Non-Hodgkin lymphoma (NHL) is a heterogeneous group of hematologic malignancies that frequently presents as disseminated disease, with tumor deposits scattered across nodal and extranodal sites throughout the body. This widespread pattern makes quantifying total tumor burden especially challenging and especially important. The metabolic active tumor volume (MATV), derived from 18F-fluorodeoxyglucose positron emission tomography (18F-FDG PET) scans, has emerged as a strong prognostic biomarker in NHL, with multiple studies demonstrating its ability to predict treatment outcomes, progression-free survival (PFS), and overall survival (OS).
FDG-PET and the Deauville scale: 18F-FDG PET/CT imaging is the clinical standard for staging and response assessment in NHL. The Deauville five-point scale compares FDG uptake in diseased areas to background uptake in the mediastinum and liver, but this visual scoring is a coarse, semi-quantitative approach with documented poor inter-reader reliability. Contrast illusion from varying background activity can bias visual assessments, which motivates the shift toward quantitative MATV-based metrics as more objective biomarkers.
The segmentation problem: To calculate MATV, tumor regions on PET images must be delineated (segmented). Manual segmentation is precise but prohibitively time-consuming when a single NHL patient may have a dozen or more discrete tumor lesions. Semi-automated approaches using standardized uptake value (SUV) thresholds offer a more objective and scalable alternative. In SUV thresholding, all voxels with an uptake value above a defined cutoff are classified as tumor. The challenge is that no consensus method exists: different thresholds produce dramatically different MATV values, yet most existing software evaluates only one threshold at a time.
This paper addresses that gap by reviewing the literature on NHL segmentation methods and introducing the MUST-segmenter (MUltiple SUV Threshold segmenter), an open-source tool that applies nine different segmentation methods to a PET scan simultaneously, producing a full spectrum of MATV estimates in a single workflow pass.
The study combined a systematic literature review with an original software development and validation effort. The PubMed literature search was conducted on August 1, 2022, querying a combination of terms covering lymphoma types (lymphoma, NHL, DLBCL, MCL), imaging modalities (PET, FDG, SUV, standardized uptake value), segmentation terminology (segment, delineate, contour, volume measurement), and tumor burden metrics (MATV, MTV, tumor volume, tumor burden). Only studies published between 2010 and 2022 with full-text availability were considered.
Inclusion criteria: Studies were required to meet all five inclusion criteria: clinical studies, NHL patients, FDG PET-based disease evaluation, segmentation performed with multiple semi-automated SUV thresholding methods, and explicit evaluation or comparison between methods. Title and abstract screening was followed by full-text review for borderline cases. Of the 123 initial results, 12 studies were ultimately selected for the review.
Patient cohort for MUST-segmenter validation: To test the software, the authors assembled a use-case cohort of 68 patients with Large B-cell lymphoma (LBCL), the most common NHL subtype, treated at the University Medical Center Groningen (UMCG) in the Netherlands. These patients had received CD19-directed chimeric antigen receptor (CAR) T-cell therapy as third-line treatment. Two pre-infusion PET/CT scans were acquired per patient, approximately 28 days and 6 days before CAR T-cell infusion, yielding 110 total PET scans for analysis. The patient cohort had a median age of 61 years (range 20-79), was predominantly male (69.1%), and was heavily pretreated, with 58.8% at Ann Arbor stage IV and 72.1% having extranodal involvement.
Statistical analysis: MATV differences across the nine segmentation methods were assessed using paired t-tests adjusted for multiple testing via the Benjamini-Hochberg procedure. Distributions were visualized with boxplots using 5th to 95th percentile whiskers. Correlation plots of log-transformed MATVs were created across all method pairs. Low- and high-MATV groups were defined using cut-offs from the literature (200 cc and 500 cc) and compared using Cochran's Q test and McNemar's tests. All analyses were performed in R v4.2.1.
