Shapley-based saliency maps improve interpretability of vertebral compression fractures classification: multicenter...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Fracture Classification
• Radiology and Diagnostic Imaging

Background:

• Vertebral compression fractures (VCFs) are a significant clinical concern.
• Accurate classification of acute versus chronic VCFs is crucial for treatment planning.
• Traditional deep learning models like Convolutional Neural Networks (CNNs) have limitations in interpretability.

Purpose of the Study:

• To evaluate the classification performance of the Vision Transformer (ViT) model for acute and chronic VCFs.
• To assess the interpretability of ViT using Shapley significance maps.
• To compare ViT performance against established models like ResNet18.

Main Methods:

• Retrospective analysis of 942 patients with VCFs, utilizing X-ray, CT, and MRI data.
• Fine-tuning of the SimpleViT model on a training dataset (7:2:1 split for train/validation/test).
• Performance evaluation using ROC curves, sensitivity, specificity, and AUC, with statistical significance assessed by the DeLong test.

Main Results:

• ViT achieved superior performance with an accuracy of 0.880 and AUC of 0.901, outperforming ResNet18 (0.843 accuracy, 0.833 AUC).
• Shapley significance maps significantly improved diagnostic sensitivity (0.883) and specificity (0.950).
• Radiologists favored ViT's Shapley-based saliency maps over GradCAM for interpretability.

Conclusions:

• Vision Transformer demonstrates enhanced performance in classifying vertebral compression fractures.
• Shapley significance maps offer superior interpretability compared to GradCAM, aiding clinical decision-making.
• ViT represents a promising advancement over CNNs for VCF diagnosis.