Semi-Supervised Segmentation and Grading of Prostate Cancer Lesions from Multiparametric Magnetic Resonance Images

Prostate cancer is the third most common cancer in Hong Kong and the most common non-skin cancer in the United States. Multiparametric MRI (mpMRI) is currently the workhorse for assessing patients with suspected prostate cancer and MRI-targeted biopsy is a standard way for diagnosing prostate cancer. Lesion risk assessment and segmentation are the most important tasks in mpMRI interpretation. Triaging patients based on mpMRI assessment reduces the number of unnecessary biopsies. Lesion segmentation is necessary for patients selected for MRI-targeted biopsy and focal therapies, which target localized lesions and spare patients from aggressive treatments. The current manual workflow for mpMRI interpretation is time-consuming and the analysis result depends on the observer’s experience. More importantly, although radiologists can assess lesions with mpMRI, Gleason grading can only be done reliably through biopsy. Therefore, an automated, non-invasive method for grading and segmenting lesions accurately will provide a leap in patient management. A multi-task convolutional neural network is proposed for automated lesion grading and segmentation. The framework allows collaboration between the grading and segmentation components to improve the performance in both tasks. The framework is developed specifically for fusing different mpMRI modalities at each layer. Propagation of fused feature maps enriches features available to deeper layers and is expected to improve segmentation/grading performance. This proposal also addresses the difficulty of obtaining sufficient images with manual segmentation for fully-supervised learning. A semi-supervised, adversarial lesion segmentation framework is proposed to allow training with a combined set of data consisting of images with and without manual segmentation. The evaluation network drives the segmentation network towards producing similar segmentation results for images with and without manual segmentation, thereby improving the performance on both image sets. The diagnostic role of dynamic contrast-enhanced (DCE) imaging is becoming questionable, with recent investigations showing that it does not improve diagnostic accuracy. Removal of DCE from mpMRI would save imaging time, cost and eliminate patient risks associated with the use of contrast agent. The performance of the proposed framework with and without DCE will be compared in this project. The proposed framework increases the reproducibility and efficiency of mpMRI interpretation, and will have a significant impact on cancer diagnosis, treatment selection and planning. The semi-supervised framework allows calibration for small differences in MR parameters, thereby making the network trained in one centre transferrable to other medical centres/clinics. This automatic interpretation tool will especially benefit medical clinics in which mpMRI expertise is less available.