KidneyNeXt: A Lightweight Convolutional Neural Network for Multi-Class Renal Tumor Classification in Computed...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Renal tumors present diagnostic challenges in CT imaging due to overlapping features.
• Manual interpretation of CT scans is time-consuming and prone to inter-observer variability.
• Automated classification systems are needed for accurate and efficient renal tumor diagnosis.

Purpose of the Study:

• To develop and evaluate KidneyNeXt, a custom convolutional neural network (CNN) for multi-class renal tumor classification.
• To assess the model's performance on diverse computed tomography (CT) datasets.
• To provide a reliable automated tool for supporting early and accurate diagnosis of kidney tumors.

Main Methods:

• A custom CNN architecture, KidneyNeXt, was designed with multi-branch pathways and hierarchical feature extraction.
• Transfer learning using ImageNet 1K pretraining was applied for enhanced generalization.
• The model was evaluated on three distinct CT datasets: a clinical dataset, Kaggle CT KIDNEY, and KAUH: Jordan.

Main Results:

• KidneyNeXt achieved high accuracy across all datasets, exceeding 99.7% on the clinical and KAUH datasets, and 99.9% on the Kaggle dataset.
• The model demonstrated robustness against class imbalance and inter-class similarity.
• Grad-CAM visualizations were used to interpret model predictions, highlighting regions of interest.

Conclusions:

• KidneyNeXt is a lightweight and effective deep learning solution for classifying renal tumors from CT images.
• The model's consistent high performance suggests potential for real-world clinical deployment as a decision support tool.
• Future research may involve integrating clinical metadata and multimodal imaging for improved diagnostic precision.