https://atlas.rsna.org/schemas/2025-11/model.json

Deep radiomics model for osteoporosis diagnosis on hip radiographs

https://dx.doi.org/10.1148/ryai.210212

1.0

moc.liamg@20eahcdh

Supported by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute, funded by the Ministry of Health & Welfare, Republic of Korea (grant HI20C2092), and the SNUH Research Fund (grant 04-2020-2060).

Institutional Review Boards of Seoul National University Hospital and Seoul Metropolitan Government–Seoul National University Boramae Medical Center approved this retrospective study; requirement for patient informed consent was waived.

Deep feature extractor: Siamese convolutional neural network with six convolution–batch normalization–ReLU blocks and global average pooling; inputs were paired crops of bilateral proximal femurs (or left/right halves of whole image). Feature selection via LASSO. Final classifier: multilayer perceptron integrating 10 selected deep features (per patient), 16 selected texture features, and 3 clinical features. Proximal femur segmentation for cropping/texture features: U-shaped Fusion-Net.

Code available at https://github.com/SWKoreaBME/deep_radiomics_osteoporosis

Opportunistic diagnosis of osteoporosis from routine hip radiographs; improved radiologist performance when used as a second reader; potential to triage patients for confirmatory DXA.

Clinical decision support systems; second-reader assistance; opportunistic screening/triage.

Threshold selected at the point of 90% sensitivity through internal validation.

Automated extraction of deep and texture features from radiographs with automated prediction; assists (does not replace) clinician decision-making.

Adults undergoing hip or pelvis anteroposterior radiography; intended to identify patients with osteoporosis (T-score ≤ −2.5 at proximal femur by DXA) for further evaluation.

Hip/pelvis AP digital radiographs; bilateral proximal femur crops generated by automated segmentation; clinical features: age, sex, weight. Texture features (shape, first-order, second-order) extracted with PyRadiomics from original and wavelet-transformed images.

Use bilateral hip/pelvis AP radiographs as input (both femurs required). For decision support, apply the threshold tuned to 90% sensitivity from internal validation. Can be used as a second reader to assist radiologists. Ensure images have sufficient coverage and quality (no implants, fractures, tumors, or obscuring markers).

Retrospective, two-center study; potential selection bias. Classification (not regression) output—does not provide continuous BMD/T-score for treatment monitoring. Requires bilateral femurs (Siamese input); not applicable to unilateral studies or cases after hip fracture/surgery. Performance may vary with acquisition differences; texture features showed lower generalizability than deep features. External validation limited to Korean institutions; needs validation across other ethnicities and wider body habitus. Soft-tissue within crop may influence predictions; model behavior indicates possible attention to cortical and surrounding soft tissues.

Use as an opportunistic screening and second-reader tool on hip radiographs to identify patients at risk of osteoporosis and refer for confirmatory DXA.

Model code and implementation details provided (PyTorch v1.5) with external validation on an independent institution dataset; segmentation Dice 0.98 reported on independent set.