BATU: A Workflow for Multi-Network Ensemble Learning in Cross-Dataset Generalization of Skin Lesion Analysis

Abstract

The development of computer vision systems for dermatological diagnosis is often hindered by dataset heterogeneity, including differences in image quality, labeling strategies, and patient demographics. In this study, we examine how such heterogeneity affects the generalization ability of computer vision models across three public dermatology image datasets. We trained five different deep learning models on each dataset separately and evaluated their performance in both intra-dataset and cross-dataset settings. To further investigate robustness, we conducted multi-source domain generalization experiments by training models on combinations of two datasets and testing on the third unseen dataset. We observed a significant drop in performance during cross-dataset evaluations. To address this, we applied various ensemble learning methods by combining the predictions from the individual models. Our results demonstrate that ensemble approaches consistently outperform individual models, achieving accuracy improvements exceeding 4% in many cases. These findings highlight the potential of ensemble learning to address challenges related to dataset variability in dermatological image analysis.