Non-segmented unsupervised learning of multispectral whole slide images for robust analysis of tissue repair and regeneration

Analyzing whole tissue architecture remains challenging due to the inherent complexity of multicellular organization, variable morphology, and the limitations of conventional segmentation-based image analysis. Traditional approaches often rely on partial sampling or nuclear/cytoplasmic boundaries, which risk introducing bias and fail to capture the contextual interplay of diverse tissue compartments. To overcome these barriers, we developed a segmentation free framework for analyzing multispectral whole slide images (WSIs). By tiling WSIs into fixed sized regions and extracting quantitative tile features, we applied unsupervised machine learning to systematically reveal patterns of tissue organization at scale. This approach preserved spatial context without the need for cell-level delineation, recapitulating expected compartments such as epidermis, adipose, and scab, while also revealing subtle but coherent substructures within stromal and granulation regions. Applied to murine wound healing, the method distinguished wild type from diabetic repair dynamics without prior labels, uncovering both gross and nuanced differences in tissue composition. Together, this work establishes a robust, unbiased strategy for whole-tissue analysis that circumvents the limitations of segmentation, leverages unsupervised learning for discovery, and advances the study of tissue repair and regenerative pathology.