This study retrospectively collected clinical data of patients from the Second Affiliated Hospital of Wenzhou Medical University (Institution 1, Training set), Yueqing People’s Hospital (Institution 2, External test set 1), Wenzhou Hospital of Integrated Traditional Chinese and Western Medicine (Institution 3, External test set 2) and People’s Hospital of Cangnan (Institution 4, External test set 3) from January 1, 2022 to December 31, 2022, respectively. Moreover, all institutions are tertiary hospitals. The institutional review boards at all participating institutions granted approval for this retrospective multicohort study (approval number: 2023-K-179-01) and waived the requirement for written informed consent.

A total of 1126 participants from four institutions were included in the study. The inclusion criteria for this study were (1) age ≥ 50 years, (2) Unenhanced chest CT scan was performed, (3) DXA was performed, and (4) Complete medical records. The exclusion criteria for this study were (1) The interval between DXA and chest CT examination was > 3 months, (2) Artifacts on CT images, (3) Lack of medical records, and (4) Premenopausal female participants. Institution 1 registered 581 participants, comprising 284 with osteoporosis and 297 without osteoporosis. Institution 2 had 229 participants, with 92 having osteoporosis and 137 without osteoporosis. Institution 3 enrolled 198 participants, with 90 having osteoporosis and 108 without osteoporosis. Institution 4 included 118 participants, with 68 having osteoporosis and 50 without osteoporosis. The participants recruitment process for this study is shown in Fig. 1.

The recruitment process of participants in this study. CT, computed tomography; DXA, dual-energy X-ray absorptiometry; OP, osteoporosis

DXA is considered the gold standard for diagnosing osteoporosis through the measurement of BMD. For postmenopausal women and men (over 50 years old), referring to the diagnostic criteria is recommended. Patients with a t score of −2.5 or less on any lumbar spine (lumbar 1–4), femoral neck, or distal radius thirds measured by DXA were diagnosed as osteoporosis patients. Patients with a t score greater than –2.5 were diagnosed as non-osteoporosis patients.

Unenhanced chest CT images of 5 mm thickness of Institution 1, Institution 2, Institution 3, and Institution 4 were obtained from picture archiving and communication systems (PACS), respectively. In addition, all CT images were resampled to a voxel spacing of 1 × 1 × 1 mm3 in order to normalize the images from different institutions. At the same time, adjust the window level of all images to 250 and the window width to 800.

The Ethics Committee of the Second Affiliated Hospital of Wenzhou Medical University, the Ethics Committee of Yueqing People’s Hospital, Wenzhou Hospital of Integrated Traditional Chinese and Western Medicine, and the People’s Hospital of Cangnan approved this study (approval number: 2023-K-179-01; YQYY202400061; 2024-L047; 2023120). Informed consent was waived because the study was a retrospective cohort study.

According to previous studies, the threshold of skeletal muscle is −29HU to 150HU [9]. A junior clinician (5 + years of clinical experience) used 3D Slicer (version 5.2.2) software to map the thoracic vertebrae and its surrounding muscles, the region of interest (ROI), at the mid-level of the 12th thoracic vertebrae (T12). In addition, the ROI outlined by the junior clinician was rechecked by a senior clinician (10 + years of clinical work experience). All clinicians are blind to patient information when mapping and checking ROI.

This study used ImageJ (NIH ImageJ version 1.54f) to calculate the area of the bilateral psoas muscle in ROI. The obtained skeletal muscle area (cm2) is divided by the square of the patient’s height (m2) to calculate SMI (cm2/m2). With the presence or absence of osteoporosis as a state variable, the receiver operator characteristic (ROC) curve was used to analyze the efficacy of SMI in predicting osteoporosis. Figure 2a summarizes this part of the process.

The research process of this study. a The process of building the SMI model. b The process of building CNN models. SMA, skeletal muscle area; SMI, skeletal muscle index; ROI, region of interesting; Grad-CAM, gradient-weighted class activation mapping; CNN, convolutional neural network

In the T12 vertebral plane CT images, only the ROI was isolated by cropping, with non-ROI areas such as the spinal cord and lungs being excluded. These ROI images were subsequently utilized for training and validating the CNN model. The data set of Institution 1 is the training set, and the data set of Institution 2, Institution 3, and Institution 4 is the external test set 1, external test set 2, and external test set 3, respectively. In this study, five CNN models (Densent121, Inception_v3, Googlenet, Resnet50, and VGG16) are constructed. Moreover, all models were pre-trained in the ImageNet dataset. The training process encompasses both forward computation and backpropagation. Before training, the input ROI was resized to 299 × 299 pixels for Inception_v3 and 224 × 224 pixels for the other CNN models, respectively. A stochastic gradient descent (SGD) optimizer was utilized to update the model parameters, with an initial learning rate of 0.01, which was adaptively adjusted using the cosine annealing algorithm. The training process extended over 200 epochs, with a fixed batch size of 32.

In addition, we use the gradient-weighted class activation mapping (Grad-CAM) [15] technique to visualize the last convolutional layer in the CNN model. Grad-CAM enables us to better understand the critical content of deep learning models to make decision recognition image information. Figure 2b summarizes this part of the process.

The distribution of clinical baseline characteristic data was evaluated using the Shapiro-Wilk test. Patient characteristics were expressed as mean ± standard deviation for continuous variables and percentages for categorical variables. Continuous variables with a normal distribution underwent analysis using the Student’s t-test, whereas categorical variables were presented as percentages and analyzed using the Pearson Chi-square test. Performance evaluation of various models involved metrics such as sensitivity, specificity, accuracy, F1-score, and the area under the ROC curve. Model comparisons were conducted using the DeLong test on the ROC curves [16]. All data were processed using SPSS (version 26.0; SPSS et al, USA) and Python (version 3.9.7).