Lymphoma is one of the most heterogeneous groups of cancers, and [18F]FDG PET/CT imaging has become the standard modality for staging, therapy response assessment, and disease surveillance across the majority of lymphoma subtypes. The clinical value of PET/CT extends beyond visual interpretation: quantitative features such as maximum standardized uptake value (SUVmax), mean SUV (SUVmean), metabolic tumor volume (MTV), and total lesion glycolysis (TLG) carry documented prognostic value and can predict treatment outcomes in diffuse large B-cell lymphoma (DLBCL), follicular lymphoma (FL), and Hodgkin lymphoma. More recently, spatial dissemination features derived from the full tumor burden have shown independent prognostic power in DLBCL, reflecting the importance of capturing disease spread rather than a single dominant lesion.
The segmentation bottleneck: Despite this prognostic potential, quantitative PET/CT analysis is not routinely performed in clinical practice. The primary obstacle is the need for accurate lesion segmentation before any features can be extracted. Manual segmentation by expert nuclear medicine physicians is considered the gold standard, but it is labor-intensive (the authors measured a median time of 42 minutes per patient) and is subject to inherent intra- and inter-observer variability. For lymphoma patients, who often present with multiple lesions scattered throughout the entire body in anatomically challenging locations, this burden is particularly heavy.
Scope of the study: This 2023 paper from the Champalimaud Foundation (Lisbon) systematically compares nine segmentation approaches on 65 whole-body [18F]FDG PET/CT staging images containing 920 identified lymphoma lesions or clusters. The nine methods span a wide range of automation levels: two absolute/relative threshold approaches, four clustering-based algorithms (standard and self-adaptive k-means and Bayesian), a semiautomatic deep learning method, and two fully automatic 3D U-Net models. The study treats manual segmentation by two experienced observers as the ground truth and uses the Dice coefficient (DC) as the primary overlap metric, complemented by intraclass correlation coefficient (ICC) analysis of 31 extracted lesion features.
The dataset is notably more comprehensive than most prior work in this space. Eighty-four percent of patients were stage III or IV, with lesions distributed across the head and neck, thorax, abdomen, and pelvis. Lymphoma subtypes included Hodgkin lymphoma (16), follicular NHL (29), DLBCL (19), and one low-grade NHL, allowing analysis across the spectrum of FDG-avid lymphoma presentations.
The study is retrospective and single-center. Sixty-five patients naive to treatment were randomly selected from cases diagnosed with B-cell lymphoma at the Champalimaud Foundation between 2009 and 2018. The mean patient age was 64 years (range 24 to 95), and mean BMI was 25 kg/m2. All PET/CT images were acquired on a Philips Gemini TF 16 scanner following intravenous injection of 3.5 plus or minus 0.20 MBq/kg. Reconstruction adhered to EANM Research GmbH EARL1 harmonization guidelines, ensuring SUV calibration and image quality consistency. Six nuclear medicine physicians (five with 7 to 13 years of experience and one with more than 25 years) identified all visible lesions on axial PET images. Two observers then independently performed both 3D ROI delineation and slice-by-slice manual segmentation on all identified lesions.
Classical semiautomatic methods: All seven non-deep-learning methods operate inside the manually drawn 3D ROI, making them semiautomatic in the sense that they require human initialization but run automatically thereafter. The two threshold approaches used an absolute SUV cutoff of 2.5 (validated as optimal for baseline MTV in DLBCL) and a relative cutoff of 41% of SUVmax (recommended by EANM for solid tumors and used in Hodgkin and DLBCL studies). Standard k-means and Bayesian classifiers were also applied with two classes (lesion and background). All four of these methods were implemented in C++ using the ITK and CImg libraries and can be integrated into 3D Slicer 4.11.2 as an extension.
Self-Adaptive Configuration (SAC): The authors developed a SAC variant for both k-means and Bayesian clustering. The algorithm begins by dividing voxel intensities into three classes, then computes a contrast coefficient: coef = (m3 minus m1) divided by (m3 plus m1), where m3 and m1 are the mean SUVs of the highest- and lowest-intensity classes, respectively. Based on the coefficient value, it either merges the two lower classes (coef below 0.90), merges the two higher classes (coef at or above 0.94), or runs a standard two-class segmentation. This adaptive merging logic intelligently handles lesions with varying contrast relative to background, which is especially important when the lesion-to-background ratio changes across anatomical locations and patient body types.
