Adaptive CSWin SqueezeNet optimized improved Kepler YOLOv8n with unified framework for ultrasound kidney image retrieval

Abstract

Kidney stone (KS) detection using ultrasound (US) imaging has become a vital diagnostic tool in urology due to its noninvasive nature and accessibility. However, prevailing models have lacked anatomical context, making KS hard to distinguish from similar conditions. To address these limitations, this research has proposed a novel Adaptive CSWin SqueezeNet-based Improved Kepler YOLOv8n Framework with Unified Image Retrieval (AS-KY-UIR). First, to overcome preconceived notional errors caused by US limitations during augmentation, Artificial Protozoa-optimized Generative Adversarial Network (AP-GAN) has been introduced to dynamically map deviation vectors to improve image quality. For segmentation ambiguity with conditions like Nephrocalcinosis, a hybrid Adaptive Filter-Panoptic CSWin SqueezeNet (AF-PT-SN) model has integrated an adaptive self-guided Filter and panoptic CSWin transformer for enabling effective separation of KS through horizontal-vertical attention. Next, to handle the complexity in KS classification caused by variability in organic matrix compositions, the proposed Improved Kepler optimized Prewitt Relief YOLOv8n Unified Network has used wavelet and relief features to capture gradient-shadow cues, while the Improved Strategy-based Kepler Optimization Algorithm (ISKOA) has refined feature boundaries for precise classification. Finally, to solve annotator bias during image retrieval, a Unified Hierarchical Image Retriever (HIRT) has computed class-wise pixel distances using the Manhattan distance metric for similarity matching. Overall, the AS-KY-UIR framework has enhanced KS detection by improving augmentation with high classification accuracy for supporting timely recovery.