Unfiltered Sonographic Images for Kidney Study: Evaluating YOLO Variants with Attention Mechanisms for Anatomical Segmentation

Ultrasound imaging is widely used in kidney studies due to its affordability, availability, and non-invasive nature. However, the inherent noise, low contrast, and artifacts in sonographic images pose significant challenges for AI-based analysis. This study investigates the performance of three YOLO model variants, YOLOv8, YOLOv11, and YOLOv12, each integrated with different attention mechanisms: Convolutional Block Attention Module (CBAM), Local Kernel Attention (LKA), and Self-Attention (SA). The models were evaluated on a realworld, raw kidney ultrasound dataset comprising 47 images from both male and female patients, with expert-verified labels. The dataset was divided into training (27 images), validation (10 images), and testing (10 images) sets. Training was conducted using Google Colab, and performance was assessed using metrics such as inference time and segmentation accuracy across 11 key anatomical structures. Among the configurations, YOLOv12 combined with SA achieved the highest accuracy of 60%, demonstrating the benefit of incorporating spatial attention in handling unfiltered clinical images. The models consistently detected well-defined regions like the renal cortex and sinus, while more ambiguous structures such as the medulla and Column of Bertin remained difficult due to poor contrast. Notably, attention mechanisms, especially SA and LKA, enhanced detection in these challenging areas, underscoring their value in clinical applications. This study sets a benchmark for analyzing noisy ultrasound data and highlights the importance of designing tailored preprocessing filters to improve the visibility and detection of complex anatomical regions in future AI-powered kidney ultrasound diagnostics.