Deep Learning for Classification of Solid Renal Parenchymal Tumors Using Contrast-Enhanced Ultrasound.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate classification of renal tumor subtypes is crucial for effective treatment planning.
• Contrast-enhanced ultrasound (CEUS) offers a non-invasive imaging modality for kidney tumor assessment.
• Deep learning (DL) has shown potential in medical image analysis but requires validation for specific applications like renal tumor subtyping.

Purpose of the Study:

• To evaluate the efficacy of deep learning models in classifying solid renal parenchymal tumor subtypes using CEUS images.
• To compare the classification performance of different deep learning architectures (ResNet-18 and RepVGG).

Main Methods:

• A retrospective analysis of 237 CEUS-scanned kidney tumors (angiomyolipomas, clear cell RCC, papillary RCC, chromophobe RCC) was performed.
• Two deep learning models, ResNet-18 and RepVGG-A0, were trained and validated for tumor subtype classification.
• Performance metrics including accuracy, AUC, sensitivity, and specificity were calculated. Class activation mapping (CAM) was used for visualization.

Main Results:

• The RepVGG-A0 model achieved a higher overall accuracy of 84.5% compared to ResNet-18's 76.7%.
• RepVGG-A0 demonstrated superior Area Under the Curve (AUC) values across all subtypes, with the highest for clear cell RCC (0.911) and angiomyolipoma (0.906).
• Both models exhibited reliable differentiation capabilities, with CAM highlighting predictive regions.

Conclusions:

• Deep learning models, particularly RepVGG-A0, can effectively classify renal tumor subtypes from CEUS images.
• These AI-driven approaches offer a potential objective and non-invasive method for improving renal tumor diagnosis and management.
• Further validation in larger cohorts is warranted to confirm clinical utility.