Label-Free Structure Discovery in Breast Cancer Histopathology: K-Means and Fuzzy C-Means on Handcrafted Descriptors under Subtype-Disjoint Cross-Validation
Year: 2026
Type: preprint
Abstract: Background There will be over 3.2 million new breast cancer cases a year by 2050 and most of these will be in low- and middle-income countries (LMICs) where pathologists are limited. The current state-of-the-art approach to computational pathology is supervised deep learning, which relies on the availability of large collections of annotated images which are hard to be assembled in such setting. First, we explore the ability of hand-crafted morphological and textural descriptors with only unsupervised clustering of the descriptors to distinguish benign from malignant breast histopathology. Methods Seven-thousand, nine-hundred and nine images from 82 patients were analysed, with 148 features per image extracted from five feature families (GLCM, LBP, Gabor, morphological, statistical intensity) after five stages of pre-processing (resize to 224×224, Macenko stain normalisation, Gaussian filtering, median filtering, and CLAHE applied to the LAB luminance channel). Features were z-scored and reduced by PCA to 21–22 features at a 95% variance threshold (85.6% reduction), fit on training data only. K-Means and Fuzzy C-Means were compared under a subtype-disjoint 4-fold cross-validation, where each test fold held out one full benign and one full malignant tumour subtype, ensuring patient-level disjointness and assessing cross-subtype generalisation. Results K-Means selected k = 2 in three folds and k = 4 in the fourth; Fuzzy C-Means selected c = 2 in all folds. K-Means showed higher internal validity than Fuzzy C-Means (silhouette 0.2197 ± 0.0450 vs. 0.1740 ± 0.0376; Davies–Bouldin 1.8113 ± 0.1378 vs. 2.0328 ± 0.1474) and comparable agreement with diagnosis under a post-hoc oracle cluster-to-class mapping (accuracy 0.6871 ± 0.1559 vs. 0.6854 ± 0.1349), against an all-malignant constant-predictor baseline of 0.600. In 3 of 8 algorithm×fold evaluations the post-hoc mapping degenerated to a constant malignant assignment, so aggregate sensitivity figures should not be interpreted as detection performance. Conclusions Handcrafted descriptors carry a modest signal aligned with malignancy, but not enough to support diagnostic use. This is a preliminary, label-free benchmark rather than a diagnostic system: clinical validity is not established, the mapping uses test labels, and evaluation is retrospective on a single Brazilian cohort.
Source: Research Square
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204164902