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. The interpretation of FDG PET-CT studies is a lengthy and complex process, which starts with looking at the maximum intensity projection (MIP) and scrolling through the entire volume in the three anatomical planes, in conjunction with the CT images. The normal biodistribution of FDG is well-known and so are the many normal variants, such as muscle, myocardial, and bowel uptake. Pathological patterns, e.g., inflammation, healing, benign tumors, must be recognized whether they are related to the disease or not. Such interpretation process may take a various amount of time, depending on the reader’s experience and the study complexity.
The aim of this study is to develop state-of-the-art tumor detection algorithm in PET images. Rather than relying on traditional techniques used in image segmentation, the aim is to explore loss functions and model architectures that have successfully been used in the most recent object detection papers. The final goal of this paper is to develop an algorithm that not only can predict if a patient has a tumor but also can give an indication of the region where the tumor is situated (e.g., with methods like Grad-CAM). The training data for this project consists of around 1000 matching PET and CT scans that can both be used for tumor detection, such that multi-modal techniques need to be explored. All implementations should be done using the PyTorch framework.

Python
Deep learning
PyTorch