Statistical analysis: Dice coefficient distributions were compared using Friedman and Wilcoxon tests at a 5% significance level. Feature agreement between manual and automated segmentations was assessed by ICC for absolute agreement, computed both on a total tumor burden basis (summing across all lesions per patient) and on a single representative lesion basis using the lesions with the highest TLG and highest SUVmax as the disease representatives. IBM SPSS 26 and R 3.2.5 were used for all statistical computations.
All three deep learning approaches in this study were built on the nnU-Net framework (Isensee et al., Nature Methods 2021), a self-configuring biomedical segmentation framework that automatically determines optimal network architecture and hyperparameters from dataset characteristics using heuristic rules. The core architecture in each case is a 3D full-resolution U-Net composed of an encoder-decoder network with skip connections between the two paths. All networks used fivefold cross-validation with an 80/20 training-to-validation split.
Fully automatic method 1 (Blanc-Durand et al.): This publicly available 3D U-Net was trained by Blanc-Durand and colleagues on 639 whole-body DLBCL PET/CT images using both PET and CT as input channels. It had achieved a median internal validation DC of 0.79 (IQR 0.66 to 0.87) in the original publication. Notably, the ground truth for this network's training was generated using a relative threshold (41% of SUVmax) rather than manual segmentation, a detail that the authors argue has significant consequences for cross-study comparisons.
Fully automatic method 2 (trained internally): The authors trained a second 3D U-Net on 144 whole-body [18F]FDG PET images from lymphoma patients (mean age 46 plus or minus 19 years, 68 women) drawn from The Cancer Imaging Archive, specifically the autoPET MICCAI challenge dataset. All lesions in this dataset were manually segmented by a radiologist and nuclear medicine physician in consensus, making the ground truth substantially higher quality than threshold-based labels. Training ran on an Ubuntu 20.04 WSL system with an NVIDIA RTX A6000 GPU.
Semiautomatic deep learning method: A third 3D U-Net was trained on 64 x 64 x 64 voxel patches (corresponding to 256 x 256 x 256 mm3) centered on individual lesions from the same autoPET dataset. In total, 2,537 isolated lesions were used to extract training patches. At inference time, the physician identifies each lesion of interest, and a patch centered on that lesion is extracted and passed to the network for segmentation. This simulates a realistic clinical workflow where the physician indicates lesion locations and the AI provides the segmentation boundary within each indicated region. The network was not optimized for fully automatic whole-body scanning and was tested only in physician-identified patches.
A total of 920 [18F]FDG-avid lesions were identified across 65 patients. The inter-observer Dice coefficient for manual segmentation (the two expert observers segmenting the same lesions independently) was a median of 0.84, establishing a realistic baseline for human-level variability in lymphoma PET segmentation. Classical semiautomatic methods were approximately five times faster than manual segmentation (median 8 minutes versus 42 minutes per patient, p less than 0.001 by Wilcoxon test), with nearly all of that time consumed by the manual 3D ROI initialization step rather than the automatic segmentation algorithm itself.
SAC Bayesian achieves the best results: The SAC Bayesian method produced the highest Dice coefficients, with median DC of 0.89 (Observer 1) and 0.85 (Observer 2) against manual segmentation. Crucially, the inter-observer DC between the two SAC Bayesian segmentations was 0.94, significantly higher than the 0.84 inter-observer DC for manual segmentation alone (p less than 0.001, Wilcoxon test). This means the SAC Bayesian method is actually more reproducible between operators than expert manual contouring, a finding with direct implications for multicenter studies where inter-observer consistency is a major concern. No statistically significant differences in DC were found between lymphoma subtypes (p at or above 0.36, Kruskal-Wallis test).
Threshold-based methods and publicly available 3D U-Net underperform: The two fixed-threshold approaches (SUV 2.5 and 41% SUVmax) and the Blanc-Durand publicly available 3D U-Net produced the lowest Dice coefficients, with median DC ranging from 0.56 to 0.68. These results place them well below both manual inter-observer agreement and the SAC-based clustering methods. The authors attribute the threshold methods' weakness to their dependence on the absolute SUV scale, which is strongly influenced by scanner resolution, reconstruction parameters, and body weight, factors that vary across institutions even with EARL1 harmonization.
Internally trained deep learning methods show promise: The semiautomatic deep learning approach achieved median DC of 0.83 (Observer 1) and 0.79 (Observer 2), and the fully automatic internally trained model produced comparable results. Both were statistically superior to the Blanc-Durand network (p less than 0.05, Wilcoxon test). However, at the patient level, all three deep learning approaches produced DC below 0.22 in a subset of patients: 9 patients for semiautomatic DL, 11 for fully automatic DL, and 8 for the Blanc-Durand model. Seven patients showed poor results across all three DL methods simultaneously, with failures primarily attributed to the inability to correctly identify and segment lesions near normal high-uptake regions such as the brain, liver, kidneys, and bladder.
