Abstract
<title>Abstract</title> <p>An unsupervised multi-class classifier was applied to four morphometrically competitive histopathologic lymphoid tissue samples dominated by B lymphocytes. The study aimed to accurately classify cluster of differentiation 20 immunohistochemically (CD20 IHC) stained lymphoid samples with (a) B cell lineage dominant profiles, (b) prototype unusual diaminobenzidine signalling profiles and (c) visualisation of each layer of the high-performance multi-class model’s operations for explainability and interpretability. CD20 IHC was selected for cellular compartment specificity, broad spectral range, variable staining intensities and high multi-scale feature contrast relative to that of baseline haematoxylin and eosin (H&E). Our deconstructed, explainable feature extraction operations were then normalised and fused to produce a 6657-dimensional unit-hypersphere. Minkowski (MK) and Cauchy kernel density (CKD) scoring-based methods were utilised for the classification task. Three hundred and thirty five prototypes were derived from a total of 652 tiles (augmented to 808 tiles) for training. Both MK and CKD classifiers achieved identical high-performance classification accuracy of 95.7%, Cohens kappa score of 0.938 and a macro-F1 of 95.1%. Unusual diaminobenzidine signalling was demonstrated in sample 1, which was attributed to a high likelihood of immunosuppression in this subset. The manuscript has addressed the paucity of available explainable classification frameworks with abundant visualisations for pathologists to easily interpret. Directions for future investigations include the addition of a complete immunohistochemistry panel and auxiliary genetic translocation status labels to discriminate between B cell and T cell lineage lesions, which was a major limitation to the study.</p>