GDT-SwinKid: A hybrid model for precise renal lesion analysis

Abstract

Detecting and delineating renal lesions accurately remains a significant clinical problem due to the variety of kidney pathology and subtle differences in CT image interpretation. In this paper, we present the design of a next-generation hybrid model called GDT-SwinKid (Gamma Distribution-based Swin Transformer for Renal Lesions), which integrates the hierarchical feature attention mechanisms of Swin Transforms with a modified U-Net decoder and employs advanced statistical modeling (specifically through an adaptive Gamma distribution). The design of GDT-SwinKid allows for both precise extraction of fine details regarding kidney lesions, as well as achieving overall contextual awareness using cross-attention and Gamma-modulated feature refinement to address the drawbacks of existing approaches. Through extensive validation utilizing a large set of clinical datasets, GDT-SwinKid achieved better performance through segmentation and classification, obtaining Dice coefficients as high as 0.95, with AUC values approaching 0.99, when compared to leading transformers and convolutional models. An absolute improvement of 5–9% in Dice coefficient compared to conventional U-Net and Swin Transformer baselines, and an increase in AUC-ROC values approaching 0.99, outperforming existing hybrid and transformer-based methods on the same CT kidney dataset. The inclusion of explainable attention maps and deep supervision provides increased trust and accountability while enabling the rapid and robust integration of GDT-SwinKid into diagnostic pipelines for kidney imaging. GDT-SwinKid combines statistical sensitivity, hierarchical attention and clinical transparency to provide a new standard for automated kidney lesion analysis and to increase the reliability and use of newly developed AI techniques in renal imaging.