Combining 3D Convolutional Neural Networks with Recurrent Neural Networks to Improve the Accuracy of Clinically Significant Prostate Cancer Detection Using Multiparametric MRI Scans

Prostatic adenocarcinoma is ranked amongst the most widespread cancers which rank high in the list of cancer deaths in the world. Advanced imaging techniques in medicine such as multiparametric magnetic resonance imaging have improved the chances of identifying the prostatic adenocarcinoma. This has been achieved in the context of deep learning successfully applied to medical images by constructing an automated image analysis system, which consists of RNN and 3D convolutional neural networks to detect clinically significant PCa in the entire volume of mpMRI images. This paper contributes to this line of work by proposing a two-stage pipeline where 3DCNN is used in the first stage for the extraction of features from volumetric images and an RNN in the second stage for slice interdependency modelling. For our two-class dataset, 500 patients are included, out of them, 200 patients have clinically significant PCa and 300 patients are without PCa. A test set of 100 patients not used in training was employed to further validate the model. At the slice level, the pipeline achieved an AUC of 0.89 with a 95% confidence interval of 0.86-0.92 while at the patient level, the pipeline performed with an AUC of 0.86 and 95% confidence interval of 0.81-0.91. This study has brought forth evidence that prostate cancer detection using 3DCNN-RNN hybridized model is clinically feasible.