Weakly supervised deep learning for diagnosis of multiple vertebral compression fractures in CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Vertebral compression fractures (VCFs) require precise classification for effective treatment planning.
• Current methods may necessitate detailed, vertebral-level labeling, increasing workload.
• Deep learning (DL) offers potential for automated fracture detection.

Purpose of the Study:

• To develop a weakly supervised DL model for vertebral-level VCF classification.
• To evaluate the model's performance using image-level labels.
• To compare the proposed model against a traditional supervised DL approach.

Main Methods:

• Trained a weakly supervised DL model on 815 patients with image-level VCF labels.
• Compared its performance to a supervised DL model trained with vertebral-level labels.
• Evaluated models on a test set of 227 patients, comparing sensitivities at similar specificities.

Main Results:

• The weakly supervised model demonstrated comparable or superior sensitivity to the supervised model for overall L1-L5 VCF classification (0.770 vs 0.705).
• Specificities remained high across all vertebrae (L1-L5 > 0.970).
• The proposed model showed significantly better sensitivity for L3 and comparable sensitivity for L1, L4, and L5.

Conclusions:

• Weakly supervised DL can achieve effective vertebral-level VCF classification using image-level data.
• This approach holds potential for reducing radiologist workload and improving diagnostic efficiency.
• The method can identify multiple VCFs even when trained on less granular labels.