Volumetric NephroNet-3D: A Patient-Disjoint, Slice-Consistent Spatial-Temporal Transformer for Multi-Class Renal Lesion and Lithiasis Classification in Dynamic Contrast-Enhanced Abdominal CT

• Abilly Elly

  Texas University

  Author

Renal lesions and lithiasis represent a significant diagnostic challenge in abdominal imaging, with dynamic contrast-enhanced computed tomography (DCE-CT) serving as the primary non-invasive modality for characterization. However, existing deep learning approaches for renal pathology classification predominantly rely on 2D slice-level analysis or conventional 3D convolutional networks that fail to adequately model the spatial-temporal dynamics of contrast enhancement across multiple phases. This study introduces Volumetric NephroNet-3D, a slice-consistent spatial-temporal transformer architecture designed for patient-disjoint multi-class classification of renal lesions and lithiasis in DCE-CT. The proposed framework integrates depthwise-separable convolutional feature extraction with a specialized cross-phase attention mechanism that captures enhancement kinetics across corticomedullary, nephrographic, and excretory phases, while maintaining slice-level consistency through volumetric token aggregation. Using a patient-disjoint evaluation protocol on a curated dataset of 4,820 contrast-enhanced abdominal CT volumes spanning normal parenchyma, cysts, solid tumors, and urinary calculi, Volumetric NephroNet-3D achieved 89.4% overall classification accuracy, with per-class sensitivity ranging from 86.2% (lithiasis) to 93.1% (solid tumors) and specificity exceeding 94% across all categories. Comparative analysis demonstrated that the proposed architecture outperformed 2D ResNet-50, 3D U-Net-based classification, and slice-level transformer baselines by 4.7–11.3 percentage points in accuracy. The patient-disjoint validation strategy revealed that conventional slice-level splitting inflates performance metrics by an average of 7.8 percentage points, underscoring the critical importance of patient-level partitioning for clinically meaningful evaluation. These findings establish Volumetric NephroNet-3D as a robust framework for automated renal pathology characterization, with implications for reducing diagnostic variability and supporting radiologist workflow in abdominal CT interpretation.

• pdf

Copyright (c) 2026 Abilly Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.