Patch-Based Convolutional Neural Network for Differentiation of Cyst From Solid Renal Mass on Contrast-Enhanced Computed Tomography Images

• Main Authors: Fatemeh Zabihollahy, N. Schieda, E. Ukwatta
• Format: Article
• Language: English
• Published: IEEE 2020-01-01
• Series: IEEE Access
• Subjects: Renal mass; benign cyst; malignant; convolutional neural network
• Online Access: https://ieeexplore.ieee.org/document/8952718/

Automated classification of renal masses detected at computed tomography (CT) examinations into benign cyst versus solid mass is clinically valuable. This distinction may be challenging at single-phase contrast-enhanced CE-CT examinations, where cysts may simulate solid masses and where renal masses are most commonly incidentally detected. This may lead to unnecessary and costly follow-up imaging for accurate characterization. In this paper, we describe a patch-based CNN method to differentiate benign cysts from solid renal masses using single-phase CECT images. The predictions of the network for patches extracted from a manually segmented lesion are combined through the majority voting system for final diagnosis. We used a dataset comprised of single-phase CECT images of 315 patients with 77 benign (oncocytomas, and fat poor renal angiomyolipoma) and 238 malignant (renal cell carcinoma including clear cell, papillary, and chromophobe subtypes) tumors. We trained our proposed network using patches extracted and artificially augmented from 40 CECT scans. The presented algorithm was evaluated using 275 unseen CECT test images consisting of 327 renal masses by comparing algorithm-generated labels to those labeled by experts and achieved mean accuracy, precision, and recall of 88.96%, 95.64%, and 91.64%. Our method yielded accuracy of 91.21% ± 25.88% as mean ± standard deviation at the patient level. The AUC was reported as 0.804. The results indicate that our algorithm may accurately characterize benign cysts from solid masses with a high degree of accuracy and may be clinically valuable to prevent unnecessary imaging follow-up for characterization in a proportion of patients.