Delineation of tumors and reference regions in PET

AI-based tumor segmentation

AI-based approaches for automated tumor delineation allow for a quick, precise, and user-independent (and thus reproducible) segmentation of tumor lesions in PET image data. In addition, the application of automated AI algorithms enables rapid translation of new quantitative approaches into a clinical context by accelerating image evaluation procedures and making them feasible within a clinical time constraints.

In this context, we developed a novel convolutional neural network (CNN) that allows for a fully automatic delineation of metabolically active tumor volume in [18F]-FDG-PET/CT scans of patients with head and neck squamous cell carcinoma (HNSCC). This CNN is also able to differentiate between primary tumors and lymph node metastases. It uses a modified 3D U-Net architecture and includes a custom-built multi-head self-attention module for improved context awareness. Since amount and diversity of datasets are crucial for optimal performance of deep-learning-based methods, we used 698 patient scans from 3 different sites and 5 public databases for the network training. The effectiveness of the developed AI method was verified on a sizeable external dataset proving its generalizability. Ground truth delineations for the network training and testing were interactively generated by two experienced physicians as medical expertise is a curial requirement for producing high-quality training data which emphasizes the importance of collaborations with clinical partners.

We were also able to show that manual and automated AI-based delineations can be used interchangeably for quantitative image analysis: the metabolic tumor volume (a quantitative PET parameter) derived from both of them possesses similar predictive value with respect to the overall survival of HNSCC patients. Thus, our developed CNN is a capable tool able to massively accelerate and facilitate study data evaluation in large patient groups which also does have clear potential for supervised clinical application.

Our further developments in this field are focused on broadening the spectrum of different cancer types with neural networks for Non-small cell lung cancer and non-Hodgkin's lymphoma delineation being currently under development. The final goal of this project is the development of a universal AI tool which is able to handle all cancer types and is publicly available.

• Nikulin et al., EANM’24 Proceedings, EJNMMI 2024
• Nikulin et al., EJNMMI 2023

AI-based reference region delineation

Quantitative PET image analysis often requires a reference value of tracer concentration alongside the target uptake. Such reference values are usually extracted from a suitable reference region. For example, pharmacokinetic analysis of tumor metabolism relies on the knowledge of the arterial input function – the concentration of radiotracer in the patient’s blood as a function of time. Tracer concentration in the blood can be measured directly via arterial blood samples. However, such a sampling procedure is invasive, discomforts patient handling, thus is too cumbersome for routine clinical use. Alternatively, the blood signal can be extracted from PET images themselves by defining a Region-Of-Interest (ROI) within a large blood vessel such as an aorta or the heart. The non-invasiveness of this image-derived approach allows to simplify the workflow and improves patient comfort. Nevertheless, special care has to be taken to assure numerical accuracy of these quantitative PET image measurements. For example, placement of the ROI has to be performed carefully while adhering to a set of rules making the process time-consuming and error-prone.

To solve this problem, we developed a neural network to automate this image-based blood pool ROI delineation. As ground truth, manually defined ROIs within the aorta lumen were used while ensuring that these ROIs are not affected by partial volume effect (signal degradation due to limited resolution of PET), motion-induced artefacts, or spill-over from the nearby highly active regions (such as inflammations, heart, and tumor lesions). The developed CNN model was able to replicate the manual delineation process even for challenging cases. It was tested on 580 unseen patient scans from multiple clinical centres and showed a high degree of concordance to manual delineations in terms of extracted blood signal and showed an extremely low failure rate (< 1%). Therefore, the proposed AI-based solution allows to accelerate image processing workflows for procedures which require determination of blood tracer concentration. These, along with the aforementioned pharmacokinetic analyses, include the SUR computation. The CNN-based blood signal determination method simplifies the SUR-based analysis to a degree that might allow its use as a superior drop-in replacement for SUV-based quantitation without increasing the clinical workload.

Quantitative analysis of amino acid brain PET requires a different definition of a reference ROI. Here, uptake in healthy brain tissue is used as a comparison point to delineate tumor lesions or to assess therapy response. The set of rules based on which the delineation of the healthy background ROI is done is loosely defined and the actual practice varies between different clinics. This introduces severe inter-observer variability making it difficult to compare data from different clinical sites. To unify the procedure, we have developed a CNN which is able to delineate these background ROIs automatically. It was trained to reproduce the reference tracer concentrations and, interestingly, it was able to derive delineation rules on its own without ever seeing the reference ROIs themselves.

• Nikulin et al., EJNMMI 2020