Region-Focused Patch-Based Mixup for Improving Classification in Low-Resource Medical Imaging

Objective: This study aims to enhance performance of classification accuracy in grayscale medical images by addressing the limitations posed by small and imbalanced datasets. The focus is to develop a simple yet effective data augmentation technique that improves model generalization without introducing dynamic or adaptive complexities. Method: A novel augmentation technique, Region-Guided Mixup Augmentation (RGM-Aug), is proposed. The method involves blending a fixed-size patch from one image into another image of the same class at a predetermined location, introducing localized feature diversity while preserving semantic integrity. The augmented dataset trains the lightweight MobileNetV3-Small model. Performance is evaluated based on accuracy, Micro-F1, Macro-F1 scores, and computational efficiency. Findings: The application of RGM-Aug resulted in substantial improvements across all evaluated models. The proposed model achieved 98.63% accuracy. Comparative analysis with existing methods demonstrated superior performance in classification metrics and efficiency. The augmentation technique also led to faster model convergence and reduced validation loss, indicating better generalization. Novelty: Unlike existing augmentation strategies that rely on global transformations or adaptive mechanisms, RGM-Aug introduces a deterministic, region-focused approach that enriches intra-class variability with minimal computational overhead. This method is particularly effective for small-scale medical imaging datasets and is designed to be architecture-agnostic, making it adaptable to various deep learning models without additional complexity. Keywords: Liver Steatosis, Ultrasound Imaging, Region-Guided Mixup Augmentation, Deep Learning, Data Augmentation