Recommender-based bone tumour classification with radiographs-a link to the past

Area of Science:

• Orthopaedic Surgery
• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Accurate bone tumor classification is crucial for early diagnosis and treatment planning.
• Linking undiagnosed patients to similar historical cases can improve diagnostic accuracy.
• Current methods for bone tumor classification from radiographs have limitations.

Purpose of the Study:

• To develop and evaluate an algorithm for linking patients with undiagnosed bone tumors to previous radiographic histories.
• To achieve simultaneous classification of multiple bone tumor entities using deep learning.
• To enable early and specific diagnosis of bone tumors through an automated system.

Main Methods:

• A retrospective study curated data from 2000-2021, including 809 patients and 1792 radiographs.
• A ResNet and transformer model were implemented, extracting image features via deep learning.
• A hash-based nearest-neighbor recommender approach clustered similar images for classification by majority voting.

Main Results:

• The model achieved high performance metrics: 92.86% accuracy, 92.86% precision, and 34.08% recall (k=3).
• This significantly outperformed state-of-the-art models, which achieved lower scores.
• The dataset included ten common primary bone tumor entities, with Osteochondroma being the most frequent.

Conclusions:

• The novel algorithm effectively classifies bone tumors from radiographs and links to prior patient data.
• This approach surpasses current models in diagnostic accuracy and efficiency.
• The tool can support clinicians, especially those with limited experience, in bone tumor diagnosis.