Implementation of Convolutional Neural Network Architecture on 3D Multiparametric Magnetic Resonance Imaging for Prostate Cancer Diagnosis

TL;DR

This paper presents a novel two-stage multimodal CNN framework for automatic prostate cancer classification from multiparametric MRI, achieving high accuracy and outperforming existing methods, thus aiding diagnosis and reducing unnecessary biopsies.

Contribution

The study introduces a new multimodal multi-stream CNN architecture that simplifies preprocessing and improves classification performance in prostate cancer detection.

Findings

Achieved ROC AUC of 0.87 in classification

Outperformed most submitted methods in PROSTATEx Challenge

Demonstrated potential to assist in clinical diagnosis

Abstract

Prostate cancer is one of the most common causes of cancer deaths in men. There is a growing demand for noninvasively and accurately diagnostic methods that facilitate the current standard prostate cancer risk assessment in clinical practice. Still, developing computer-aided classification tools in prostate cancer diagnostics from multiparametric magnetic resonance images continues to be a challenge. In this work, we propose a novel deep learning approach for automatic classification of prostate lesions in the corresponding magnetic resonance images by constructing a two-stage multimodal multi-stream convolutional neural network (CNN)-based architecture framework. Without implementing sophisticated image preprocessing steps or third-party software, our framework achieved the classification performance with the area under a Receiver Operating Characteristic (ROC) curve value of 0.87. The result outperformed most of the submitted methods and shared the highest value reported by the PROSTATEx Challenge organizer. Our proposed CNN-based framework reflects the potential of assisting medical image interpretation in prostate cancer and reducing unnecessary biopsies.