The application of multi-instance classification technology based on microscopic hyperspectral imaging in the analysis of microscopic hyperspectral breast cancer dataset

Microscopic hyperspectral imaging (MHSI) of unstained tissue offers a promising, quantitative alternative to traditional histopathology but is challenged by low-contrast morphology and high-dimensional data. Traditional pathological diagnosis, when applied to unstained tissue sections, often relies on the analysis of individual patches. The fatal flaw in this approach, however, lies in the inherent low contrast of unstained sections. This can render any single patch informationally ambiguous, making it difficult for models to accurately differentiate between paracancerous and tumor tissue. To fundamentally address this issue, we propose a new diagnostic paradigm: to shift away from relying on unreliable individual patches and instead perform a holistic diagnosis by intelligently integrating information from all patches across an entire slide. Based on this principle, we designed the Multi-Scale Hierarchical Attention Network (MS-HAN). The core advantage of MS-HAN lies in its hierarchical integration mechanism. It begins by extracting features from each independent patch, but critically, it refrains from making judgments based on any single one. Instead, it employs a multi-head self-attention aggregator to evaluate the importance of all patches within the slide, fusing them into a weighted, global feature vector that represents the slide's overall state. This design enables the model to automatically disregard uninformative or misleading patches and focus its attention on the most diagnostically critical regions.We validated the superiority of this method on a dataset from 60 breast cancer patients. By integrating whole-slide information, MS-HAN achieved an accuracy of 86.2\% and an AUC of 0.91. Furthermore, interpretability analysis confirmed that the model makes its final prediction precisely by identifying and focusing on key diagnostic regions within the slide. These results demonstrate that our integrated diagnostic framework successfully overcomes the limitations of single-patch dependency, offering a more robust and precise solution for stain-free computational pathology.