MultiScaleKANNet: a hybridCNN-KAN-Transformer architecture forradiographic bone-loss risk stratification fromknee X-rays

Osteoporosis is underdiagnosed because dual-energy X-ray absorptiometry (DXA) is costly and scarce. We present \textbf{MultiScaleKANNet}, a hybrid deep-learning architecture for radiographic bone-loss risk stratification from routine knee X-rays, combining convolutional feature learning, learnable nonlinear transformations via Kolmogorov--Arnold Network (KAN) layers, and Transformer-based multi-scale attention. We evaluated MultiScaleKANNet on 2,035 knee radiographs from four public Kaggle sources with three categories (Healthy, Osteopenia, Osteoporosis). The labels are \emph{proxy labels}---some derived from quantitative ultrasound T-scores rather than DXA---so results represent radiographic risk stratification, not clinical diagnosis. On a stratified held-out test set ($n{=}407$), the model achieved 97.30\% accuracy (95\% CI: 95.3--98.6\%; Cohen's $\kappa{=}0.9584$; MCC${=}0.9585$). A source-held-out evaluation yielded 89.52\% binary accuracy ($\kappa{=}0.7903$), suggesting in-distribution metrics may partly reflect dataset homogeneity. Ablation studies confirm synergistic gains from KAN layers (+2.46\%), multi-scale processing (+4.17\%), and Transformer attention (+4.91\%), with 40\% parameter reduction versus ResNet-18. This is a \emph{methodological feasibility study}; prospective DXA-confirmed validation is required.