Segmentation overlap measured by Dice coefficient captures geometric accuracy, but the ultimate clinical purpose of segmentation is to extract meaningful quantitative biomarkers from the delineated volumes. This study quantified 31 features per lesion, grouped into three categories: intensity features (SUV scale), geometric features (shape and size), and spatial dissemination features (eight measures of tumor spread across the body). ICC for absolute agreement was computed between features derived from manual segmentation and from each automated method, both on a total tumor burden basis and for individual representative lesions.
SAC Bayesian feature agreement: The SAC Bayesian method produced excellent ICC values (ICC at or above 0.92) for all five of the most clinically used quantitative features: SUVmax, SUVmean, SUVpeak, MTV, and TLG when computed on total tumor burden. For the inter-observer comparison of SAC Bayesian features, the agreement was equal to or higher than between the two manual segmentations for every one of the 31 features, confirming that the method's higher reproducibility translates from geometric overlap to quantitative outputs. The eight spatial dissemination features were particularly reproducible, with ICC at or above 0.97 for both manual and SAC Bayesian segmentations.
Representative lesion analysis: When the analysis focused on a single representative lesion per patient rather than total tumor burden, choosing the lesion with the highest TLG produced excellent ICCs (at or above 0.94 for all five key features). However, using the lesion with the highest SUVmax as the representative produced inferior ICCs for features depending on volume: MTV and TLG dropped to 0.85 and 0.76, respectively (intra-observer) and 0.60 and 0.72 (inter-observer). The authors attribute this to the fact that the highest-SUVmax lesion is sometimes part of a conglomerated cluster, where segmentation decisions about connected versus disconnected neighboring lesions change the volume substantially between methods and observers. The lesion with the highest TLG is therefore a more reliable representative for disease characterization.
Deep learning feature analysis excluded: For all three deep learning methods, the authors decided not to report quantitative feature analysis beyond the Dice coefficient. The reason was that in 20 out of 65 patients (31%), at least one non-malignant region with a higher SUVmax than the highest actual lymphoma lesion was incorrectly identified and segmented, rendering the SUVmax-dependent features unreliable. This false-positive segmentation in high-uptake anatomical regions is a recognized challenge for fully automatic deep learning models and represents a practical obstacle to autonomous clinical deployment.
The authors situate this work as the first study to compare manual, semi-, and fully automatic segmentation methods on whole-body lymphoma PET/CT using as many as 31 lesion features, and with a patient and lesion count larger than most prior clinical lymphoma segmentation studies. The core finding, that SAC Bayesian outperforms all other methods including deep learning in terms of both segmentation accuracy and feature reproducibility, reflects an important practical reality: semiautomatic methods that leverage human expertise for initialization can outperform fully automatic models on datasets that differ from the training distribution.
Why the Blanc-Durand network underperformed: The authors provide a compelling explanation for the Blanc-Durand network's poor external performance (DC 0.56 to 0.68, similar to fixed-threshold methods). Although the model was trained on a large dataset of 639 DLBCL images, the ground truth labels used for training were generated using the 41% SUVmax threshold rather than manual segmentation. When evaluated against manual segmentation as the external test standard, the network effectively inherited the systematic biases of the threshold method it was trained to replicate. This finding underscores that training data quality is as important as training data quantity when developing deep learning segmentation models.
Why threshold-based methods are inadequate: Threshold-based segmentation methods tie the lesion boundary to the absolute or relative SUV scale, which is strongly scanner-dependent. SUVmax in particular is strongly influenced by the physical spatial resolution of the scanner and reconstruction parameters, varying across institutions even when EARL1 protocols are followed. Using a fixed threshold (whether absolute at 2.5 or relative at 41%) therefore introduces systematic segmentation errors that propagate directly into feature extraction and can yield MTV and TLG values that differ substantially from manual gold standard measurements. The authors explicitly recommend against threshold methods in clinical studies, particularly when no harmonized reconstruction protocol is enforced across sites.
