CyclicAugment: Optimized Medical Image Analysis via Adaptive Augmentation Intensity

Computer-aided diagnosis (CADx) systems play a crucial role in accurately diagnosing and monitoring diseases through medical imaging. However, there are many challenges, such as data scarcity and complex structural patterns, limiting the performance of deep-learning models. Although conventional data-augmentation techniques can help mitigate these limitations, excessive augmentation may distort critical diagnostic information, leading to suboptimal performance. To address this issue, this study proposes CyclicAugment, a novel data-augmentation strategy that dynamically adjusts augmentation intensity in a cyclic manner throughout training. The method gradually transitions from using original data as weak augmentation, subsequently applying strong augmentation before reverting to weak augmentation again, reintroducing the original data. This process ensures stable training through an adaptive learning rate. By cyclically adjusting augmentation intensity, our approach enhances data diversity while preserving essential diagnostic information, thereby improving the generalizability of deep learning models. We validated CyclicAugment across six different medical imaging datasets and multiple deep learning architectures, including convolutional neural networks, vision transformers, and multilayer perceptrons. Our framework achieved a maximum accuracy improvement of 8.8%, reaching a peak performance of 0.927 and thereby demonstrating its effectiveness in enhancing medical image analysis.