The Diagnostic Image Analysis Group (http://www.diagnijmegen.nl) at the Department of Radiology, Radboud University Nijmegen Medical Centre, is offering a PhD position.
The Diagnostic Image Analysis Group is a research division of the Department of Radiology of the Radboud University Nijmegen Medical Centre. Nijmegen is the oldest Dutch city with a rich history and one of the liveliest city centers in the Netherlands. Radboud University has over 17,000 students. Radboud University Nijmegen Medical Centre (RUNMC) is a leading academic centre for medical science, education and health care with over 8,500 staff and 3,000 students.
The focus of the Diagnostic Image Analysis group is the development and validation of novel methods in a broad range of medical imaging applications. Research topics include image analysis, image segmentation, machine learning, and the design of decision support systems. Application areas include breast, brain, lung, retina and prostate imaging. Key to the success of the group is close cooperation with clinical partners and a disease oriented approach. Currently the group consists of twenty researchers including thirteen PhD students and support from four scientific programmers.
Detection of breast cancer in mammograms is the most researched computer-aided detection (CAD) application. It is also the only commercially successful application of CAD. In the United States, health insurance companies reimburse screening organizations for using CAD and this has created a large market for mammography CAD. The majority of mammograms in the US are nowadays read with CAD. Several companies are active in this area, but despite intense research over the past twenty years, the detection of masses and architectural distortions in mammograms by computers still lags behind human experts.
More and more screening programs are adopting fully digital mammography equipment and this has resulted in very large (and continuously growing) digital databases of mammograms that could be used to improve CAD systems. The goal of this project is to investigate if novel, state-of-the-art techniques in pattern recognition and machine learning, in combination with very large training databases, can be used to further improve CAD systems up to and beyond the level of a human expert.
The starting point for this project is an advanced CAD system for masses and architectural distortions in mammograms and a large digital database taken from the Dutch breast cancer screening program. This reference CAD system has been developed over the past twenty years within the Diagnostic Image Analysis Group by Dr. Nico Karssemeijer and co-workers. See Earlier detection of breast cancer by computer assisted decision making in screening for some background information. This system has a candidate lesion detector that we intend to use unaltered. For each candidate a set of about 60 features is computed. A neural network is used for classification and trained with a database of scanned films. Using the reference CAD system and the new digital database a training set of around 1,000 proven cancers and around 100,000 false positive lesion candidates will be compiled. Our previous research suggests that with the current feature set and classifier, only limited performance increase can be obtained by retraining the CAD system on this larger training database and that potential for performance gain by using other classifiers is also limited. The first step in the project is therefore to implement a number of novel sets of features, mostly based on existing literature.This will increase the number of features to around 500. The main part of the project will be to investigate novel classification strategies, for example random forest classifiers, that can deal with this type of machine learning problem: a large number of features (102-103), a large skewed training database (~103 positive examples and ~106 negative examples) and a difficult classification problem. The goal is to improve CAD performance compared to the reference CAD system. Data collection and observer studies to measure CAD performance stand-alone compared to that of experienced radiologists are part of the project.
In a later stage the developed method will be applied to other important CAD applications. These will include nodule detection in chest CT and chest X-ray. For these applications we expect data sets of similar size to be available.
You should be a creative and enthusiastic researcher with an MSc degree in Computer Science, Physics, Engineering or Biomedical Sciences or similar, with a clear interest in pattern recognition, machine learning and medical image analysis applications. Good communication skills, and expertise in software development, preferably in C++, are essential.
You are appointed as a PhD student you will get the standard salary and secondary conditions for PhD students in the Netherlands. Your performance will be evaluated after 1 year. If the evaluation is positive, the contract will be extended by 3 years. The research should result in a PhD thesis.
This position is funded by the NWO VIDI project Towards Intelligent Machines: Design of Dynamic Computer-Aided Diagnosis Systems. You will work at DIAG and will closely collaborate with the Machine Learning Group of Prof. Dr. Tom Heskes and various clinical collaborators. Dr. Nico Karssemeijer and Dr. Bram van Ginneken will be promotors. Dr. Clarisa Sanchez will be your daily supervisor and co-promotor.
For more information please contact Bram van Ginneken by e-mail.
Send applications as a single pdf file to firstname.lastname@example.org. This pdf file should contain your CV, a letter of motivation, a list of followed courses and grades and preferably a reprint of your Master thesis or any publications in English you have written.