A Curvature-Controlled Contextual–Perceptual Feature Fusion Framework for ILD Detection From Respiratory Sounds

Interstitial lung disease (ILD) screening from respiratory sounds (RSs) remains challenging due to the subtle acoustic differences between pathological and healthy patterns, compounded by limited labeled medical audio data. This paper proposes a curvature-controlled (CC) contextual–perceptual feature fusion framework for automated detection of ILD from RSs. High-level contextual representations extracted using a pre-trained wav2vec 2.0 model are combined with perceptually relevant acoustic features obtained from the encodec neural audio codec. The extracted representations are fused using a geometry-aware fusion strategy operating in a CC latent space. A selective fine-tuning strategy is employed to adapt higher-level acoustic representations to the target task while preserving generalization capability. Experimental results demonstrate that the proposed approach consistently outperforms conventional fusion methods with an accuracy of $86.2 \pm 1.5 \%$, highlighting the effectiveness of curvature-aware multimodal representation learning for RS analysis.