Journal Article

Detection of collagen band–associated regions in H&E-stained colonic biopsies of collagenous colitis patients using superpixel-based feature extraction and neural network classification

Kupčinskas, Juozas, Poškienė, Lina, Petrolis, Robertas, Šabanas, Povilas, Ramonaitė, Rima, Jančiauskas, Dainius, Meilutytė-Lukauskienė, Diana, Kiudelis, Vytautas, Čerapaitė-Trušinskienė, Reda, Kriščiukaitis, Algimantas

2026

Overview

Background Collagenous colitis (CC) is diagnosed histologically and is characterised by a thickened subepithelial collagen band together with inflammatory and epithelial changes. Although routine haematoxylin and eosin (H&E) staining is sufficient for diagnosis in most cases, visual assessment of the collagen band can be challenging in borderline or heterogeneous specimens. Additional stains may be required in diagnostically difficult situations. The aim To develop a machine-learning–based algorithm for detecting subepithelial collagen band-associated regions in routine H&E-stained colonic biopsy images as a decision-support tool for histopathological assessment. Methods H&E-stained colonic biopsy specimens from 36 patients with histologically confirmed CC were imaged at 20 × magnification (1392 × 1040 pixels). Images were segmented into 1,000 superpixels using the Simple Linear Iterative Clustering (SLIC) algorithm. Superpixels overlapping with expert-provided rough annotations of the collagen band were labelled and characterised using normalised RGB histograms. A feed-forward neural network classifier (three hidden layers, 10 neurons per layer) was trained to distinguish collagen band–associated from non-collagen regions. Class imbalance was addressed by data augmentation of minority-class superpixels. Post-processing with connected-component size filtering was applied to enforce spatial continuity. Superpixel-level performance was evaluated quantitatively, and image-level outputs were assessed using expert acceptability scoring. Results The classifier achieved a superpixel-wise accuracy of 0.928 (sensitivity 0.898, specificity 0.953). Size-based post-processing substantially reduced isolated false-positive detections. At the image level, the final algorithm achieved an acceptability accuracy of 0.846 according to expert evaluation. The model successfully highlighted subepithelial collagen band–associated regions consistent with expert annotations but did not model additional diagnostic features required for complete CC diagnosis. Conclusion Our superpixel-based neural network highlights collagen-rich regions in H&E-stained colonic biopsies, offering decision support for pathologists. As diagnosis of collagenous colitis requires broader histopathological and clinical context, this method is intended as a decision-support tool rather than a stand-alone diagnostic solution.

• Publisher: BioMed Central,BioMed Central Ltd,Springer Nature B.V,BMC
• Subject: Acceptability / Algorithms / Annotations / Biopsy / Classification / Clustering / Colitis / Collagen / Data augmentation / Decision support systems / Diagnosis / Diarrhea / Fecal incontinence / Inflammatory bowel disease / Machine learning / Medical imaging / Medicine / Medicine & Public Health / Morphology / Neural networks / Pathology / Performance evaluation / Quality of life