The knowledge distillation-assisted multimodal model for osteoporosis screening.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Osteoporosis Research

Background:

• Osteoporosis screening rates before fractures are low despite high prevalence and severe consequences.
• Challenges in osteoporosis screening include selecting representative CT slices and handling incomplete multimodal data.

Purpose of the Study:

• To develop an effective osteoporosis screening model using a multimodal approach.
• To address challenges in CT image slice selection and incomplete data utilization.

Main Methods:

• A reinforcement learning-driven knowledge distillation-assisted multimodal model was developed.
• The model integrates demographic data, laboratory indicators, and CT images.
• Novel components include a deep reinforcement learning-based image selection module (DRLIS) and a knowledge distillation-assisted multimodal model (KDAMM).

Main Results:

• The multimodal model achieved 88.65% accuracy and 0.9542 AUC, outperforming existing models.
• Effectiveness of individual framework components was demonstrated.
• SHAP values were used to assess feature importance.

Conclusions:

• A knowledge distillation-assisted multimodal model for opportunistic osteoporosis screening was presented.
• The model combines demographic, laboratory, and CT image data.
• The framework achieved state-of-the-art performance on self-collected datasets.