Morphological saliency prior-guided band-aware multi-modal disease diagnostic fusion network for vibrational spectroscopy.

Multimodal vibrational spectroscopy provides a powerful biochemical probe for non-invasive clinical diagnosis. However, when mapping high-dimensional spectral features, existing deep neural networks are highly susceptible to overfitting high-variance noise, including baseline drift and amorphous broadband signals. Because underlying physicochemical constraints are rarely incorporated into representation learning, such diagnostically driven paradigms often lack prior-based interpretability, thereby limiting their reliability in complex medical decision-making. Here, we propose a Morphological Saliency Prior-guided Band-aware multimodal fusion network, termed MSPBA. In contrast to previous studies on spectral disease diagnosis that primarily emphasize multimodal fusion architectures or improvements in classification performance, our work focuses on constructing a Morphological Saliency Prior jointly defined by peak prominence, full width at half maximum and local residual fluctuations. This prior is embedded into both the model input and the early feature-extraction stage, enabling the network to preferentially attend to physically reliable spectral bands during the learning of high-dimensional serum spectral representations. By transforming spectral regions with both high response intensity and narrow peak width into prior masks, the proposed framework guides the neural network towards highly ordered core bands at the early stage of feature extraction, thereby reducing interference from complex broadband noise from a physicochemically informed perspective. In addition, we designed a lightweight band-aware synergy module to adaptively capture complementary feature associations across modalities. Validation on two clinical serum cohorts comprising more than 570 cases, collected from different hospital sources, instrument platforms and spectral dimensionalities, demonstrated that MSPBA achieved strong diagnostic performance and cross-cohort stability in both three-class thyroid tumour classification and binary thyroid dysfunction classification. Specifically, the model achieved an accuracy of 97.97% and an AUC of 0.9961 in the thyroid tumour cohort, and an accuracy of 97.62% and an AUC of 0.9938 in the thyroid dysfunction cohort. Furthermore, MSPBA enabled physically prior-matched band identification for interpretability analysis. By incorporating spectroscopic prior knowledge into the representation-learning process, this study improves the classification performance of multimodal spectral diagnostic models in complex classification tasks, while enhancing discriminative stability and feature interpretability across distinct clinical diagnostic settings.