Performance of Deep CNN and Radiologists in Prostate Cancer Classification: A Comparative Pilot Study

Abstract

In recent years multiple deep-learning solutions have emerged that aim to assist radiologists in prostate cancer (PCa) diagnosis. Most of the studies however do not compare the diagnostic accuracy of the developed models to that of radiology specialists but simply report the model performance on the reference datasets. This makes it hard to infer the potential benefits and applicability of proposed methods in diagnostic workflows. In this paper, we investigate the effects of using pre-trained models in the differentiation of clinically significant PCa (csPCa) on mpMRI and report the results of conducted multi-reader multi-case pilot study involving human experts. The study aims to compare the performance of deep learning models with six radiologists varying in diagnostic experience. A subset of the ProstateX Challenge dataset counting 32 prostate lesions was used to evaluate the diagnostic accuracy of models and human raters using ROC analysis. Deep neural networks were found to achieve comparable performance to experienced readers in the diagnosis of csPCa. Results confirm the potential of deep neural networks in enhancing the cognitive abilities of radiologists in PCa assessment.