A Hybrid Convolutional Recurrent Multidomain Attention Generative Adversarial Network (CRMDA‐GAN) Framework for Low‐Dose Computed Tomography Denoising in Public Healthcare

Computed tomography, a noninvasive medical procedure for high‐resolution detailed cross‐sectional images of the internal body organs exposes a patient to hazardous health risks like cancer due to high radiation over time, making it a growing concern. Low‐dose computed tomography (LDCT) is a safe alternative with reduced radiation exposure but the generated CT images are corrupted with noise and artifacts that critically degrade the image quality, impacting the diagnostic accuracy and interpretation. Despite extensive research in LDCT denoising, it is still a challenging and an open area of active investigation that aims to strike an enhanced tradeoff between noise reduction and CT image quality. This paper introduces a hybrid deep learning framework in which the power of convolutional neural network (CNN) U‐Net is complemented by a bottleneck encompassing a recurrent gated block to capture interslice correlations and an attention block that focuses on multidomain features to learn global anatomical context. To prioritize pixel‐level correctness and realism in the denoised CT image, the network is trained with least absolute deviations and adversarial loss. The LDCT image denoising capability of the proposed model is experimentally evaluated on the Mayo Clinic TCIA (Public Dataset) and a methodically and strategically self‐collected private dataset. The model achieves 40.0 db for peak‐signal‐to‐noise ratio (PSNR), 0.95 for Structural Similarity Index (SSIM), and 0.0133 for root mean square error (RMSE) and generates visually coherent images outperforming the performance of the benchmark denoising networks with potential for effective LDCT denoising in public healthcare.