The knowledge distillation-assisted multimodal model for osteoporosis screening

Osteoporosis is characterized by reduced bone mass and deterioration of bone structure, yet screening rates prior to fractures remain low. Given its high prevalence and severe consequences, developing an effective osteoporosis screening model is highly significant. However, constructing these screening models presents two main challenges. First, selecting representative slices from CT image sequences is challenging, making it crucial to filter the most indicative slices. Second, samples lacking complete modal data cannot be directly used in multimodal fusion, resulting in underutilization of available data and limiting the performance of the multimodal osteoporosis screening model.

In this paper, we propose a reinforcement learning-driven knowledge distillation-assisted multimodal model for osteoporosis screening. The model integrates demographic characteristics, routine laboratory indicators, and CT images. Specifically, our framework includes two novel components: 1) a deep reinforcement learning-based image selection module (DRLIS) designed to select representative image slices from CT sequences; and 2) a knowledge distillation-assisted multimodal model (KDAMM) that transfers information from single-modal teacher networks to the multimodal model, effectively utilizing samples with incomplete modalities. The codes are published on: https://github.com/AImedcinesdu212/Osteoporosis-Predictionhttps://github.com/Hidden-neurosis/osreoporosis.git.

The proposed multimodal osteoporosis screening model achieves an accuracy of 88.65 % and an AUC of 0.9542, surpassing existing models by 2.85 % in accuracy and 0.0212 in AUC. Additionally, we demonstrate the effectiveness of each novelty within our framework. The SHAP values are calculated to assess the importance of demographic characteristics and routine laboratory test data.

This paper presents a knowledge distillation-assisted multimodal model for opportunistic osteoporosis screening. The model incorporates demographic characteristics, routine laboratory indicators (including blood tests and urinalysis), and CT images. Extensive experiments, conducted on self-collected datasets, validate that the proposed framework achieves state-of-the-art performance.