Deep Learning Auto-segmentation of Prostate Cancer from MRI data


Session type:

David Gillespie, Ian S. Boon, Connah Kendrick, Cheng S. Boon1, Moi Hoon Yap
1Clatterbridge Cancer Centre



Magnetic Resonance Imaging (MRI)-based prostate cancer radiotherapy workflow are increasingly explored with aims of improving prostate cancer outcomes and reduce late toxicities of treatment (1). MRI-only radiotherapy workflow requires daily adaptive radiotherapy planning and will add significant demands on oncology services. Artificial intelligence (AI) models using deep learning algorithms can be trained for radiotherapy auto-segmentation to aid MRI-only radiotherapy prostate workflow. 


This study uses publicly available MRI prostate cancer datasets from Promise 12, NIC ISBI, ProstateX database. Deep learning algorithms Mask RCNN, Detectron2 and U-Net were used for the auto-segmentation of prostate gross tumour volume (GTV) on MRI images.  We trained new AI models, validated and evaluated the performance of these AI models against ground truth.


The training model and validation were performed on a MR prostate dataset of 132 patients. Promise12 (n= 43), ISBN (n=25) and ProstateX (n=64) dataset was split into three groups with a 0.6, 0.2, 0.2 ratio for training, validation and test split respectively. 

Four deep learning AI networks were trained and validated on the publicly available MRI datasets consisting of 132 patients with prostate cancer. The trained models were able to achieve an average of 0.89 to 0.92 Dice Similarity coefficient (DSC).


This proof of concept study shows that new AI models using deep learning algorithms can be developed to produce prostate GTV contours with good similarity to ground truth provided. Future work will require collaboration of oncologists and  machine learning to develop novel metrics to assess clinically acceptability of contours prior to validation for potential clinical use.


1) Kerkmeijer LGW, Maspero M, Meijer GJ, et al. Magnetic Resonance Imaging only Workflow for Radiotherapy Simulation and Planning in Prostate Cancer. Clin Oncol (R Coll Radiol). 2018;30(11):692-701. doi:10.1016/j.clon.2018.08.009

Impact statement

Magnetic Resonance Imaging (MRI)-only prostate cancer radiotherapy workflow requires daily adaptive radiotherapy planning where oncology workload may be mitigated by application of deep learning algorithms for auto-segmentation of radiotherapy contours.