MyUncleVision: An Interpretable Lattice-Based Framework for Unsupervised Anomaly Discovery in Whole-Slide Histopathology Images
Year: 2026
Type: preprint
Abstract: Manuscript prepared for Computer Methods and Programs in Biomedicine (Elsevier), with cover letter and highlights (docx and pdf). Abstract. Unsupervised discovery in whole-slide images (WSI) has to separate unusual tissue organization from artifacts and routine histology, and its outputs must be checkable by experts. We present MyUncleVision, an open pipeline that places patch embeddings on a tissue lattice and measures their organization with the quantities of the maintenance-lattice framework MyUncle: local disorder, neighbourhood entropy, transitions between tissue states, and a three-axis novelty score (appearance rarity, state-wise disorder, contact rarity). Regions are clustered across slides, ranked, mapped and explained with measured numbers. The order and entropy definitions generalize MyUncle's from two to d dimensions and reproduce its local disorder and von Mises entropy exactly at d = 2, which automated tests verify against the unmodified core. On eight synthetic slides with planted anomalies and artifacts, the novelty score reaches an AUROC of 0.90 (slide-bootstrap interval 0.86–0.94) with a handcrafted encoder but 0.22 with DINOv2, which is below chance: the planted states cover 3.1% of the patches, more than the 1% rarity scale, and DINOv2 embeds them more homogeneously than normal tissue. The appearance axis alone recovers with DINOv2 once the scale exceeds the prevalence (0.91 at a scale of 10%). Quality control removed 42% of the planted artifact patches. The platform has not been run on kidney tissue and no pathologist has reviewed any output; the evaluation protocol for the Kidney Precision Medicine Project images is specified. Code, tests and results are open. Status. Validated on synthetic slides with planted anomalies and artifacts only; it has not been run on kidney tissue (Kidney Precision Medicine Project images), and no pathologist has reviewed any output. On the synthetic benchmark the novelty score reaches an AUROC of about 0.90 with a handcrafted encoder and 0.22 (below chance) with DINOv2 at the default settings; the manuscript analyses why. Code: https://github.com/sandlerleon/MyUncleVision. Built on MyUncle (doi:10.5281/zenodo.21223570), vendored unmodified. A generative AI tool (Claude, Anthropic) assisted with code, analyses and drafting; the author reviewed the content and takes full responsibility.
Author: Sandler, Leon
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225645347