Dimensionality reduction of medical hyperspectral images based on GPCA-multiscale tensor low-rank graph embedding

Early pathological diagnosis plays a vital role in disease treatment. Medical microscopic hyperspectral imaging technology provides a new perspective for the early detection of diseases. However, the large number of spectral bands introduces a wealth of spectral features, which leads to data redundancy and noise. This not only significantly affects image recognition and classification performance but also increases computational and storage demands. To address this problem, this paper proposes a novel dimensionality reduction framework based on grouped principal component analysis multiscale low-rank graph embedding (GPCA-MLRGE). Specifically, Grouped Principal Component Analysis (GPCA) and skip-connection module achieve a balance between feature compression and local information retention, which significantly reduces the computational complexity; the multi-scale tensor low-rank graph embedding module effectively extracts the spatial-spectral synergistic features of cancerous regions by using the hierarchical feature fusion mechanism. Finally, the extracted spatial spectrum features are fed into a commonly used Support Vector Machine (SVM) classifier to verify the dimensionality reduction effect. The experimental results on the precancerous lesions in gastric cancer (PLGC) and cholangiocarcinoma (CCA) datasets validates the effectiveness of the method.