Evaluation of Threshold and Gradient-Based PET Segmentation Algorithms using New Anthropomorphic XCAT Phantom with Scalable Lymphatic System: Application to Nodal Non-Hodgkin’s Lymphoma

Objectives: The Deauville 5-point scale (Deauville 5PS) is an internationally-recommended metric for routine clinical reporting in a number of lymphomas. Tracer uptake of each [¹⁸F]FDG-avid tumour is visually assessed relative to the mediastinum and liver using standardized uptake values (SUV), based on a semi-quantitative 1-to-5 scale. However, there is evidence that quantitative imaging metrics can enhance the prognostic value of PET and its ability to effectively guide treatment decisions. For example, total metabolic tumour volume (TMTV) is a well-documented predictor of therapy response and overall survival in lymphoma patients. However, TMTV accuracy can be impacted by the selected reconstruction parameters and segmentation algorithm. We aim to evaluate the quantification accuracy of lesion activity and volume while using different fixed-threshold and gradient-based segmentation methods. We will use a realistic lymphoma patient scenario simulated with a 4D-extended cardiac torso phantom (XCAT) and newly embedded lymphatic system. We focus on primary mediastinal B-cell lymphoma (PMBCL), which classically presents as large, heterogeneous tumour masses located within the mediastinum, though our work is also being extended to include diffuse large B-cell lymphoma (DLBCL). Methods: Ten PMBCL patients with heterogeneous mediastinal tumours were simulated using the XCAT phantom with an integrated lymphatic system. Organ activity concentrations were specified based on analysis of [¹⁸F]FDG PET/CT images from 5 PMBCL patients. Tumours with 4-100mL volumes were generated with activity concentration of 21.5 kBq/ml. This concentration was determined from an analysis of 10 PMBCL patient images using segmentations performed by a nuclear medicine physician. To create the tumours in the XCAT phantom, lymph node morphology and function was altered (i.e. lymph nodes were expanded, stretched asymmetrically, converged within the mediastinum, and heterogeneous activity concentration was established). Activity and attenuation files were generated using the XCAT phantom framework. PET data was generated using the MATLAB-based PET simulation and reconstruction tool (Ashrafinia et al., 2017), which simulated the GE Discovery RX scanner (OSEM with 2 iterations, 24 subsets) with multiple Poisson noise-realizations. Regions-of-interest were drawn using MIM (MIM Software Inc.) using 20%, 25%, 30%, 40%, 50% of SUV_max fixed threshold and MIM’s PET Edge+ gradient-based algorithm. Total metabolic tumour volume (TMTV) and total lesion glycolysis (TLG) were calculated for each of the tumours. Results: TMTV percent bias, obtained with a 25% fixed threshold, for the different tumour volumes was 4.4% (13mL), -0.7% (39mL), -11.4% (71mL), and -14.7% (100mL). Percent bias in TMTV with PET Edge+ was 22.1% (13 mL), 11.9% (39 mL), 26.3% (71 mL), and 20.5% (100 mL). For TLG, the 25% threshold, led to percent bias of -15.6% for the 100mL lesion and -8.5% at 13mL. PET Edge+ TLG percent bias was -5.0% at 13mL, and only deviated -1.6% for the 100mL lesion. Conclusion: Our results suggest that the 25% fixed-edge threshold provides better accuracy for TMTV quantification for PMBCL tumours. PET Edge+ overestimated tumour volume, but provided more accurate and consistent activity values. In a clinical setting, 25% threshold is recommended for accurate TMTV delineation, while gradient algorithms appear to be more suitable for determining TLG. The use of the XCAT phantom with the added lymphatic system also expands the possibilities for development of novel segmentation methods that lead to automated reporting of TMTV in lymphoma.

• Download figure
• Open in new tab
• Download powerpoint