Can Machine Learning Models Based on Radiomic and Clinical Information Improve Radiologists' Diagnostic Performance...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Oncology
• Radiology
• Machine Learning in Healthcare

Background:

• Accurate preoperative diagnosis of bone tumors is crucial for effective treatment planning.
• Traditional diagnostic methods rely heavily on physician interpretation of radiographic and clinical data.
• Machine learning (ML) offers potential to augment diagnostic capabilities in medical imaging.

Purpose of the Study:

• To evaluate the efficacy of ML models in improving the diagnostic performance of physicians for bone tumors.
• To assess the utility of interpretable ML models as aids for clinical decision-making in bone tumor diagnosis.

Main Methods:

• Retrospective collection of radiographic and clinical data from bone tumor patients.
• Development and validation of multiple ML models, with extreme gradient boosting (XGBoost) selected based on Area Under the Curve (AUC).
• Two reading experiments involving seven physicians were conducted to compare independent reading versus ML-assisted reading using an independent test set.

Main Results:

• The XGBoost model, incorporating clinical information and radiomics features, achieved the highest classification performance (AUC = 0.905).
• Interpretable algorithms identified Gray Level Co-occurrence Matrix (GLCM) features as highly predictive.
• ML-assisted reading significantly improved physician diagnostic performance, with a mean AUC increase of 0.037 (P=0.047) compared to independent reading.

Conclusions:

• ML models integrating radiomics and clinical data from knee X-ray images can effectively assist clinicians.
• These models aid in the preoperative diagnosis of benign and malignant bone tumors, enhancing diagnostic accuracy.
• The study demonstrates the value of AI as a supportive tool for radiologists in bone tumor characterization.