Imbalance-aware spectral classification via logit-label adjustment and multi-expert mutual learning

The integration of deep learning with molecular vibrational spectroscopy has been widely investigated in domains such as disease diagnosis and food safety. However, existing spectral recognition methods assume a balanced class distribution in the training set, where each class contains an equal number of samples. In real-world scenarios, such as rare disease diagnosis, class imbalance is far more common, posing significant challenges to model generalization. To tackle this issue, we propose Spec-ImC, a novel approach specifically designed to address class imbalance in spectral classification and better align with practical application needs. Specifically, we introduce a Logit-Label Adjustment (LLA) strategy, which recalibrates both logits and labels to mitigate the bias in decision boundaries toward many-shot classes. To further enhance model robustness, we propose a Cosine-Annealed Data Augmentation (CADA) strategy together with a novel Consistency Regularization (CR) technique, which stabilizes gradient updates even under strong data augmentation. Furthermore, we develop a Multi-Expert Mutual Learning (MEML) mechanism, enabling multiple expert models to exchange knowledge, thereby fostering feature diversity and boosting recognition performance, particularly for few-shot classes. Extensive experiments on two bacterial spectral datasets, namely Bacterial ID and Bacterial Stain, as well as one cancer spectral dataset, namely Cancer, demonstrate that our method consistently outperforms state-of-the-art methods under various simulated imbalance scenarios, indicating its effectiveness and potential applicability to class-imbalanced spectral analysis tasks.

The proposed framework effectively mitigates the adverse effects of class imbalance in spectral classification and provides a robust solution for molecular vibrational spectroscopy applications with long-tailed data distributions. Its superior performance and generalizability indicate strong potential for real-world spectral analysis tasks, particularly in applications involving rare categories and limited training samples.