The accurate detection of prostate cancer is an interesting but still challenging task. Multiparametric MRI is the essential method for diagnosing prostate cancer. Over the past two decades, computer-aided diagnosis (CAD) has been a rapidly growing and active area of research. The purpose of this study is to evaluate the performance of a CAD algorithm for identifying prostate cancer in multiparametric MRI.

The prototype CAD algorithm using convolutional neural network has been developed with 1,500 cases data set for training and test with 5-fold cross validation method (4 training and 1 validation sequentially). This CAD algorithm performed auto segmentation of prostate gland and detection of prostate cancer. Then it merged to show result visualization as probability map for prostate cancer. The diagnostic performance of CAD algorithm for identifying prostate cancer from MRI was performed. 352 external validated multiparametric MRIs (T2-weighted, diffusion-weighted, and apparent diffusion coefficient images, b value of 0, and 1000 s/mm²) at two institutions were enrolled for this study. All data were acquired on 3T MRI scanners with pelvic phased-array coils. Using pathologic diagnosis after radical prostatectomy whole mount specimen as gold standard or ground truth, the...

...read more

The diagnostic performance of all prostate gland was a sensitivity of 71.1%, specificity of 82.1%, positive predictive value of 73.1%, negative predictive value of 80.6%, and accuracy of 77.7%.

CAD algorithm in multiparametric prostate MRI using convolutional neural network can have compatible diagnostic performance for prostate cancer. Comprehensive further studies for different experienced radiologists to evaluate the performance with/without CAD algorithm are needed.

In view of its compatible performance, we believe CAD algorithm in multiparametric MRI using convolutional neural network can be used as a useful supplementary tool in patients with suspicious of prostate cancer.