Prostate cancer detection from multi-institution multiparametric MRIs using deep convolutional neural networks

Multiparametric magnetic resonance imaging (mpMRI) of the prostate aids in early diagnosis of prostate cancer, but is difficult to interpret and subject to interreader variability. Our objective is to generate probability maps, overlaid on original mpMRI images to help radiologists identify where a cancer is suspected as a computer-aided diagnostic (CAD). We optimized the holistically nested edge detection (HED) deep convolutional neural network. Our dataset contains T2, apparent diffusion coefficient, and high $b$ -value images from 186 patients across six institutions worldwide: 92 with an endorectal coil (ERC) and 94 without. Ground-truth was based on tumor segmentations manually drawn by expert radiologists based on histologic evidence of cancer. The training set consisted of 120 patients and the validation set and test set included 19 and 47, respectively. Slice-level probability maps are evaluated at the lesion level of analysis. The best model: HED using $5\times 5$ convolutional kernels, batch normalization, and optimized using Adam. This CAD performed significantly better ( $p<0.001$ ) in the peripheral zone ( $\mathrm{AUC}=0.94\pm 0.01$ ) than the transition zone. It outperforms a previous CAD from our group in a head-to-head comparison on the same ERC-only test cases ( $\mathrm{AUC}=0.97\pm 0.01$ ; $p<0.001$ ). Our CAD establishes a state-of-the-art performance for predicting prostate cancer lesions on mpMRIs.