Deep Learning for Detecting and Subtyping Renal Cell Carcinoma on Contrast-Enhanced CT Scans Using 2D Neural Network...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Renal cell carcinoma (RCC) detection and subtyping are crucial for patient management.
• Accurate differentiation between clear cell RCC (ccRCC) and non-ccRCC subtypes impacts treatment strategies.
• Computed tomography (CT) is a primary imaging modality for RCC evaluation.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) algorithm for RCC detection and subtyping using CT.
• To investigate the efficacy of a 2D neural network architecture with feature consistency modules for ccRCC versus non-ccRCC classification.
• To offer a computationally simpler and accurate DL approach for RCC characterization.

Main Methods:

• Retrospective analysis of CT scans from 196 histopathologically proven RCC patients (143 ccRCC, 53 non-ccRCC).
• Development and testing of 2D DL architectures, notably FocalNet-DINO, incorporating spatial and class consistency modules.
• Performance evaluation using metrics including recall, specificity, accuracy, F1 scores, and area under the curve (AUC).

Main Results:

• The FocalNet-DINO architecture achieved a high recall rate of 0.823 at 0.025 false positives per image (FPI) for RCC detection.
• Integration of consistency modules improved recall by 0.2% at 0.025 FPI and enhanced accuracy and AUC by 0.1% for RCC classification.
• The DL approach improved cancer detection in 21 slices and reduced false positives in 126 slices.

Conclusions:

• Deep learning algorithms utilizing 2D neural networks and feature consistency modules demonstrate high performance in RCC detection and classification.
• This approach offers a novel, computationally efficient, and accurate method for characterizing RCC subtypes on CT.
• The findings support the potential of DL for improving diagnostic accuracy in renal cell carcinoma management.