KIDNEY CANCEROUS TUMOR PREDICTION USING CNN SYSTEM ARCHITECTURE

Abstract

The study addresses the challenge of interobserver variability in the treatment decisions for kidney cancers, a concern highlighted by the anticipated 73,750 kidney cancer diagnoses in the United States in 2020. This variability often arises due to the subtle differences in imaging characteristics of tumor subtypes. To address this issue, we propose an endto-end deep learning model leveraging multi-phase CT scans to differentiate between five primary histologic subtypes of kidney cancers, encompassing both benign and malignant tumors. The proposed model demonstrates remarkable precision in identifying kidney cancers, even those of minimal size. In preparing the data for analysis, we divided it into training and validation test sets. The training set was used to employ the random forest method for ranking potential predictors based on their predictive importance. The model’s performance was then validated on the test set using leave-one-out cross-validation. This study utilized convolutional and recurrent neural networks to predict kidney cancer outcomes. We used the models to classify adenoma, adenocarcinoma, and non-neoplastic whole slide images (WSIs). The evaluation of our models was conducted using three distinct test sets. The results showed area under the curve scores of 0.97 and 0.99 for distinguishing between cancerous tumors and adenomas and 0.96 and 0.99 for differentiating between kidney cancer and adenomas, respectively. These findings suggest that our models are not only generalizable but also hold significant potential to integrate and deploy into realistic pathological diagnostic workflows of kidney cancer.