Keywords: prostate cancer, multi-parametric MRI, ordinal regression, prostate cancer detection, prostate cancer grading, deep learning, U-Net
TL;DR: Simultaneous detection and Gleason Grade Group scoring of prostate cancer lesions directly from multi-parametric MRI using soft labels.
Abstract: The goal of this work is to detect prostate cancer (PCa) from multi-parametric MRI (mpMRI) and to simultaneously predict the Gleason Grade Group (GGG) of the detected tumors. We used the ProstateX-2 dataset, for training, validation and testing. The challenge training set contains 99 patients and 112 lesions. The challenge test set contains 63 patients and 70 lesions. T2-weighted and apparent diffusion coefficient images were used as input for a U-Net model. For each tumor a GGG was assigned based on biopsy and pathologic analysis. Segmentation maps of the tumors that were multiplied with a scaled value of the GGG, were used as target of the network, turning the problem into a binary soft-label ordinal regression problem. Using 5-fold cross validation, a voxel-wise quadratic-weighted kappa score of 0.391 ± 0.062 and a DSC (GGG ≥ 2) of 0.321 ± 0.039 were achieved. In order to evaluate using the challenge test set, the voxel-wise predictions were converted into a single GGG prediction per lesion. Our method ranks higher than 30 out of 43 participants of the challenge with a lesion-wise quadratic weighted kappa score of 0.082 ± 0.272 when evaluating on the ProstateX-2 test set. Despite solving a more difficult, but more clinically relevant, problem than the original ProstateX-2 challenge, a relatively high score was achieved using an approach to both grade and detect prostate cancer directly from mpMRI.
Code Of Conduct: I have read and accept the code of conduct.