The clinical readiness question for deep learning: Both internally trained DL models achieved overall Dice coefficients close to SAC-based methods, but failed in 7 to 11 patients with catastrophically low DC values (below 0.22). The authors note that clinical use of an autonomous system that fails this severely in roughly 10 to 15% of patients is unacceptable. The failures were not random but concentrated near normal high-uptake regions (brain, liver, kidneys, bladder), suggesting that the training dataset, while containing manual labels, did not include sufficient examples of lesions in these anatomical contexts to allow the network to generalize robustly.
Single-center and single-scanner design: All 65 patients were scanned on the same Philips Gemini TF 16 PET/CT scanner at a single institution following standardized EARL1 protocols. While this controls for acquisition variability within the study, it means the results may not generalize to institutions using different scanner hardware or reconstruction software. The SAC Bayesian method, which is contrast-based and calibrated for lesion-to-background ratios, may behave differently on PET images acquired with different scanner resolutions or reconstruction filters. The same concern applies even more strongly to the deep learning models, which are known to be sensitive to acquisition domain shifts.
Training data mismatch for deep learning: The two internally trained 3D U-Net models used the autoPET MICCAI challenge dataset for training, which was acquired on Siemens Biograph mCT scanners at different institutions, with different age and sex distributions (mean age 46 vs 64 years in the test set) and different physician segmentation protocols. This domain gap between training and test data is a direct contributor to the observed failure cases. External validation with independent datasets has been shown in systematic reviews to reduce AI radiology model performance by 5 to 15% on average, and this study provides a concrete example of that phenomenon in lymphoma segmentation.
Lesion identification variability not assessed: The study measured the agreement between observers in segmenting the lesions, but did not assess inter-observer agreement in the prior step of lesion identification, that is, whether the six nuclear medicine physicians who identified lesions from the reports would have identified the same set of lesions if each had reviewed the scans independently. Variability in lesion identification could affect any downstream segmentation comparison, and the authors acknowledge this as an important limitation that would make the method comparisons fairer if addressed.
Dataset size for DL training: The 144 training images used for the internally trained models, while publicly available, are modest by deep learning standards. Lymphoma PET segmentation is particularly challenging because of the diverse anatomical distribution of lesions, the variable number of lesions per patient, and the presence of normal high-uptake regions that can mimic malignancy. Larger training datasets with diverse scanner types, acquisition protocols, and patient demographics are likely needed for the fully automatic models to reach the reliability required for autonomous clinical use.
The authors conclude that the SAC Bayesian classifier is the most robust and reproducible segmentation method evaluated in this study. It is five times faster than manual segmentation (median 8 vs 42 minutes per patient), achieves higher inter-observer Dice coefficients than manual-to-manual comparison (0.94 vs 0.84, p less than 0.001), and produces feature values that agree excellently with manual ground truth for all clinically relevant quantitative biomarkers (ICC at or above 0.92 for SUVmax, SUVmean, SUVpeak, MTV, and TLG). The paper explicitly recommends SAC Bayesian as a viable replacement for manual segmentation in clinical lymphoma studies and as the preferred approach when building large annotated datasets.
Strategic role in deep learning dataset construction: The most forward-looking recommendation in the paper is that the SAC Bayesian method can serve as a practical tool for building the large, high-quality, consistently labeled datasets that deep learning models require but currently lack. The scarcity of accurately labeled whole-body lymphoma PET/CT datasets is one of the most significant bottlenecks for DL model development in this disease area. Using a reproducible semiautomatic method with near-manual accuracy to pre-annotate large institutional archives could accelerate dataset construction dramatically compared to fully manual annotation, potentially unlocking the training data scale needed for robust deep learning segmentation.
Deep learning prognosis: The authors do not dismiss deep learning but frame its current limitations clearly. Both internally trained models showed overall performance close to SAC-based methods across the full cohort but failed unacceptably in a meaningful minority of patients due to false-positive segmentation near normal high-uptake structures. The authors advocate for continued development with larger and more diverse training datasets, better handling of the brain, liver, kidney, and bladder regions as exclusion zones or separate classification targets, and mandatory human verification of all segmented lesions before clinical use. They predict that fully automatic methods will eventually surpass semiautomatic approaches as training data scales up.
Clinical implications: For lymphoma quantification to become standard of care, the field needs reproducible segmentation methods that can be adopted across institutions without requiring expert manual contouring for every patient. The SAC Bayesian method, with its open-source implementation available as a 3D Slicer extension, represents a practical near-term solution. Standardizing segmentation methodology across multicenter studies would also enable the construction of larger validated prognostic models that use the full suite of 31 lesion features, including the eight spatial dissemination measures that showed excellent reproducibility (ICC at or above 0.97) across both manual and SAC Bayesian segmentations.