A Dual-Selective Channel Attention Network for Osteoporosis Prediction in Computed Tomography Images of Lumbar Spine

Abstract

Osteoporosis is a common systemic bone disease with insidious onset and low treatment efficiency. Once it occurs, it will increase bone fragility and lead to fractures. Computed tomography (CT) is a non-invasive medical examination method that can identify the bone condition of patients. In this paper, we propose a novel channel attention module, which is subsequently integrated into the supervised deep convolutional neural network (DCNN) termed DSNet, which can perform feature fusion from two different scales, and use the method of quadratic weight calculation to enhance the interconnection among feature map channels and improve the detection and classification performance for the bone condition in lumbar spine CT images. To train and test the proposed framework, we retrospectively collect 4805 CT images of 133 patients, using DXA as the gold standard. According to the T-value diagnostic criteria defined by WHO, the vertebral bodies of L1 - L4 in CT images are labeled and classified into osteoporosis, osteopenia and normal bone mineral density. Meanwhile, the training set and test set are constructed in the ratio of 4:1. As a result, the DSNet achieves a prediction accuracy of 83.4% and a recall rate of 90.0% on the test set, indicating that the proposed model has the potential to assist clinicians in diagnosing individuals with abnormal BMD and may alert patients at high risk of osteoporosis for timely treatment.