Whole-body positron emission tomography - computed tomography (PET-CT) imaging is routinely used worldwide in many clinical indications. Nuclear imaging is now a key procedure in the diagnosis, (re)staging, and assessment of treatment response in many cancer types and inflammatory diseases. However, due to the high cost and significant radioactive exposure, PET scans are only performed when there are strong indications, leaving their potential for early diagnosis of for example tumors still largely untapped. Especially children and pregnant women have very limited access to PET scans.
The aim of this master thesis is to investigate the use of adversarial learning for improving the quality of low-dose PET scans. Hereto, nuclivision has collected matching pairs of low-dose and high-dose PET scans of the same patient. Using supervised image-to-image techniques, deep learning models can be trained that take as input a low-dose PET scan and predict what the matching high-dose PET scan would look like. Traditionally, measures such as the RMSE or SSIM are used during training, but some successful cases of adversarial training have been reported. The aim of this master thesis project is to implement a baseline model using adversarial learning, where a discriminator model tries to distinguish enhanced images from real high-dose PET images. Then, in a second phase the student should replace the discriminator function with a “judge” network that can quantify the image quality of a PET image. All implementations should be done using the PyTorch framework.

This topic can also be chosen as or extended to a master thesis.

References:
1. https://arxiv.org/abs/2312.04382
2. https://arxiv.org/pdf/2108.10772v2.pdf
3. https://arxiv.org/pdf/2304.00451.pdf

Python
Deep Learning
Torch