Semi-supervised segmentation of blood vessels in synchrotron μm-CT scans

Abstract

This masters thesis handles segmentation of blood vessels in Wistor-rat lungs, scanned by a µm-CT scanner. These scans are very detailed 3D-volumes and have been rescaled to a more manageable size of 1024 × 1024 × 1024 voxels per volume.

The main goals of the thesis is to find a more efficient way to perform segmentation in medical images and thus semi-supervised segmentation techniques have been explored.

The most effective method found is comprised of two parts, first a rough segmentation performed by the classic image segmentation technique FMM which is then used as the masks for training an ML-model. This ML-model is an Attention ResUnet. The FMM is performed in 3D and requires seed points, from which it propagates to create the vessel-segmentation. The ML-model is then fed 2D-slices from the images and masks. For the prediction, the volumes are fed through the
ML-model from the x-, y-, and z-axes ensuring that the final prediction takes the 3D-structure into account.

Due to the lack of ground truth masks it is difficult to make a quantitative evaluation of the results. However by studying the segmentations a qualitative analysis shows that this method gives good results, both in regard to the quantity of identified vessels and the quality of the segmentation of said vessels. Furthermore, the method significantly improves upon segmentations created through classic methods. The method also proves to be much more time efficient than
manual segmentation.

The code used for this thesis is presented in a digital appendix where the pre-trained weights can be used for segmentation of blood vessels, or to train a new model.

Popular Abstract

Technology and its application is a cornerstone of modern medicine. An area of particular interest is image analysis and modelling. Leveraging modern tools and methods can increasingly improve, and automate medical practice. Using these, medical practices can examine subjects in greater detail and speed, further improving healthcare. This is however not easily achieved, where data, practices, or technology can hamper success. An area with several of these thresholds is segmentation of medical images.

An accurate segmentation enables modelling of the object, making foundations for 3D models and other useful visuals. In order to accomplish these visuals, one has to either apply image analysis methods which leverage gradients and pixel intensities, or utilise deep-learning-algorithms. Both of these directions have specific requirements for an accurate segmentation.

Methods based on image analysis often require clean images with limited noise. One such method is the fast-marching method (FMM) which propagates from manually placed points, in directions with low intensities and gradients. This makes FMM suitable for segmenting simple and clean structures, however its lack of complexity means it has difficulties capturing finer details. This limitation is a general trend for these image based methods. This is an issue in the medical context, as it often is in the details that information can be extracted.

Machine learning (ML) methods do not have the same limitations as the image based methods. Instead of propagating from placed points, the ML learns structures by updating weights in a function to capture the features of a region of interest (ROI). This process relies on the labels to give enough "examples" of the ROI to achieve a representative function. Without these labels, the ML-algorithm is not provided enough information to learn what is, and is not part of a certain class. This infers a early stage limitation, as creating these labels are incredibly time-consuming.

The available segmentation methods can, even with their respective limitations, efficiently be deployed on simpler medical images. However, for more complex images these can be too resource demanding or simply lack performance. Segmentation of lung tissue is a prime example of where these limitations are yet to be overcome. This thesis tackles the case specific problems that occur when segmenting blood vessels in medical images of rat lungs.

The main problems in this case are the CT-scans’ complex and noisy nature, combined with the vessels’ open ended characteristics. This means that along with a difficulty in differentiating blood vessels from alveoli and background, the open borders at the end of the vessels cause issues in image analysis.

In order to mitigate these issues a two-step approach can be applied, leveraging strengths of classic image segmentation and ML-techniques. Using FMM, a rough segmentation can be performed which is sufficient for the next step. The results from this can then be used as a mask, semi-automating the annotation process. Even with the low bar to clear for this segmentation, constraints need to be made to increase the robustness and segmentation quality. The output from this Constrained-FMM produces an accurate enough segmentation to be used as a mask for an ML-algorithm’s training.

Using the FMM’s simplicity, several masks can be created cheaply. These can then be used to train an ML-model, which can learn enough to be able to capture a larger majority of the blood vessels. These identified vessels are not only many more compared to the FMM-segmentation, but also of higher quality. By combining each of the methods strengths, the model can segment with greater accuracy and efficiency whilst being cheaper in operation.

• author: Svedberg, Adam and Petersson, Sebastian
• supervisor: Ted Kronvall
• organization: Mathematical Statistics
• course: FMSM01 20251
• year: 2025
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics; Technology and Engineering
• keywords: Machine Learning, ML, Attention ResUnet, deep learning, neural network, CNN, Fast Marching Method, FMM, Constrained-FMM, image segmentation, medical segmentation, CT, blood vessel, lungs
• publication/series: Master's Thesis in Mathematical Sciences
• report number: LUTFMS-3530-2025
• ISSN: 1404-6342
• other publication id: 2025:E77
• language: English
• id: 9202547
• date added to LUP: 2025-06-24 11:40:31
• date last changed: 2025-06-24 11:40:31