Ensemble of Convolutional Neural Networks for the Detection of Prostate Cancer in Multi-parametric MRI Scans

Abstract

Prostate MP-MRI scan is a non-invasive method of detecting early stage prostate cancer which is increasing in popularity. However, this imaging modality requires highly skilled radiologists to interpret the images which incurs significant time and cost. Convolutional neural networks may alleviate the workload of radiologists by discriminating between prostate tumor positive scans and negative ones, allowing radiologists to focus their attention on a subset of scans that are neither clearly positive nor negative. The major challenges of such a system are speed and accuracy. In order to address these two challenges, a new approach using ensemble learning of convolutional neural networks (CNNs) was proposed in this paper, which leverages different imaging modalities including T2 weight, B-value, ADC and Ktrans in a multi-parametric MRI clinical dataset with 330 samples of 204 patients for training and evaluation. The results of prostate tumor identification will display benign or malignant based on extracted features by the individual CNN models in seconds. The ensemble of the four individual CNN models for different image types improves the prediction accuracy to 92% with sensitivity at 94.28% and specificity at 86.67% among given 50 test samples. The proposed framework potentially provides rapid classification in high-volume quantitative prostate tumor samples.