The MUST-segmenter was developed as a Python-based plugin for 3D Slicer, a widely used open-source medical image computing platform. The tool accepts PET and CT images in DICOM format, converts PET data to SUV maps, and presents axial, coronal, and sagittal views for navigation. Its key innovation is the ability to extract MATV using multiple SUV thresholding methods simultaneously from a single segmentation workflow, eliminating the need to re-segment each patient for every threshold approach. The source code is publicly available at GitHub (kyliekeijzer/Slicer-PET-MUST-segmenter).
Workflow steps: The segmentation pipeline follows five stages: (1) loading PET/CT imaging data alongside pre-defined avoidance regions such as brain, bladder, kidneys, liver, and heart contours obtained from Mirada's Embrace Atlas Contouring tool; (2) user identification of NHL lesions by clicking seed points at the center of each tumor on the PET/CT images; (3) semi-automatic region-growing segmentation using a 6-connected voxel neighborhood that expands outward from each seed point until adjacent voxels no longer meet the SUV threshold criterion; (4) visualization of resulting contours overlaid on PET images and in 3D view; and (5) extraction and export of MATV values.
Type 1, fixed thresholds: Three absolute SUV cutoffs were implemented: SUV2.5, SUV3.0, and SUV4.0. All connected voxels from the seed point meeting or exceeding the fixed absolute SUV value are classified as tumor. These methods are common in clinical research and are straightforward to interpret, but they do not account for inter-patient variability in background uptake or scanner calibration.
Type 2, relative thresholds: Four relative methods were implemented. The 41suvMax method uses 41% of the maximum SUV value within user-defined bounding boxes around lesions, a threshold recommended by the European Association of Nuclear Medicine (EANM) for solid tumors. The liverSuvMax method uses the maximum SUV in a 1.2 cm diameter sphere placed in the right lobe of the liver as the threshold. The PERCIST method uses 1.5 times the average liver SUV plus 2 standard deviations, derived from the same liver sphere. The SUVbrain method uses 41% of average brain SUV with an age-correction factor, acknowledging that brain metabolic activity declines with age (correction = (patient age - 20) x 0.125 / 50). Two majority-voting methods (MV2 and MV3) classify voxels as tumor if they are identified as such by two or more, or three or more, of the fixed and relative methods respectively.
The 12 included studies covered NHL patient cohorts ranging from 12 to 239 patients (mean 91, median 95). All 12 included NHL patients, with three additionally including Hodgkin lymphoma (HL) cases. DLBCL was the dominant subtype studied. Six studies reported EANM Research Ltd. (EARL) accreditation status for their PET scanners, with two including both EARL-accredited and non-accredited images to test reconstruction robustness. The most commonly used segmentation software across studies was ACCURATE, followed by PLANET Onco, Hermes, and MATLAB-based tools.
Most investigated thresholds: The SUV2.5 fixed threshold was used in 9 of 12 studies, making it the most common method. The 41suvMax relative method appeared in 10 of 12 studies. SUV4.0 appeared in 5 studies. In contrast, liver-based thresholds were rarely investigated (liverSuvMax in 1 study, PERCIST in 3 studies). Majority-voting approaches appeared in 5 studies. Adaptive thresholding (AT) algorithms that incorporate background intensity were tested in 8 studies, and one study compared thresholding methods against deep learning segmentation.
Predictive performance findings: Five studies evaluated how well MATV calculated with different thresholds predicted OS and PFS. Two studies found significant associations between MATV and OS across all methods tested (SUV2.5, 41suvMax, liverSuvMax, PERCIST). Three studies found significant MATV-PFS associations across all methods. For OS and PFS, three studies concluded that SUV2.5 achieved the best predictive performance, while one study favored 41suvMax for PFS. Notably, studies recommending SUV2.5 did not include SUV4.0 in their evaluation; when SUV4.0 was included, it was frequently preferred. Optimal MATV cut-offs to separate low- and high-risk patients varied considerably across studies, ranging from 295 to 552 cc for PFS using SUV2.5.
