Combining deep learning with statistical shape modelling enables automated lower limb measurements with...

Accurate anatomical landmarking is crucial for assessing lower limb alignment, diagnosing deformities, planning surgeries and monitoring treatment outcomes. Traditional methods rely on manual measurements from 2D standing radiographs, which fail to capture 3D bone morphology and are influenced by patient positioning. Weight-bearing computed tomography (WBCT) enables 3D evaluations under physiological loading conditions, but manual landmark identification on WBCT is time-consuming and subject to observer variability. This study aims to leverage deep learning (DL) and statistical shape modelling (SSM) for automated assessment of lower limb alignment and morphology.

A hybrid DL-SSM model automatically calculated 28 lower limb alignment and morphology measurements using 30 full-leg WBCT scans. The DL model was trained in a five-fold cross-validation setting. It automatically segmented the femur, patella, tibia, talus, calcaneus and second metatarsal. A cascaded SSM-fitting methodology automatically identified the necessary 3D landmarks to derive the 28 measurements. The automated measurements were statistically compared to manual measurements performed by three experienced raters on both WBCT scans and 3D bone models.

The DL segmentation model achieved high accuracy, with a mean dice similarity coefficient exceeding 0.96. The proposed method corresponded well to manual assessments, with the magnitude of detected differences generally matching the interobserver reliability of the manual method. The mean absolute error for the angular measurements ranged from 0.35° ± 0.39° to 5.53° ± 4.68°.

The hybrid DL-SSM methodology for automated assessment of lower limb alignment demonstrated comparable reliability to manual methods. This method provides an observer-independent method for 3D lower limb alignment and morphology assessment under weight-bearing conditions.

Level III.