Accessible cartilage tumor malignancy prediction via vision-language foundation model adaptation

Area of Science:

• Musculoskeletal Oncology
• Artificial Intelligence in Radiology
• Medical Imaging Analysis

Background:

• Cartilage tumors, such as enchondromas and chondrosarcomas, require accurate malignancy assessment for appropriate treatment.
• Distinguishing between benign enchondromas and malignant chondrosarcomas can be challenging based on imaging alone.
• Integrating non-imaging data with radiographic analysis may improve diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate a vision-language foundation model for predicting cartilage tumor malignancy.
• To assess the model's performance using radiographic images combined with demographic information.
• To determine the model's utility as a non-invasive, cost-efficient, and scalable diagnostic tool.

Main Methods:

• A dataset of 3336 radiographs from 455 patients with enchondroma or chondrosarcoma was curated.
• An adapted vision-language foundation model, based on CLIP, was fine-tuned with Medical Knowledge Adapters.
• The model was evaluated using 10-fold patient-level cross-validation, incorporating plain radiographs and demographic data.

Main Results:

• The model achieved an Area Under the ROC Curve (AUC) of 0.91 using radiographs alone.
• Incorporating demographic information improved the AUC to 0.94.
• The model demonstrated robust performance across tumor grades, including distinguishing atypical cartilaginous tumors (AUC 0.91) and high-grade chondrosarcomas (AUC 0.95).

Conclusions:

• The vision-language foundation model shows strong performance in predicting cartilage tumor malignancy.
• Combining radiographic images with demographic data enhances diagnostic accuracy.
• This approach offers a promising non-invasive, cost-effective, and scalable solution for cartilage tumor assessment in clinical practice.