Quantitative parameters for the evaluation of oncological PET
Automatic detection of tumor boundaries in PET data
Accurate tumor delineation in PET image data is essential for the integration of PET in the planning of radiation and in the assessment of therapy response. Tumors are quite often delineated by manually drawing of the tumor boundaries. However, manual delineation is prone to substantial interobserver variability, as the limited resolution and statistical accuracy of PET data make it difficult to for the observer to clearly identify the actual object boundary. The large interobserver variability of manual delineation can be especially problematic when evaluating follow up studies by different observers. In addition, purely manual delineation is inherently tedious and time-consuming and thus inappropriate for routine use.
We have therefore developed an algorithm for tumor delineation which only requires a sub-volume of the PET image containing one or more tumors. The algorithm then iteratively determines the tracer concentration in the background of the tumor and finally applies a background adjusted threshold in percent of the maximum concentration of the tumor. This leads to robust results as long as the tracer distribution within the tumor is not strongly heterogeneous. We have therefore extended the algorithm to also precisely delineate such metabolically heterogeneous tumors. In this extension, a threshold value adapted to the background is calculated for each tumor-suspected voxel, whereby the background and the maximum value are only computed in the close vicinity of the voxel. This local method is slower than the global method (delineation time: 10s to 1min, depending on tumor and voxel size), but is also able to precisely delineate very heterogeneous tumors.
However, both algorithms cannot be used to delineate lesions with low or defuse tracer uptake. Therefore, we have extended our research to AI-based lesion delineation. So far we have trained and validated a neural network for the delineation of head and neck lesions. We are also currently developing and evaluating AI-based networks for the delineation of lung cancer and lymphoma.
All algorithms are essential parts of the PET software ROVER (see below). They are used for tumor delineation by researchers in Dresden and by many ROVER users in preclinical and clinical research.
Key publications:- Nikulin et al., EJNMMI 50, 2023
- Hofheinz et al., Med Phys 40 (8), 2013
- Hofheinz et al., Nuklearmedizin 51 (1), 9-16, 2012
Accurate quantification of tumor metabolism
A high metabolism of cancer cells is an indication of fast tumor growth and can therefore provide important information for therapy planning and therapy outcome. PET allows the quantification of tumor metabolism, but the absolute quantification is very time-consuming and limited to a small proportion of patients. Therefore the the easier to determine semi-quantitative parameter standardized uptake value (SUV) is typically used in clinical routine as surrogate parameter for tumor metabolism. However, SUV quantification has well-known limitations and correlates rather weak with the actual metabolic rate.Therefore, we have recently introduced the parameter standardized uptake ratio (SUR), which is the ratio of tumor SUV and blood SUV, corrected for variations in the uptake time (the time between injection of the radiotracer and PET imaging). The Tumor SUV can be determined, for instance, using one of the delineation algorithms mentioned above. For the determination of the blood SUV, we have trained a neural network which determines the SUV in the descending aorta (see here).
We were able to show that the SUR correlates much better with tumor metabolism than the SUV, which translates into a better prediction of therapy outcome in patients with esophageal cancer, lung cancer, and cervical cancer.
The figure below shows an example of outcome prediction in the case of esophageal cancer where a baseline PET and a restaging PET were acquired and evaluated. Patients were treated with definitive radiochemotherapy.
Shown are Kaplan-Meier curves with respect to event free survival (EFS) for normal tissue reaction on radiation, defined as changes of SUV from baseline PET to restaging PET in the non-tumor affected esophagus(left), for tumor SUR derived in the restaging PET (middle) and for the combination of both (right). Both parameter are significant prognostic factors for EFS with notable effect size. The combination of both parameters further increases the effect size. On the other hand tumor SUV did not show a significant effect in this group of patients.
In ongoing studies, we are currently investigating weather the SUR is also suitable for better predicting the outcome of treatment in patients with other typed of cancer.
Key publications:- Cegla, Hofheinz, et al., Sci Rep 13 (1) 2023
- Zschaeck, Hofheinz, et al. IJC 147 (5) 2020
- Hofheinz et al., EJNMMI 46, 2019
- Hofheinz et al., EJNMMI research 6 (1), 1-9, 2016
- van den Hoff et al., EJNMMI research 4 (1), 18, 2014
- van den Hoff et al., EJNMMI research 3 (1), 77, 2013
Tumor shape as a predictor of therapy outcome
Intratumoral spatial variation of cellularity, angiogenesis, necrosis, etc. are typical for aggressive tumors and, if known, can have influence on therapy planning. Since such variations lead to a heterogeneous distribution of tumor metabolism, they can in principle be quantified with PET by analyzing the distribution of the tracer uptake rate inside the tumor. However, the limited spatial resolution of PET leads to strong partial volume effects which strongly affects the measured tumor heterogeneity.
On the other hand, it can be hypothesized that the variation of cellularity etc. may also result in an irregular shape of the region of interest that delineates the increased tracer uptake. In order to characterize such irregular variation of the shape of a lesion, we have introduced the parameter tumor asphericity (ASP). ASP describes the deviation from a spherical shape and can be computed as following:
where S is the surface area of the tumor and V its volume. For spherical geometries the ASP is zero. For all other 3D geometries, ASP is larger than zero and quantifies the relative increase in surface area of a 3D structure compared to the surface area of a sphere with same volume.
We were able to show that ASP has the potential to predict therapy outcome in various tumor diseases (head and neck cancer, esophageal cancer, neuroendocrine tumors, neuroblastoma, cervical cancer, and prostate cancer) and thus provide important information for therapy planning.
The figure below shows an example of ASP-based outcome prediction in the case of nasopharyngeal cancer, where only a baseline PET was acquired and evaluated. Patients were treated with definitive radiochemotherapy.
Shown are Kaplan-Meier curves with respect to the overall survival (OS) for the metabolically active tumor volume (left), for tumor ASP (middle) and for the combination of both (right). Both parameter are significant prognostic factors for OS with notable effect size. The combination of both parameters further increases the effect size.
We are currently investigating if these results can be confirmed in large patient groups.
Key publications:- Andela, Hofheinz et al., Rad Onc 19 (1) 2024
- Zschaeck, Hofheinz et al., Sci Rep 13 (1) 20840, 2023
- Cegla, Hofheinz et al., Sci Rep 13 (1) 8423, 2023
- Zschaeck, Hofheinz et al., PLOS one 15 (7) 2020
- Hofheinz et al., Eur J Nucl Med 42 (3), 429-437, 2015
- Apostolova, Hofheinz et al., Eur Radiol 24 (9), 2077-2087, 2014
Translation to clinical research and application
Image evaluation algorithms and new PET parameters are integrated into the software ROVER as soon as possible. In this way a fast translation into preclinical and clinical research and finally into clinical application is achieved. For example all above mentioned research results are based on evaluations of the corresponding PET data performed with ROVER.
ROVER is an evaluation program designed for the fast interactive visualization, analysis and easy data and image export of tomographic image data. It combines several functionalities related to the efficient evaluation of these data with the emphasis on (human and small animal) PET. Nevertheless many features are also useful with other imaging modalities (e.g. CT, MRI).
Development of ROVER has started in 2002 by our department. In 2005 ROVER became commercially available and is distributed by our cooperation partner ABX GmbH, Radeberg, Germany. Currently about 50 institutions world wide are using ROVER for preclinical and clinical research.
References: