Akturk, B.Bilgin, G.2026-02-102026-02-102025978151069265797815106905619781510699663978151069330297815106922519781510692275978151069308197815106910019781510690844978151069146910.1117/12.31014712-s2.0-105026949203https://doi.org/10.1117/12.3101471https://hdl.handle.net/20.500.14627/1422The classification of histopathology image data using deep learning algorithms is a pivotal aspect in cancer diagnosis, offering crucial benefits in terms of accuracy and rapid decision-making for pathologists, technologists, clinicians, and researchers. However, the complex process of data structuring and labeling often involves the presence of noisy or outlier labels and images, which inevitably diminish the training and predictive efficacy of classification models. This study introduces an innovative approach for classifying histopathology images with noisy characteristics, presenting a learning process resilient to noise. In this dataset, where classification performance is inherently challenged, popular transfer learning models are integrated with Bayesian deep learning and ensemble learning algorithms. This integration helps mitigate common problems such as overfitting and underfitting in outlier data predictions, yielding more stable outcomes. These hybrid methods, unaffected by the noise and uncertainty in data structures, demonstrate performance comparable to, or even surpassing, transfer learning algorithms. Consequently, the viability of addressing real-world noisy dataset scenarios is enhanced. Experimental results indicate that the proposed ensemble Bayesian transfer learning algorithms match the success of pre-trained transfer learning algorithms and offer an alternative to time-consuming model validation techniques, underscoring the practical application of ensemble Bayesian transfer learning in enhancing uncertainty assessments. · © 2025 SPIE · 0277-786X ·eninfo:eu-repo/semantics/closedAccessBayesian Deep LearningEnsemble LearningHistopathology Image ClassificationNoisy LabelsTransfer LearningConference Object