Comprehensive Summary
This study presents an osteoporosis screening model. The model addresses two challenges when constructing a model. The first issue is related to the difficulty of selecting representative slices of computed tomography (CT) images. The researchers propose a deep reinforcement learning-based image selection framework (DRLIS) to pick the best image slice. The second issue is that samples lacking complete data cannot be directly used in multi-modal fusion. The researchers therefore develop a knowledge distillation assisted multimodal model (KDAMM) that allows the use of sample with incomplete modalities. The model achieved an accuracy of 88.65% and an Area under the Curve (AUC) of 0.9532. This outperforms existing models by 2.85% in accuracy and 0.0212 in AUC.
Outcomes and Implications
Osteoporosis is common in elderly individuals, and is one of the main causes of disability and death. Due to the asymptomatic nature of the disease, high risk populations often find out that they are compromised after suffering an accident. There is, therefore, a demand for an osteoporosis screening model. This study presents a model that has superior performance against preexisting models. Nevertheless, there are still advancements to be made such as addressing imbalance in convergence speeds, generating modalities that introduce less noise or segmenting large models to improve performance.