KDH-Net: Explainable Medical AI for Multiclass Kidney Disease Characterization from CT Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nephrology

Background:

• Accurate kidney disease differentiation from CT scans is challenging due to visual similarities and data variability.
• Existing deep learning studies often use unreliable data splitting and focus solely on accuracy.
• This limits the clinical applicability of current AI models for kidney CT analysis.

Purpose of the Study:

• To develop and validate KDH-Net, a hybrid deep learning framework for multiclass kidney disease characterization.
• To ensure robust performance evaluation using patient-level data splitting and advanced validation techniques.
• To enhance the reliability and interpretability of AI-driven kidney CT analysis.

Main Methods:

• KDH-Net integrates EfficientNetB0, ResNet50, and MobileNetV2 via feature-level fusion.
• A two-stage training strategy was employed for improved optimization stability.
• Patient-level evaluation, calibration analysis, and Grad-CAM were used for realistic assessment and interpretability.

Main Results:

• KDH-Net achieved 0.93 overall accuracy and a 0.91 macro-average F1-score on the patient-level dataset.
• The model demonstrated balanced performance across all kidney disease classes.
• Confidence analysis and Grad-CAM visualizations confirmed prediction reliability and highlighted relevant anatomical regions.

Conclusions:

• KDH-Net offers a stable, reliable, and interpretable framework for kidney CT characterization.
• The system provides trustworthy predictions under realistic evaluation conditions.
• KDH-Net is designed to support, not replace, clinical decision-making in nephrology.