Computer-Aided Design (CAD) techniques have been developed to assist nephrologists by optimising clinical workflows, ensuring accurate results and effectively handling extensive datasets. The proposed work introduces a Dilated Bottleneck Attention-based Renal Network (DBAR-Net) to automate the diagnosis and classification of kidney diseases like cysts, stones, and tumour. To overcome the challenges caused by complex and overlapping features, the DBAR_Net model implements a multi-feature fusion technique. Two fold convolved layer normalization blocks \(\:({\text{C}\text{L}\text{N}}_{\text{b}1}\) & \(\:{\text{C}\text{L}\text{N}}_{\text{b}2})\) capture fine-grained detail and abstract patterns to achieve faster convergence and improved robustness. Spatially focused features and channel-wise refined features are generated through dual bottleneck attention modules \(\:{(\text{A}}_{\text{b}\text{a}\text{m}1})\:\&\:{(\text{A}}_{\text{b}\text{a}\text{m}2})\) to improve the representation of convolved features by highlighting channel and spatial regions resulting enhanced interpretability and feature generalisation. Additionally, adaptive contextual features are obtained from a dilated convolved layer normalisation block \(\:\left({\text{D}\text{C}\text{L}\text{N}}_{\text{b}}\right)\) , which effectively captures contextual insights from semantic feature interpretation. The resulting features are fused additively and processed through a linear layer with global average pooling and layer normalization. This combination effectively reduces spatial dimensions, internal covariate shifts and improved generalization along with essential features. The proposed approach was evaluated using the CT KIDNEY DATASET that includes 8750 CT images classified into four categories: Normal, Cyst, Tumour, and Stone. Experimental results showed that \(\:\text{t}\text{h}\text{e}\) improved feature detection ability enhanced the performance of DBAR_Net model attaining a F1 score as 0.98 with minimal computational complexity and optimum classification accuracy of 98.86%. The integration of these blocks resulted in precise multi-class kidney disease detection, thereby leading to the superior performance of DBAR_Net compared to other transfer learning models like VGG16, VGG19, ResNet50, EfficientNetB0, Inception V3, MobileNetV2, and Xception.
