Introduction

Magnetic resonance imaging (MRI) provides a noninvasive assessment of the prostate that improves the detection of prostate cancer and can reduce unnecessary biopsies. The excessive variation in the performance and interpretation of MRI is, however, a major barrier to its widespread acceptance and use. In this study, we employed computer-aided diagnosis with a deep neural networks algorithm for prostate cancer detection using multiparametric MRI (mpMRI).

Methods

Between 2010 and 2015, 354 patients underwent extended systematic prostate biopsy with MRI-targeted biopsy (MTB), and 209 patients with negative mpMRI underwent systematic prostate biopsy (SB). All patients with PSA levels of less than 20 ng/ml and negative findings of digital rectal examination were included as subjects. mpMRI was interpreted by an experienced radiologist. For the supervised training using deep neural network architecture, we selected 163 mpMRI-positive patients who were diagnosed with prostate cancer by MTB and 135 mpMRI-negative patients who were not diagnosed with prostate cancer by SB. We chose representative T2-weighted and diffusion-weighted (DW) MRI images from both mpMRI-positive and mpMRI-negative patients. The 298 pairs of T2-weighted and DW images labeled as &[Prime]cancer&[Prime] or &[Prime]no cancer&[Prime] were randomly divided into 248 training and 50 test datasets, and the measure of diagnostic accuracy was calculated. The structure of the deep neural network model we used contains an input layer, three fully connected hidden layers, and an output layer (figure 1). The layers in the networks have ReLu non-linear activation units, and their learning rate was 0.01 for 10 epochs with a dropout ratio of 0.5.

Results

In the 50 hold-out validation test datasets, the mean area under the curve and the accuracy were 0.84 (0.72-0.99) and 0.81 (0.72-0.96). The mean positive predictive value, negative predictive value, sensitivity, and specificity of the algorithm were 0.89 (0.74-1.00), 0.75 (0.50-1.00), 0.76 (0.50-1.00), and 0.88 (0.18-1.00), respectively.

Conclusions

Computer-aided diagnosis with a deep neural networks algorithm for prostate cancer with mpMRI can provide reproducible interpretation and a greater level of standardization and consistency.

Funding

none