Amalgamating 3D Convolutional Neural Networks with Recurrent Neural Networks to Enhance Detection of Clinically Notable Prostate Cancer Using Multiparametric MRI Scans

Abstract

Prostatic adenocarcinoma is one among the prevalent cancers contributing to top causes of cancer deaths worldwide. Techniques of advanced imaging, like multiparametric magnetic resonance imaging, efficiently enhances the accuracy of prostatic adenocarcinoma detection. We take our cue from the huge success deep learning has had with medical image analysis by developing an automated pipeline combining RNN with 3DCNN for the identification of clinically significant PCa within whole-organ mpMRI scans. In this paper, we present a two-stage pipeline where the first stage uses 3DCNN for feature extraction from volumetric images, while the second stage uses an RNN for modelling sequential dependencies across slices. Our two-class dataset includes 500 patients; 200 have clinically significant PCa and 300 are free of PCa. A test set of 100 patients not used in training validated this model. At the slice level, this pipeline produced an AUC of 0.89 with a 95% confidence interval of 0.86–0.92, and at the patient level, it had an AUC of 0.86 with a 95% confidence interval of 0.81–0.91. The results have demonstrated that the integration of 3DCNN with RNN has very good potential for the clinical application of prostate cancer detection.