DiagNeXt: A Two-Stage Attention-Guided ConvNeXt Framework for Kidney Pathology Segmentation and Classification

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Nephrology Diagnostics

Background:

• Accurate kidney pathology segmentation and classification are challenging due to morphological variations and class imbalance.
• Existing computer-aided diagnosis methods struggle with complex kidney image analysis.

Purpose of the Study:

• To introduce DiagNeXt, a two-stage deep learning framework for enhanced kidney pathology segmentation and classification.
• To address challenges in medical image analysis for computer-aided diagnosis of kidney diseases.

Main Methods:

• Developed DiagNeXt, a two-stage framework using attention-enhanced ConvNeXt architectures for segmentation (DiagNeXt-Seg) and classification (DiagNeXt-Cls).
• Incorporated Enhanced Convolutional Blocks (ECBs), spatial attention, Atrous Spatial Pyramid Pooling (ASPP), Context-Aware Feature Fusion (CAFF), and Evidential Deep Learning (EDL).
• Utilized a boundary-aware compound loss and attention-guided skip connections for precise segmentation and feature preservation.

Main Results:

• Achieved 98.9% classification accuracy, surpassing state-of-the-art by 6.8% on a large kidney CT dataset.
• Demonstrated near-perfect AUC scores for Normal (1.000), Tumor (1.000), Cyst (0.999), and Stone (0.994) pathologies.
• Showcased 6.2× faster inference speed and provided clinically interpretable uncertainty maps and attention visualizations.

Conclusions:

• DiagNeXt offers superior diagnostic accuracy and computational efficiency for kidney pathology analysis.
• The framework's interpretability and uncertainty estimation enhance its clinical applicability.
• DiagNeXt shows strong potential for integration into clinical systems for kidney disease diagnosis and treatment planning.