The segmentation of medical images plays a crucial role in diagnoses, treatment planning, and clinical decision-making. The performance of segmentation models is affected by several reasons, including a lack of labeled medical images and misinterpretation of image features, which affects accuracy. This paper introduces a new technique, Dynamic Contextual Feature Activation (DFCA), that dynamically selects and suppresses image features based on their contextual importance. Two models are presented; DFCA-GAN incorporates DFCA in Generative Adversarial Networks (GAN) to generate high-quality synthetic medical images and masks for data augmentation. DFCA-U-Net is an improved version of U-Net that uses DFCA to emphasize critical regions in image segmentation. The experiments on the ISIC 2016, ISIC 2017 skin lesion datasets, and the Chest X-ray Dataset for Tuberculosis demonstrate that DFCA-U-Net with DFCA-GAN data augmentation outperforms traditional U-Net and DFCA-U-Net, achieving 90.42%, 90.22%, and 79.02% in the Dice coefficient. This study demonstrates that DFCA improves deep learning models and data augmentation increases performance by reducing data scarcity, thereby enhancing the robustness and generalizability of the model.
Rais et al. (Fri,) studied this question.