Automation and quantification in Pelvic Magnetic Resonance Imaging

PelvicMRISegmentation.png ProstateMRDeepLearningT2Texture.png

Pelvic magnetic resonance imaging (MRI) is used in the diagnosis of prostate, bladder, rectum, cervix, pancreas cancer and various other diseases such as pain due to adhesions. MRI excels in the depiction of the many soft-tissue structures in the pelvis (bladder, rectum, prostate, cervix, intestine, fat, muscle). Pelvic MRI reading requires good expertise and time to properly read all the images in an imaging study and may contain information yet unexplored for diagnostic use. Research in automation and quantification of pelvic MRI tries to reduce the expertise level and reading time and make readily available all information for diagnostic use.

Automation of Prostate MRI is a prominent research project. Prostate MRI has been shown to help reduce the current number of biopsies by 50% and also find 10-20% of the cancers earlier in a stage where curative treatment is very feasible. We have been exploring automated tools to help read prostate MRI. Deep learning allows us to improve conventional image analysis methods and to explore approaches that were infeasible before. We, therefore, have several projects related to the application of Deep Learning to prostate MRI.

As examples, a recent master project resulted in a convolutional neural network (CNN) deep learning system that can analyze T2-weighted prostate MR (see image on top). It was trained to reproduce the PIRADS score of a radiologist. The system was able to successfully identify suspicious regions. Another example project used a 3D U-Net to segment anatomical structures in pelvic MRI (see image on top). This was a pilot project showing feasibility.

Available student projects

  • Automated pelvic MRI segmentation. Deep learning (DL) models exist that can segment the prostate with an overall DICE of 0.80. This is not enough. Many zonal areas are still incorrectly segmented. The segmented boundary of the prostate is still too inaccurate for diagnostic and interventional use. We continue to seek improvement. Example research projects are: Publication?bibkey=Geld18, Mooi18.
  • Normalization of MRI. MRI scalar values are vendor, sequence, coil and position dependent. Normalized MRI can be used for quantitative analysis. We have been successful in using multireference tissue technology to normalize prostate MRI. First manually, and recently completely automatically. This has until only been shown to work on t2-weighted images. A stronger improvement is expected if this is shown for all sequences (e.g. DWI, DCE, T1). Furthermore, a multivariate model is likely to outperform a univariate model. Finally, this can be extended to any MRI application.
  • Detection of prostate cancer in MRI. We have DL models detecting cancer. Yet we still miss cancers, why? How can this be improved? More data, smarter DL? Adding normalization? Adding segmentation?

We have several project grants under review. Ambitious students may be able to apply for a Ph.D. position.

A large database of ten-thousand annotated pelvic/prostate MR is available from the Computer aided detection of prostate cancer database. You will be able to build upon existing code for the analysis of prostate MR and extensive knowledge in our group on deep-learning.

Institute

The work will be executed in the Diagnostic Image Analysis Group (DIAG) of the Radboud University Nijmegen Medical Centre. DIAG is a leading research group in computer-aided detection and diagnosis (CAD). Radboudumc has one of the most internationally renowned groups in prostate MR radiology.

Tasks

  • Familiarize yourself with deep learning within DIAG.
  • Research and develop a DL model and algorithms for a specific task in pelvic MRI image analysis
  • Write paper

Requirements

  • Good grades
  • Working knowledge of deep learning and medical imaging
  • Writing skills

Information

  • Project duration: 6-12 months (can be tuned to your requirements)
  • Supervision: Henkjan Huisman
  • Location: Radboud University Medical Centre, Nijmegen, The Netherlands
  • For more information please contact Henkjan Huisman, henkjan.huisman@radboudumc.nl.