Robustness and clinician-based evaluations: Four studies evaluated inter-observer robustness, with two recommending SUV2.5 and one each recommending SUV4.0 and liver-based methods. Two studies assessed robustness across PET image reconstruction methods: one found SUV4.0 best aligned across EARL and high-resolution reconstruction, while another found no significant difference. Clinician quality-score assessments in three studies consistently recommended SUV4.0, with two of those also recommending majority-voting methods (MV2, MV3). Three studies reported high linear correlations among different thresholding methods overall.
Applying all nine segmentation methods to 110 PET scans from the 68 LBCL patient cohort revealed strikingly large variation in MATV estimates. Median MATV values ranged from 35 cc (41suvMax) to 211 cc (SUVbrain) across methods. Nearly all pairwise comparisons between methods yielded statistically significant differences (paired t-test p less than 0.05), with only two exceptions: SUV4.0 and PERCIST were not significantly different from each other, and neither were liverSuvMax and MV2, suggesting these pairs produce clinically similar tumor volume estimates.
Method-specific variability: The largest interquartile range was observed for SUVbrain (IQR 64.56 to 501.55 cc), reflecting the sensitivity of this age-corrected brain-referenced threshold to individual variation in cerebral metabolism. The smallest IQR was found with 41suvMax (IQR 10.62 to 86.72 cc), indicating this method produces the most consistent volume estimates but also the smallest volumes. As expected, fixed thresholds showed a monotonic inverse relationship: higher fixed SUV cutoffs produce smaller defined tumor volumes. This pattern was also reflected in the majority-voting methods, since MV3 (requiring agreement from more methods than MV2) tends to include fewer voxels.
Correlation structure: Log-transformation of MATVs was applied to establish approximate normality before computing pairwise correlation plots. Fixed thresholding methods (SUV2.5, SUV3.0, SUV4.0) were strongly correlated with each other, and majority-voting methods followed a similar pattern. The 41suvMax method showed the greatest divergence from all other methods in the correlation analysis, consistent with its consistently low MATV estimates. The boxplot distributions of liverSuvMax and MV2 were visually near-identical, as were those of SUV4.0, PERCIST, and MV3, suggesting functional equivalence within these groupings for this NHL cohort.
Zero MATV cases: Some patients recorded MATV values of zero for the SUV3.0 and SUV4.0 methods, meaning that no lesion voxels exceeded those fixed thresholds. This reflects cases where tumors had low metabolic activity, a finding consistent with the clinical reality that not all NHL lesions are highly FDG-avid, particularly after prior treatment.
Beyond the absolute MATV values, the clinical impact of the segmentation choice was assessed by examining how different methods affect patient classification into low-MATV and high-MATV risk groups. Two MATV cut-offs commonly found in the NHL literature were applied: 200 cc and 500 cc. At the 200 cc cut-off, the proportion of patients classified as high-MATV ranged from just 5% (41suvMax) to 51% (SUV2.5). At the 500 cc cut-off, the range was 1% (41suvMax) to 25% (SUVbrain). For both cut-offs, Cochran's Q test confirmed highly significant differences in patient proportions across methods (p less than 0.0001).
Pairwise comparisons: McNemar's tests for pairwise comparison of patient proportions between methods showed significant differences between nearly all method pairs at both cut-offs. The 41suvMax method diverged most substantially from all others, consistently producing the smallest number of high-MATV patients. This has direct clinical consequences: if a clinical study or treatment algorithm uses a 200 cc high-MATV threshold, the proportion of patients identified as high-risk would differ by up to 46 percentage points depending solely on the choice of segmentation method.
Why 41suvMax differs so much in NHL: In solid tumors, 41suvMax was recommended by the EANM because tumors tend to have high, relatively homogeneous FDG uptake. In NHL, however, tumor FDG uptake is highly heterogeneous: some lesions show intense focal uptake while others are only mildly avid. In a heterogeneous tumor, the maximum SUV can be very high, making 41% of that maximum a stringent threshold that excludes large portions of the tumor periphery. A previous study by Driessen et al. found the opposite pattern in classical Hodgkin lymphoma, where 41suvMax produced the largest MATVs, because lower overall SUV values in HL result in a lower absolute threshold and therefore more included voxels.
