Unsupervised separation of benign and malignant tissue in Nigerian breast histopathology photomicrographs: an exploratory evaluation of K-means and fuzzy C-means clustering

Computer-aided diagnosis for breast histopathology has been developed almost exclusively using whole-slide images from high-income countries, whereas many laboratories in sub-Saharan Africa operate with conventional light microscopes and camera-captured photomicrographs. It is not known whether such images contain unsupervised structure that corresponds to the benign–malignant distinction. We investigated whether partitional clustering can identify coherent tissue-pattern structure in this type of data. Breast slides from a tertiary hospital in Nigeria were stained with haematoxylin and eosin and photographed at ×4, ×10 and ×40. After quality control, 14,084 tiles from 3,244 photomicrographs of 186 patients were analysed. One hundred and eighteen handcrafted texture, morphological and intensity descriptors were computed per tile and standardised. Swarm feature selection, run on 151 development patients with 35 patients withheld and an objective combining patient-grouped cross-validated classification accuracy (so label-informed) and the silhouette coefficient, retained 45 descriptors. K-means and Fuzzy C-means were evaluated using internal validity indices, a sweep over two to fifteen clusters, and agreement (adjusted Rand index) with the diagnostic label inherited from the parent slide. Both algorithms yielded two clusters with weak internal structure. K-means gave a silhouette coefficient of 0.2866, a Davies–Bouldin index of 1.3209 and a Calinski–Harabasz index of 5,759.30; Fuzzy C-means gave 0.2794, 1.3704 and 5,676.43 (partition coefficient 0.6308). The silhouette coefficient was highest at two clusters across the sweep. Agreement with the inherited label was weak (adjusted Rand index 0.0231 and 0.0226), and purity for both algorithms (0.5806) equalled the majority-class baseline, i.e. no improvement. Cluster membership was more strongly associated with patient identity than with diagnosis (normalised mutual information 0.0419 vs. 0.0109) and only weakly with magnification (0.0073). In this exploratory analysis, K-means and Fuzzy C-means clustering of a handcrafted, label-informed 45-descriptor representation did not recover the inherited slide-level diagnostic label. The partition was more strongly associated with patient identity than with diagnosis; because no technical factors were measured, this cannot be attributed to technical rather than biological variation. The results do not show that unsupervised methods in general cannot separate benign from malignant tissue, and any quality-control use of clustering remains unvalidated.