An Attention-Enhanced Yolov8 Model For Accurate Multi-Class Kidney Abnormality Detection In CT Imaging

Abstract

Kidney problems like cysts, stones, and tumors are a huge health issue worldwide. The tricky part? A lot of people don’t even know they have them until things get pretty serious, mostly because it’s tough to get specialized care. Deep learning models—take YOLOv8, for example—are starting to help doctors spot these issues in medical images. But they still struggle, especially when the data is unbalanced or the images are just plain complicated. In this paper, we lay out a better way to classify different kidney problems using CT scans. We’ve built on top of YOLOv8 by adding a Convolutional Block Attention Module (CBAM), so the model learns to zero in on the parts of the kidney that really matter. To tackle the problem of class imbalance, we use a Focal Loss function, which makes the training process focus more on the tough, less common cases. And to make our model even stronger, we’ve thrown in a smarter data augmentation pipeline that uses both mixup and cutmix strategies.

We trained and tested our system on a large public dataset—12,446 labeled CT images split into four groups: cyst, tumor, stone, and normal. The results stand out. Our model doesn’t just beat the baseline YOLOv8; it jumps ahead, hitting 91.47% accuracy overall. Precision clocks in at 90.32%, recall at 88.76%, and the F1-score lands at 89.21%. The big story here is the tumor class. Tumors are tricky to catch, but recall for that category soared from 30.41% all the way up to 82.34%. So, yeah, this approach really moves the needle. It gives clinicians a reliable, automated tool for spotting kidney problems early, which can make a real difference for patients.