SPRN: Spectral-Partitioned Residual Network with Homogeneous Pixel Detection for Hyperspectral Cancer Diagnosis

Hyperspectral Imaging (HSI) technology has emerged as a critical tool for non-invasive cancer detection and diagnosis, eliminating the need for tissue biopsies while providing detailed diagnostic insights. With advancements in artificial intelligence, the analysis of HSI data and applications in precision medicine have seen significant breakthroughs. While traditional image recognition methods can offer preliminary diagnostic outcomes, they often fail to leverage the nuanced cellular and biochemical information embedded in heterogeneous tissue regions. In contrast, deep learning approaches excel by automatically extracting features and adaptively training on spatial-spectral HSI data, thereby enhancing cancer diagnostic accuracy. This paper presents a Spectral-Partitioned Residual Network (SPRN) based on Convolutional Neural Networks (CNNs). The proposed framework integrates a spectral partitioning architecture with a homogeneous pixel detection module to optimize cancer classification. The spectral-partitioned residual structure mitigates spectral redundancy and prevents overfitting, while the homogeneous pixel detection module minimizes noise from interfering pixels. Furthermore, an attention mechanism is introduced to dynamically prioritize spectral bands most relevant to classification, thereby boosting model efficacy. Experimental validation demonstrates that SPRN achieves 98.8% accuracy and a 97.7% Kappa coefficient in detecting grade IV glioblastoma multiforme cancerous transformation, significantly outperforming conventional methods.