Subject

Micro-calcifications (MCs) are an important indicator in the context of breast cancer. The resolution of in vivo CT-scans does not allow identifying them in early growth stages: detection and segmentation (delineation) occur at the limits of the resolution. For in vivo imaging the spatial distribution of the MCs is one of the main parameters. Nevertheless, several publications mentioned also morphological characteristics as discriminating features between malignancy and benign cases, even if the MCs are only a few voxels in size. Since segmentation and defining shape of small voxel volumes is inevitably heuristic in nature, results suffer from robustness and are usually not replicable.

At VUB-ETRO we collaborate with the Radiology and Anatomopathology departments in UZ-Jette, and hence we were able to reconstruct in vitro biopsies with micro-CT scanners at a high resolution (in collaboration with the University of Antwerp). By taking tissue biopsies, the spatial structure of the 3D distribution of MCs is disturbed, but we are in a good position to analyse shape and even internal texture. Two of our studies have already been published in top journals [1,2]. The original approach proposed in this master thesis is based on other types of standard image processing and classification approaches but has never appeared in the literature before for MC analysis. Therefore, we target a new publication.

Kind of work

We first look for some characteristic seed points in the 3D micro-CT scans (e.g., maximum intensity values, or the centroid of an initial MC segmentation). If we take the centroid, then there is only one seed point for each (pre)segmented MC. On the contrary, usually there will be multiple maximum intensity points within one microcalcification and hence multiple seed point. We start up a region growing model (e.g., connected component growth) starting from the seed points.

In case of a single seed point, the main research question is finding an acceptable stop criterion on top of the detection of region merging with another MC region.

In case there are multiple seed points within one initially segmented MC, the previous method becomes a little bit more complex because we will detect the iteration number where growing regions from the same initial MC segmentation are merging and continue to grow as one new region. We can then construct a graph of this type of merges for further analysis. Assuming the growth is a proxy of the real physical growth, both the growth of the volume and the time dependent (iteration dependent) curves of several classical morphological features (and eventually also texture features) can be extracted and serve as extra features for the classification.

The novelty lies in the intelligent combination of techniques. The student will need high analytic and programming skills, he/she will familiarise himself with the nature of 3D medical images, generic classical image processing techniques, goal oriented algorithmic development and programming, understanding performance metrics. Together, we will design an algorithmic integrated processing flow dedicated to the specific application of MC analysis.

Besides region growing we will also use another classical segmentation technique from the world of mathematical morphology: the watershed transform. This approach gives the same type of outputs as region growing. Hence, we will be able to compare both.

The evaluation will be made with classical metrics comparing the similarity of segmentations and of course the characterisation of classification in benign/malignant MCs.

For the classification we will use off-the shelve machine learning approaches and use them in first instance as black boxes

Framework of the Thesis

The developments will be performed, building on previous know-how.

Project will consist of:

· (1) Thorough literature study (2) Situate the approach among other approaches used for MC segmentation and classification and make the link between in vivo mammography and in vitro biopsy scans (e.g., methodologies, resolution, data sets, evaluation metrics). (3) In parallel the candidate must start exercising his programming skills (e.g., Python, SimpleITK, Sklearn, Pandas, Numpy, Bokeh, ETE toolkit, etc.

· (4) Familiarisation with data access and Hydra computing environment.

· (5) Study and understand the algorithms that will be used.

· (6) Development of the first workflows including quantitative and qualitative validation of obtained results.

· (7) Batch implementation of the workflow/s.

· (8) Optimization of the workflows by iterative improvements applied to the method and additional feedback from supervisors.

Number of Students

1

Expected Student Profile

(1) Very good programming (Python) and analytic skills
(2) Previous experience with: 3D medical images, generic classical image processing techniques, classification algorithms.