Abstract

Purpose: According to the statistics, up to 15-20% of removed solid kidney tumors turn out to be benign in postoperative histopathological examination, despite having been identified as malignant by a radiologist. The aim of the research was to limit the number of unnecessary nephrectomies of benign tumors.

Methods or Background: We propose a machine-aided diagnostic system for kidney tumor malignancy identification that can serve as a second opinion source to prevent potentially unnecessary surgical procedures. CT images of a cross-section of a tumor were used to identify tumor malignancy. The system utilizes a deep convolutional neural network as a feature extractor. The network was trained on approximately 15,000 CT images of arterial phase coming from 349 solid kidney tumors. We introduced medical image colorization to improve the knowledge transfer, as the network was pretrained on natural RGB images. We used subtype max-pooling to map predicted histopathological subtype to malignant/benign classification, to improve the network’s accuracy over binary classification.

Results or Findings: The algorithm achieved F1-score of 90% and accuracy of 87% for the identification of benign and malignant tumors. Achieved specificity of 100% was essential to the study, as the goal of the research was to help reduce the number of tumors misdiagnosed as malignant. We discovered that image colorization can improve accuracy up to 10 pp.

Conclusion: We present a machine-aided diagnostic system to classify malignancy of renal tumors. Our method has proved to be a valuable source of second opinion assisting in the decision whether or not nephrectomy should be performed. Limitations: We are planning to explore how the colorization affects images coming from other domains.

Ethics Committee Approval: Ethics committee approval was not required.

Funding for this study: "e-Pionier - wykorzystanie potencjału uczelni wyĪszych na rzecz podniesienia innowacyjnoĞci rozwiązaĔ ICT w sektorze publicznym"

Citations

Authors (7)