Deep Learning Algorithm for Fully Automated Detection of Small (≤4 cm) Renal Cell Carcinoma in Contrast-Enhanced...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Renal cell carcinoma (RCC) is frequently discovered incidentally via abdominal computed tomography (CT).
• Early detection of small RCCs (≤4 cm) is crucial for effective treatment.
• Automated detection methods can improve diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for fully automated detection of small RCCs (≤4 cm).
• To assess the algorithm's performance using a multicenter dataset of contrast-enhanced CT images.

Main Methods:

• Retrospective selection of contrast-enhanced CT images from 7 centers (January 2005-May 2020).
• Development of a 2-step deep learning algorithm involving kidney and tumor segmentation.
• Internal validation using 453 patients (dataset A) with 10-fold cross-validation.
• External validation using 132 patients (dataset B).

Main Results:

• The algorithm achieved high performance metrics on both internal and external validation datasets.
• Accuracy: 88.3% (internal), 87.5% (external).
• Sensitivity: 84.3% (internal), 84.8% (external).
• Specificity: 92.3% (internal), 90.2% (external).
• Area Under the Curve (AUC): 0.930 (internal), 0.933 (external).

Conclusions:

• The deep learning algorithm demonstrates high accuracy, sensitivity, specificity, and AUC for detecting small RCCs.
• The algorithm shows potential for contributing to the early diagnosis of small renal cell carcinomas.
• The multicenter validation supports the algorithm's generalizability and clinical utility.