An accurate deep learning-based computer-aided diagnosis system for early diagnosis of prostate cancer

This chapter introduces a system for early diagnosis of prostate cancer using a convolutional neural network (CNN) from diffusion-weighted imaging (DWI) acquired at seven b-values (100, 200, …, 700 s/mm2). The proposed system has three main steps. First, prostate segmentation is performed to separate the region of interest from the background. Prostate segmentation is done using a level-set model that depends on three features for enhanced accuracy. These features are intensity, shape prior, and spatial features. These three features are fused using a nonnegative matrix factorization approach. Second, the apparent diffusion coefficients (ADCs) of the segmented regions at each b-value are computed and used to train the classification model of the final step. Finally, a CNN is trained using the ADC maps to distinguish malignant subjects from benign ones. The DWI datasets used to evaluate the performance of the proposed system were collected from 45 subjects (20 benign and 25 malignant) at two different magnetic field strengths which are 1.5 and 3 Tesla (T). An average accuracy of 93.7%±3% is achieved.