These findings make clear that MATV cut-offs derived from one segmentation method cannot be applied to another method without recalibration. Studies using different thresholding approaches are therefore not directly comparable, and meta-analyses pooling MATV-based risk estimates across studies using different methods should be interpreted with caution.
No consensus from the literature: Despite reviewing 12 published studies, no single SUV thresholding method emerged as uniformly preferred across all evaluation criteria. Studies recommending SUV2.5 for predictive performance did not include SUV4.0, which was consistently preferred when evaluated. Studies using quality scores recommended SUV4.0. The best-performing method also appears to depend on the intended clinical application, with different thresholds potentially optimal for prognostic stratification versus radiation therapy target delineation. This means the search for a universally optimal method remains unresolved.
Small patient cohorts in the literature: The median sample size across the 12 reviewed studies for survival modeling was 118 patients, which is modest for developing and validating prognostic models. Small cohorts increase the risk of overfitted models whose performance estimates may not generalize to broader NHL populations. The variability in recommended methods across studies likely reflects, at least in part, statistical noise from underpowered comparisons.
3D Slicer learning curve: Although the MUST-segmenter itself is open-source and requires no coding experience, it runs within 3D Slicer, which has a non-trivial interface and requires familiarization before effective use. For clinicians or researchers without prior 3D Slicer experience, onboarding time could limit practical adoption. Seed placement for patients with widely disseminated disease remains cumbersome, as each tumor site requires a separate seed point click.
No inter-observer variability assessment: The MUST-segmenter validation was performed without evaluating inter-observer variability, meaning the reproducibility of seed placement and the resulting MATV estimates across different users was not quantified. While SUV-based methods are more objective than manual segmentation, seed point placement and bounding box adjustment still introduce some user-dependent variability. This limitation is acknowledged as relevant for future studies. Additionally, the segmentation methods do not incorporate CT information, meaning necrotic low-uptake tumor regions may be missed.
Expanding the MUST-segmenter: The open-source architecture of the MUST-segmenter was designed to accommodate additional thresholding methods as the field evolves. Future iterations could incorporate CT-based tumor boundaries to capture low-FDG-avid regions such as necrotic tumor cores. The current SUV thresholding approach segments only metabolically active tissue, which is well-suited for prognostic applications but may underestimate anatomical tumor extent relevant for radiation therapy planning. Combining PET and CT information in a multimodal segmentation framework, potentially using deep learning, could address this gap.
Deep learning segmentation: The authors note that deep learning approaches for automated PET segmentation are increasingly promising. Fully automatic convolutional neural networks, such as the U-Net and nnU-Net architectures applied to FDG PET/CT, have shown strong performance for lymphoma lesion detection and segmentation. Incorporating deep learning options into the MUST-segmenter or parallel workflows would allow direct comparison between threshold-based and learning-based MATV estimates, which is needed to determine whether deep learning can resolve the inter-method variability documented in this study.
Larger multicenter validation: The predictive performance of different segmentation methods needs to be evaluated in larger, multi-institutional NHL cohorts to generate reliable recommendations. The MUST-segmenter is positioned to facilitate this type of research, since it dramatically reduces the time required to extract MATV across multiple thresholds. For a patient with 12 tumor lesions, the MUST-segmenter completed segmentation with 3 SUV thresholds in approximately 6 minutes, compared to 29 minutes with the ACCURATE software, representing an approximately fivefold efficiency gain.
Other image biomarkers: Beyond MATV, the MUST-segmenter can extract additional PET image biomarkers from segmented tumor volumes, including radiomics features capturing tumor heterogeneity, total lesion glycolysis, and dissemination metrics. The same segmentation method uncertainty that affects MATV will also affect these derived features, making standardization equally important for radiomics research in NHL. Future work should examine how the choice of segmentation method propagates into downstream radiomics-based predictive models.