumor Society and American Academy of Orthopedic Surgeons -- Radiographic recommendations
2. CLAIM Checklist -- Artificial Intelligence in Medical Imaging