P061 Multimodal deep learning with cross-attention for femoral neck bone mineral density estimation

Abstract

Abstract Background/Aims Poor bone health is a major public health issue, and low bone mineral density (BMD) increases fracture risk. Dual-energy X-ray absorptiometry (DXA) is the clinical standard for BMD assessment, but cost and limited availability restrict widespread use, especially in low-resource settings. By contrast, plain radiography is far more accessible globally, offering a pragmatic route to support early diagnosis and screening. We therefore present XAttn-BMD, a multimodal deep learning framework that aims to predict femoral neck BMD from plain hip radiographs while integrating health-related clinical metadata to supplement prediction. Methods Data were drawn from the Hertfordshire Cohort Study (HCS; n = 233; age 71.6-80.6 years; 48.9% female). Hip anteroposterior radiographs and structured clinical metadata (including age at x-ray and BMD scan, height, weight, body mass index, alcohol consumption, physical activity, diet quality score, sex, and smoking status) were evaluated using stratified 10-fold cross-validation (per fold: 90% training, 10% validation). The novel model architecture comprises an image backbone and a metadata Multilayer Perceptron fused via bidirectional cross-attention. Regression performance was assessed using Mean Squared Error (MSE), Mean Absolute Error (MAE) and Coefficient of Determination (R² score), and potential screening utility was examined by classifying at clinically relevant femoral neck thresholds (BMD 0.899 g/cm²; T-score = −1 for Lunar femoral neck). Results Through extensive experiments, XAttn-BMD improved regression accuracy and robustness over baseline models. Compared with direct feature concatenation, the cross-attention fusion reduced MSE by 16.7%, MAE by 6.03%, and increased R² score by 16.4%, indicating more effective multimodal integration. In threshold-based screening, the model demonstrated a favourable sensitivity-specificity balance. Across all subjects, it achieved an accuracy of 0.773, with weighted precision 0.831, recall 0.748, and F1-score 0.787, indicating reliable discrimination between individuals with and without osteopenia (−2.5 < T-score < −1.0) and osteoporosis (T-score ≤ −2.5). Conclusion In summary, fusing hip radiographs with health-related metadata via bidirectional cross-attention provides a principled and scalable approach to femoral neck BMD estimation. The method yields consistent gains over direct feature concatenation baselines and facilitates downstream screening applications, particularly where DXA access is constrained. Disclosure Y. Zhang: None. R. Attar: None. N.R. Fuggle: None.

Ähnliche Arbeiten