Synthetic Q-Space Learning With Deep Regression Networks For Prostate Cancer Characterisation With Verdict

Abstract

Traditional quantitative MRI (qMRI) signal model fitting to diffusion-weighted MRI (DW-MRI) is slow and requires long computational time per patient. Recently, q-space learning utilises machine learning methods to overcome these issues and to infer diffusion metrics. However, most of q-space learning studies use simple multi layer perceptron (MLP) for model fitting, which might be sub-optimal when estimating more complex diffusion models with many free parameters. Previous works only investigate the application of q-space learning on diffusion models in the brain. In this work, we explore q-space learning for prostate cancer characterization. Our results show that while simple MLP is adequate to estimate parametric maps on simple models like classic VER-DICT, deep residual regression networks are needed for more complex models such as VERDICT with compensated relaxation (R-VERDICT).

Differentiating False Positive Lesions from Clinically Significant Cancer and Normal Prostate Tissue Using VERDICT MRI and Other Diffusion Models

“…The creation of the training set and training of the DNN (which was performed only once) took approximately 200 s (1.1 GHz Dual-Core Intel Core M processor). Prediction of the trained DNN for the whole unmasked DW-MRI dataset (roughly 5 × 10 5 voxels) took approximately 50 s for each subject [ 27 , 36 , 43 ]. The image analysis pipeline is shown in Figure 1 b.…”

“…We fitted the diffusion models to the VERDICT-MRI data and obtained the ADC from the mp-MRI. The model fitting procedure used deep neural networks (DNNs) for ultra-fast and robust parameter estimation [ 27 ]. We compared parameter estimates between both false positives and clinically significant cancer and normal tissue and false positives using statistical tests.…”

Differentiating False Positive Lesions from Clinically Significant Cancer and Normal Prostate Tissue using VERDICT MRI and Other Diffusion Models

“…The DNN was trained using synthetic data, which has been proven to achieve equivalent robustness to real data for deep learning model fitting [42]. We generated 100,000 syn- [27,36,43]. The image analysis pipeline is shown in Figure 1b.…”