Exploiting biochemical data to improve osteosarcoma diagnosis with deep learning.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Early and accurate osteosarcoma (OS) diagnosis is crucial for patient outcomes.
• Current machine learning (ML) models for OS diagnosis primarily use X-ray images, often lacking generalization and explainability.

Purpose of the Study:

• To explore deep learning models for improved accuracy, explainability, and generality in primary OS diagnosis.
• To evaluate the added value of integrating biochemical data (alkaline phosphatase and lactate dehydrogenase) with imaging data.

Main Methods:

• A deep learning model was designed to incorporate numerical features of alkaline phosphatase (ALP) and lactate dehydrogenase (LDH) with visual features from X-ray imaging.
• A late fusion approach in the feature space was employed to combine diverse data types.
• The model was evaluated on a real-world dataset of 848 patients (2608 cases) aged 4 to 81.

Main Results:

• The integrated model achieved a diagnostic accuracy of 97.17%, a significant improvement over the baseline accuracy of 94.35%.
• Simultaneous incorporation of ALP and LDH via a late fusion approach proved effective.
• Grad-CAM visualizations demonstrated model explainability, aligning with orthopedic specialists' assessments.

Conclusions:

• Integrating biochemical markers (ALP, LDH) with X-ray imaging using a late fusion deep learning model enhances osteosarcoma diagnostic accuracy and explainability.
• This multimodal approach offers a more robust and generalizable solution for primary OS diagnosis compared to image-only methods.