Comparative Evaluation of Twenty Vision Transformer Variants for Six-Class Gastrointestinal Endoscopic Image Classification
Computers and Electronics in Medicine · 30.07.2026 · 3 (2) · 162-173
DOI: 10.69882/adba.cem.2026077 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computers and Electronics in Medicine, online erschienen am 30.07.2026. Verfasst von Yousuf Islam und 3 weiteren Autor:innen. Die Arbeit nennt 52 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
This study compares 20 vision transformer variants for six-class gastrointestinal endoscopic image classification using a labeled subset of HyperKvasir. The evaluated classes were cecum, polyps, pylorus, retroflex-rectum, retroflex-stomach, and z-line. Repeated row totals from six confusion matrices indicated an evaluation set of 768 images. The experiment was configured for five-fold cross-validation with a maximum of 50 epochs and a random seed of 42; however, the numerical fold-level file contained results only for Fold 1. Learning curves supplied qualitative information for selected runs in Folds 2–5 and complete graphical coverage for EViT-B0, EViT-B1, and EViT-B3. In the available numerical results, MaxViT-Tiny achieved the highest accuracy (99.09%), F1 score (99.10%), precision (99.12%), and recall (99.09%), whereas TinyViT-21M achieved the highest AUC (99.98%). MaxViT had the highest mean family accuracy, while TinyViT had the highest mean family AUC. Accuracy and F1 were strongly correlated, but the association between accuracy and AUC was weak. The learning curves showed rapid convergence with model-dependent validation fluctuations, and the confusion matrices located most errors in cecum–polyp and polyp–retroflex-rectum pairs. These findings support further evaluation of hybrid local–global transformer designs. Complete fold-level results, patient-level split information, calibration analysis, and external validation are required before drawing conclusions about cross-fold or clinical generalizability.
Zitieren
APA
Islam, Y., Aouiti, C., Ladjeroud, A., & Ahmad, M. (2026). Comparative Evaluation of Twenty Vision Transformer Variants for Six-Class Gastrointestinal Endoscopic Image Classification. Computers and Electronics in Medicine, 3(2), 162-173. https://doi.org/10.69882/adba.cem.2026077
Vancouver
Islam Y, Aouiti C, Ladjeroud A, Ahmad M. Comparative Evaluation of Twenty Vision Transformer Variants for Six-Class Gastrointestinal Endoscopic Image Classification. Computers and Electronics in Medicine. 2026;3(2):162-173. doi:10.69882/adba.cem.2026077
BibTeX
@article{Islam2026,
author = {Islam, Yousuf and Aouiti, Chauoki and Ladjeroud, Asma and Ahmad, Meshkat},
title = {Comparative Evaluation of Twenty Vision Transformer Variants for Six-Class Gastrointestinal Endoscopic Image Classification},
journal = {Computers and Electronics in Medicine},
year = {2026},
volume = {3},
number = {2},
pages = {162--173},
doi = {10.69882/adba.cem.2026077}
}
EUVIMED