Efficient and Precise Classification of CT Scannings of Renal Tumors Using Convolutional Neural Networks

Abstract

We propose a new schema for training and use of deep convolutional neural networks for classification of renal tumors as benign or malign from CT scanning images. A CT scanning of a part of the human body produces a stack of 2D images, each representing a slice at a certain depth, and thus comprising a 3D mapping. An additional temporal dimension may be added by injection of contrast fluid with CT scannings performed at certain time intervals. We reduce dimensionality – and thus computational complexity – by ignoring depth and temporal information, while maintaining an ultimate accuracy. Classification of a given scan is done by majority voting over the classifications of all its 2D images. Images are divided into training and validation sets on a patient basis in order to reduce overtraining. Current experiments with scans for 369 patients, yielding almost 20,000 2D images, demonstrate an accuracy of 93.3% for single images and 100% for patients.