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

Deep Learning BMD (L1 ROI) Tool for CT-based Osteoporosis Assessment

https://pubs.rsna.org/doi/10.1148/ryai.220042

1.0

Perry J. Pickhardt, email: gro.htlaehwu@2tdrahkcipp

Authors declared no funding for this work. Disclosures note J.W.G. NIH grant R01 LM013151 and R.M.S. CRADA grant from PingAn; other individual disclosures listed.

Single-center, retrospective study approved by the institutional review board; HIPAA-compliant; informed consent waived.

Convolutional Neural Network; U-Net variant (TernausNet) with VGG11 encoder; implemented in PyTorch; Adam optimizer; BCEWithLogitsLoss.

Not specified in the article (no public code or download link provided).

Fully automated, explainable CT-based BMD assessment enabling opportunistic osteoporosis screening from routine abdominal/chest CT without additional cost, time, or radiation.

Clinical decision support and workflow optimization for opportunistic screening; automated measurement integrated into radiology workflow/reports (as envisioned by authors).

Primary analysis used L1 trabecular attenuation <100 HU to indicate osteoporosis; alternative operating points discussed (e.g., 140 HU for single-slice/feature-based; 120 HU for seven-slice median) to trade sensitivity/specificity.

Fully automated ROI localization and HU extraction (DL body part regression for L1 level; automated ROI segmentation and measurement).

Adults undergoing abdominal or chest CT for any clinical indication; opportunistic assessment of vertebral trabecular density for osteoporosis screening in the radiology reading environment.

Axial abdominal/chest CT images including L1 level (DICOM); heterogeneous acquisition parameters with and without IV/oral contrast; primarily 120 kVp; various kernels and slice thickness (mostly 5 mm).

Tool identifies L1 level via automated body part regression, segments anterior trabecular ROI using a DL U-Net (TernausNet), and computes HU statistics. Median (or mean) HU can be taken from single slice at mid-L1 or minimum across 3 or 7 contiguous slices (L1±1 or L1±3) to tune specificity vs sensitivity for osteoporosis detection using a chosen HU threshold (e.g., 100 HU). Visual verification of ROI placement is possible.

Single-center development/validation; training labels derived from an older feature-based tool (manual masks not available); performance by technical factors (kVp, contrast, vendor) not stratified; osteoporosis defined by <100 HU manual L1 threshold (not universal standard); study focused on supine scans (nonsupine examples shown but not formally included); severe L1 compression fractures excluded from manual reference standard.

Use single-slice median HU for higher specificity or minimum-of-multislice HU (e.g., 7 slices) for higher sensitivity depending on clinical objective; integrate output into reporting workflow for opportunistic screening.

Training details reported: 24,318 image pairs (from 14,290 CT scans) with 80/20 split; preprocessing to [-500,500] HU -> [0,255]; augmentations (shift, rotation up to 185°, scale); trained 15 epochs with early stopping; LR 1e-4; batch 32; single NVIDIA A100 40GB; ~16 hours training with pretrained encoder.

Training required ~16 hours on single NVIDIA A100 40GB GPU; runtime characteristics not reported.