Subject

Medical imaging with fluorine-18 fluorodeoxyglucose ([18F]FDG) positron emission tomography / computed tomography (PET/CT) is very sensitive and can be used to visualize disease burden in many different types of cancer. However, physiological background activity may hamper the detection of metabolically active lesions.

For cancers that can metastasize anywhere in the body, like melanoma, the brain in particular is problematic due to its high glucose consumption. Though the brain is included in the whole-body PET/CT imaging acquired at UZ Brussel and brain metastases are an important biomarker for survival, a volumetric measurement can currently not be extracted. Brain metastases may appear as hypo- or hypermetabolic on a PET image but may also be indistinguishable from the background activity. Patients suspected of having brain metastases will typically receive magnetic resonance imaging (MRI) to confirm. However, this modality is less available and more expensive. Detecting and segmenting brain lesions on the whole-body [18F]FDG PET/CT image may be preferred as every melanoma patient receives this exam anyway and automated assessment of the brain may lead to early detection.

Kind of work

A dataset of 69 melanoma patients has been collected with annotations on brain metastases provided as a binary indicator at image level. The goal of the project is to investigate if unsupervised and weakly supervised anomaly detection methods [2-4] can be used for detection and segmentation of brain metastases on whole-body [18F]FDG PET/CT images. Ultimately, this will lead to a better estimation of a patient’s tumour burden and survival chances.

Framework of the Thesis

The developments will be performed as an extension to existing software developed within the ETRO research group. The algorithms will be implemented in Python, using open- source image processing and machine learning.
The project will involve:

• Literature study.


• Implementation of machine learning methods for anomaly detection.


• Implementation of image processing methods for extraction of the volume of brain metastases.


• Training and application of the tools for different patient sequences.


• Thesis writing.

References
[1] I. Dirks, M. Keyaerts, B. Neyns, I. Dirven, and J. Vandemeulebroucke. "Development and Validation of a Predictive Model for Metastatic Melanoma Patients Treated with Pembrolizumab Based on Automated Analysis of Whole-Body [18F]FDG PET/CT Imaging and Clinical Features". Cancers, 15(16):4083, 2023. doi: https://doi.org/10.3390/cancers15164083.
[2] Chen, X., Konukoglu, E.: ‘Unsupervised detection of lesions in brain MRI using constrained adversarial auto-encoders.’ arXiv preprint arXiv:1806.04972 (2018)
[3] Zimmerer, D., Kohl, S.A., Petersen, J., Isensee, F., Maier-Hein, K.H.: Contextencoding variational autoencoder for unsupervised anomaly detection. arXiv preprint arXiv:1812.05941 (2018)
[4] Wolleb, J.: ‘Diffusion models for medical anomaly detection.’ Medical Image Computing and Computer Assisted Intervention–MICCAI 2022: 25th International Conference, Singapore, September 18–22, 2022, Proceedings, Part VIII. Cham: Springer Nature Switzerland, 2022.

Number of Students

1

Expected Student Profile

Following a MSc in a field related to Biomedical Engineering or Applied Computer Science - Digital Health.