MSc Thesis: Medical Image Segmentation Topics
Hendrik Moeller - Feb 20, 2025
Automatic segmentation is an essential tool in medical imaging. From segmentation masks, one can localize and extract important structures and features from images, such as distinguishing multiple vertebrae in the spine.
I have a multitude of interesting projects to offer.
Topics:
Soft Segmentation: Soft segmentation is a subgroup of segmentation tasks that trains from non-binarized data and thus also outputs meaningful non-binarized segmentation masks. Research Question: How can binarized segmentation masks best be utilized for soft segmentation training? This would build upon a previous thesis, which had an interesting outlook. Can we minimize the volume error by scaling the input after binarization? Can we shape the softness however we want for training?
Label Masking and Pseudo Training Labels: When having 10 segmentation classes, is it better to have one sample with all 10 classes annotated or 10 samples with one class annotated each? This project explores this direction by incorporating label masking (similar to loss masking) and generating pseudo-labels during training on fully annotated classes to analyze the performance differences. Sources: https://arxiv.org/html/2403.01909v1
A possible benefit is to train two sequences (MRI and CT) that share some segmentation classes but not all of them and then train them in a way that transfers one class to the other sequence.
Resolution: The SOTA for handling different image resolutions is to resample an input dataset to a fixed resolution, train the model on that resolution, and then resample back to the individual input resolution. How do we find the optimal resolution for training? Is there a meaningful difference between model resolution performance and native resolution performance? Can we set the training resolution higher than is reflected in our data to account for the resampling error? What about the patch-size in relation to the resolution?
A different direction could be to simulate low-dimensional data by downsampling and utilizing softness or smoothing algorithms after upsampling again to try to make higher-resolution segmentation tasks more precise.
POI Prediction: A previous thesis used segmentation to predict points-of-interest on vertebra surfaces. Unfortunately, the method failed due to insufficient data quality. Now that we are equipped with revised and quality-ensured annotations, I want to see how well a deep learning approach can replace and subsequently speed up the point annotation process.
Analyzing Training Behavior for the Task of Segmentation: We have shown that by looking at the behavior of gradients during training, we can have solid interpretations of the same training cases and detect mislabels for classification tasks. This project will apply this to the task of semantic segmentation and find out whether this is possible and how well it works.
Extrapolation in segmentation training: Simple structures such as bones in CT are clearly visible and separable. In this project, we want to investigate if a model can learn from limited field-of-view annotations and extrapolate the segmentation to bones the model has never seen.
Prerequisites:
We are seeking a motivated Master's student in Computer Science or a related field. Prof. Daniel Rueckert will provide supervision with me being the advisor.
- Interest and motivation in machine learning and a topic above.
- Advanced knowledge of deep learning with imaging data.
- Good background in Python coding, using Pytorch (lightning).
- At least basic understanding and experience with segmentation masks.
- Beneficial: Experience with imaging data such as CT or MRI.
- Beneficial: Experience with topic-related field (training segmentation model, ...).
What we offer:
- Exciting topics with potential for publication
- Hardware for scientific computing
- Close and regular supervision (30min, once a week at least)
- Uncomplicated work environment
How to apply:
Send an email to Hendrik Moeller, with your CV, grade report, and a small introduction about you and your motivation. Please also specify which of the topics above interest you.
Beneficial: Start the subject of the mail with "[MA]" (then I will see it faster).
Preferred start date: Flexible (but rather sooner than later)