Deep Learning Radiomics Model Based on Computed Tomography Image for Predicting the Classification of Osteoporotic...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Osteoporotic vertebral fractures (OVFs) are a significant cause of disability in older adults.
• Accurate diagnosis and classification of OVFs are crucial for effective management.
• Advancements in CT imaging, radiomics, and deep learning offer potential for improved OVF classification.

Purpose of the Study:

• To evaluate the efficacy of a deep learning radiomics model for accurate OVF classification using CT imaging.
• To compare the performance of models pretrained on RadImageNet versus ImageNet.
• To interpret the predictive rationale of the optimal model using SHAP analysis.

Main Methods:

• Analysis of 981 patients (1098 vertebrae) with OVFs classified using the ASTLOF system (Classes 0, 1, 2).
• Development of a deep transfer learning model (ResNet-50) combined with radiomics features, refined by LASSO regression.
• Performance assessment using machine learning classifiers, ROC metrics, and the "One-vs-Rest" approach; comparison via DeLong test; interpretation via SHAP analysis.

Main Results:

• The RadImageNet-based fused model achieved superior predictive performance across validation sets (macro-average AUCs: 0.837 internal, 0.773 external, 0.852 prospective).
• The model showed strong performance in classifying Class 2 fractures (AUC=0.907), followed by Class 0 (AUC=0.829) and Class 1 (AUC=0.794).
• SHAP analysis identified key features like cluster shade, mean, and large area low gray level emphasis influencing predictions.

Conclusions:

• The RadImageNet-based deep learning radiomics model demonstrates significant utility for OVF classification using CT data.
• This model shows superior predictive performance compared to the ImageNet-based model, aiding clinical decision-making.
• The model's ability to accurately classify different OVF classes supports its application in treatment planning.