Superpixel-Based Endmember Extraction for Hyperspectral Endometriosis Detection

Abstract Hyperspectral Imaging (HSI) is a promising tool for assisting medical diagnostics, as it enables precise tissue characterization and differentiation through detailed spectral analysis. However, HSI data analysis faces challenges in reliably identifying relevant regions (e.g., tumors, vascular structures) due to spectral variability, complicating universal algorithm development, especially with limited ground truth data. Focusing on endometriosis as a medical use case, we generate plausible reference spectra (endmembers) for affected tissue using HSI data, despite limited and weakly annotated datasets. Our processing pipeline includes Savitzky-Golay (SG) smoothing, Standard Normal Variate (SNV) standardization, Principal Component Analysis (PCA), superpixel segmentation and the Pixel Purity Index (PPI). Validation via Intersection over Union (IoU) achieves 0.94 accuracy in lesion detection.