Vendor equalization with adversarial learning in breast mammograms

Breast cancer is the most frequent cause of cancer-related death in women worldwide. The past decades have seen a steady increase in cancer incidence, but not in mortality due to improved treatment and screening. In order to further decrease the mortality rate it is paramount to detect breast cancer in an as early stage as possible. In this project we will work with a large annotated dataset from the Dutch breast cancer screening program where the images are produced on different full-field digital mammographs (FFDM) by different vendors.

An important challenge in the automated detection of breast cancer in mammograms is to achieve a comparable performance on images produced by different machines. In mammography it is important to detect very small structures which makes it difficult to transfer the models to images generated on machines with a different resolution and noise profile. Secondly, the collection of labelled data is expensive, certainly in the medical domain.

In this project we will work with a large expert annotated mammography database made in a breast screening setting on one of our mammographs. Our aim is to produce a system which achieves state-of-the-art results in the detection of breast cancer lesions in mammographic images, and not only on images produced by the same machine but also on unlabelled data produced by different machines.

This project is of particular interest to students who want to obtain skills in deep learning.

References

1. Kooi, T., Litjens, G., van Ginneken, B., Gubern-Mérida, A., Sánchez, C. I., Mann, R., … Karssemeijer, N. (2017). Large scale deep learning for computer aided detection of mammographic lesions. Medical Image Analysis, 35, 303–312. https://doi.org/10.1016/j.media.2016.07.007

2. Tzeng, Eric, et al. "Adversarial discriminative domain adaptation." arXiv preprint arXiv:1702.05464 (2017) https://arxiv.org/abs/1702.05464.

Institute

The work will be executed in a high level environment for medical image analysis, at the Diagnostic Image Analysis Group (DIAG) of the Radboud University. DIAG is a leading research group in computer-aided detection and diagnosis (CAD).

Tasks

  • Do a literature research on existing methods or similar approaches, including the algorithms developed in our group [1-2].
  • Implement neural network to detect and classify calcifications
  • Validate the detection and classifications by comparing the automatic detection to the annotations made by experts.
  • Write papers and/or thesis report.

Requirements

  • Students with a major in computer science, biomedical engineering, physics, (applied) mathematics, physics, artificial intelligence or any related area in the final stage of master level studies are invited to apply. PhD candidates who want to work on this project as an intern are also welcome to apply.
  • Interest in deep learning
  • Affinity with programming.

Information