Work

RGVT: retrieval-guided Vision Transformer for CT-based kidney lesion classification and segmentation

Year: 2026
Type: article
Abstract: Kidney lesion classification and segmentation from computed tomography images remain challenging because cysts, stones, tumors, and normal renal tissue may show overlapping visual characteristics, low-contrast boundaries, and inter-patient variability. This study proposes RGVT, a retrieval-guided Vision Transformer framework for CT-based kidney lesion classification and segmentation. The proposed framework extracts global image features using a Vision Transformer encoder and retrieves visually similar reference-case embeddings from a training-set memory bank. These retrieved embeddings are fused with current-case image features through cross-attention to improve lesion representation before classification and segmentation. The framework was evaluated on a CT kidney lesion dataset 1 containing cyst, normal, stone, and tumor classes and was compared with U-Net, U-Net + HC, and Swin Transformer baselines. In the original image-level evaluation, RGVT achieved an accuracy of 96.40%, precision of 96.20%, recall of 96.00%, F1-score of 96.10%, and macro-AUC of 0.977. For binary kidney-stone segmentation on KSSD2025, RGVT achieved a Dice score of 94.20% and an IoU of 89.50%. An additional group-disjoint analysis was performed using the same RGVT architecture and is reported in the Supplementary Information to examine the effect of examination-level separation on classification performance. The retrieval memory was restricted to training-group representations in this additional analysis. These results are intended to provide complementary evidence regarding the behavior of retrieval-guided visual feature fusion under different data-partitioning conditions.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7226007044
FWCI:
Cites: 5
Cited by:
Related to: 10