Evaluating Data Augmentation Strategies for Robust Cross-Dataset Generalization in Wound Classification

This study investigates the impact of data augmentation strategies on the cross-dataset generalization of deep learning models for wound classification. Using the AZH dataset for model training, we implemented and optimized various augmentation strategies-categorized as photometric, geometric, elastic, cutout, and automated methods (RandAugment and TrivialAugmentWide)-to enhance model robustness. Each augmentation's parameters were tuned using Optuna to maximize performance. The optimized augmentations were then evaluated on the Medetec dataset to analyze generalization performance across datasets. Results indicate that photometric and cutout transformations, particularly when combined, significantly improve cross-dataset accuracy and F1-scores. TrivialAugmentWide and RandAugment provide flexible, competitive alternatives, achieving robust generalization without extensive tuning. These findings underscore the importance of both tailored and automated augmentation strategies for enhancing model performance in diverse clinical imaging environments.