Prostate cancer detection based on deep convolutional neural networks and support vector machines: a novel concern level analysis

In this paper, four different strategies are explored for the classification task. Both ResNet50 and VGG-16 are utilized and re-trained to classify our diffusion-weighted magnetic resonance imaging (DWI) database to clarify the existence of prostate cancer (PCa) or not. Transfer learning and data augmentation are applied for ResNet50 and VGG-16 to solve the problem of lack of tagged data and increase system efficiency. The last fully connected layer is replaced by the Support Vector Machine (SVM) classifier to achieve a better accuracy. Both transfer learning and data augmentation are performed for SVM to increase the performance of our framework. In addition to, a k-fold cross validation is applied to test our models performance. Our proposed techniques are trained and evaluated on a given DWI dataset that involves 1765 patients of which 845 are with PCa and 920 are without. This paper employs end-to-end fully convolutional neural networks without any prepossessing or post-processing. The proposed technique based on ResNet50 hybridized with SVM achieves the best performance with 98.79 % accuracy, 98.91 % area under the curve (AUC), 98.43 % sensitivity, 97.99 % precision, 95.92 % F1 score and a computational time of 2.345 s.