An Attention-Enhanced YOLOv8 Model For Accurate Multi-Class Kidney Abnormality Detection in CT Imaging

Abstract

Kidney diseases such as cysts, stones, and tumors are a major global health burden, often remaining undiagnosed until advanced stages due to limited access to specialized care. While deep learning models like YOLOv8 have shown promise in automating detection from medical images, their performance is often hindered by class imbalance and poor feature discrimination in complex clinical data. This paper introduces an enhanced deep learning framework for the accurate multi-class classification of kidney abnormalities from computed tomography (CT) scans. Our novel approach integrates a Convolutional Block Attention Module (CBAM) into the YOLOv8 architecture, enabling the model to focus diagnostically relevant regions within the kidney. To address significant class imbalance, we adopt a Focal Loss function, which prioritizes difficult and underrepresented cases during training. We further enhance model robustness through an advanced data augmentation pipeline incorporating mixup and cutmix strategies. The proposed system is trained and evaluated on a publicly available dataset of 12,446 annotated CT images across four categories: cyst, tumor, stone, and normal. Experimental results demonstrate a substantial improvement over the baseline YOLOv8 model, achieving an overall classification accuracy of 91.47%, with precision, recall, and F1-scores of 90.32%, 88.76%, and 89.21%, respectively. Notably, the recall for the critical and often-missed tumor class improved from 30.41% to 82.34%. This work presents a significant step toward reliable, automated diagnostic support, offering a tool that can assist clinicians in early and accurate detection of renal pathologies, thereby improving patient outcomes.