Automated Kidney Disease Detection Using a Dual-Attention and Transformer-Driven Deep Learning Model

Abstract

Kidney diseases are one of the significant health challenges in the world, and they need to be properly diagnosed and in time to avoid the development of the disease and its complications. Computed tomography (CT), ultrasound, and magnetic resonance imaging (MRI) are medical imaging modalities that are important in the evaluation of kidney disease, but manual interpretation is time-consuming and prone to inter-observer variation. This research aims to overcome these issues by establishing KDAT-Net, a new hybrid deep learning architecture that incorporates convolutional neural networks, dual attention mechanisms, and transformer-based global feature modeling to detect kidney diseases automatically. The proposed model uses an InceptionV3 backbone to extract multi-scale spatial information of kidney images. Efficient Channel Attention (ECA) is used to learn inter-channel dependencies with the aim of improving discriminative representation learning, and a spatial attention mechanism to focus attention on disease relevant parts of the anatomy is incorporated. In contrast to the more traditional CNN-based methods which use only local receptive fields, the refined feature maps are converted to a sequence of tokens and encoded by transformer encoder blocks, which allows the modeling of long-range spatial relations as well as global contextual information. The resulting characteristics are then grouped and categorized into various kidney disease types through fully connected layers. Substantial experimental assessments show that the suggested framework is more effective than baseline CNN models and attention-enhanced variants in accuracy, precision, recall, and F1-score. Moreover, the qualitative analysis based on the gradient-based class activation maps demonstrates that the model can emphasize clinically important kidney areas, which makes it easier to interpret and increase its credibility. The proposed KDAT-Net is a powerful and scalable solution to automated kidney disease detection and can assist in clinical decisions in real-world diagnostic processes.