Total metabolic tumor volume (TMTV) has become one of the most clinically important prognostic biomarkers across the major lymphoma subtypes. Derived from whole-body 18F-FDG PET/CT imaging, TMTV quantifies the aggregate metabolic burden of all tumor deposits in the body at baseline. Multiple large multicenter trials have demonstrated that high TMTV at diagnosis is independently associated with inferior progression-free survival and overall survival in Hodgkin lymphoma (HL), diffuse large B-cell lymphoma (DLBCL), and follicular lymphoma (FL). Despite this prognostic value, TMTV has not yet been incorporated into standard clinical risk scoring systems, partly because manual calculation is impractical at scale.
The manual delineation bottleneck: Computing TMTV requires a physician or trained technician to manually draw regions of interest (ROIs) around every tumor deposit visible on the PET scan, across the entire body from skull base to feet. Lymphoma patients often have 15-25 or more individual tumor foci distributed across nodal and extranodal sites. Manual segmentation is time-consuming, taking 30 minutes to several hours per patient, and is highly variable between observers, particularly for lesions with heterogeneous or diffuse uptake patterns. This interobserver variability introduces noise into TMTV-based risk stratification.
Thresholding methodology disagreements: Compounding the segmentation challenge is the lack of consensus on which mathematical threshold to use when converting continuous SUV (standardized uptake value) maps into binary tumor masks. The three most commonly used methods each draw different boundaries. The 41% SUVmax method sets the threshold per lesion at 41% of that lesion's maximum SUV. The SUV 2.5 absolute cutoff includes all voxels above 2.5 SUV. The SUV 4.0 cutoff uses a higher fixed value. The Otsu method is a histogram-based automatic approach. Each method yields a different TMTV value for the same patient, making cross-study comparison difficult and hindering standardization for clinical trials.
This paper from Revailler et al. at the Institut Universitaire du Cancer de Toulouse addresses both problems simultaneously: training a deep learning convolutional neural network (CNN) on the largest manually labeled baseline lymphoma PET/CT dataset assembled to date, and doing so in a way that remains agnostic to any single thresholding methodology.
The study population was drawn from seven multicenter clinical trial datasets conducted under the Lymphoma Study Association (LYSA): AHL2011 (n=703), GAINED (n=573), RELEVANCE (n=277), REMARC (n=217), FLIP (n=125), LNH2007-3B (n=61), and PVAB (n=74). Together these trials contributed 2,030 baseline FDG PET/CT scans spanning three FDG-avid lymphoma subtypes: Hodgkin lymphoma (HL, n=777), DLBCL (n=851), and follicular lymphoma (FL, n=402). Patients ranged across Ann Arbor stages I through IV, with the majority of HL cases at stages IIB-IV and most DLBCL and FL cases at stages II-IV. The diversity of trial protocols, imaging centers, and scanner hardware across seven studies provides meaningful heterogeneity that strengthens generalizability.
Image preprocessing pipeline: All PET and CT DICOM files underwent a standardized preprocessing pipeline. Images were first converted from DICOM to NIfTI (.nii) format and then resampled to a uniform 128 x 128 x 256 matrix at 4.0 x 4.0 x 4.0 mm isotropic voxel resolution using linear interpolation. This fixed physical voxel grid standardized acquisition field coverage up to 102 cm while preserving the whole-body image without cropping or tiling. PET and CT were co-registered at the same spatial origin to construct 4D PET/CT input tensors. Input values were normalized to the range 0 to 1, corresponding to CT Hounsfield unit range (-1000 to 1000) and PET SUV range (0 to 25).
Dataset split: The 2,030 scans were partitioned into 60% training (n=1,218), 20% validation (n=405), and 20% test (n=407) sets. The training set was used for model weight updates, the validation set for monitoring training progress and early stopping, and the test set was held out entirely for final performance evaluation. Data augmentation applied random translations, rotations, and scaling per patient, generating one additional augmented image per training case to reduce overfitting without expanding the annotation burden.
The central methodological innovation in this study is the construction of a probability-weighted ground truth rather than forcing the CNN to learn a single binary segmentation methodology. For each of the 2,030 cases, 13 expert physicians trained in TMTV calculations at LYSA centers manually delineated all tumor uptakes using the open-source Fiji PET/CT viewer. These manual ROIs defined the spatial extent of disease based on physician judgment, independent of any mathematical threshold rule.
Generating the TMTVprob map: After manual delineation, four different thresholds were independently applied to the manual ROIs: 41% SUVmax, SUV greater than 2.5, SUV greater than 4.0, and the Otsu histogram-based method. For each voxel within the manually drawn ROI, the fraction of the four threshold methods that classified that voxel as tumor was computed, producing a per-voxel probability between 0 and 1 in four discrete steps (0, 0.25, 0.5, 0.75, 1.0). This TMTV probability map, called TMTVprob, reflects genuine biological uncertainty about the tumor boundary: voxels deep in the core of a lesion, where all four thresholds agree it is tumor, get probability 1.0, while edge voxels where thresholds disagree get intermediate values.
Why this matters: By training the CNN to predict TMTVprob rather than any single binary mask, the model learns a generic representation of tumor probability that is not locked to one methodological convention. Post-processing can then apply any of the three standard thresholding algorithms to the predicted probability map to recover TMTV estimates consistent with each clinical convention. This two-step architecture cleanly separates lesion detection (the CNN's job) from segmentation thresholding (a rule-based post-processing step), making the model reusable across clinical contexts that use different TMTV methodologies.
Model architecture and training: The chosen architecture was a 3D V-NET, a fully convolutional network specifically designed for volumetric medical image segmentation. The V-NET used 4 resolution levels and 8 feature channels at the first level, producing a progressively deeper feature hierarchy across the whole-body volume. Training ran on a dual NVIDIA 1080 Ti GPU system for 93 hours over 100 epochs. The optimizer was stochastic gradient descent with a learning rate of 0.001 and momentum of 0.9. The loss function was soft dice loss (1 minus dice coefficient), chosen because it handles the severe class imbalance between background voxels and tumor voxels without requiring class-weighted adjustments. The final output layer used sigmoid activation to produce per-voxel probabilities matching the TMTVprob target.
The primary validation metric was the Dice similarity coefficient, which measures the overlap between predicted and ground truth segmentation masks on a 0 to 1 scale, with 1 indicating perfect agreement. Against TMTVprob on the training set, mean Dice reached 0.836 - a strong result indicating that the model successfully learned to reproduce expert-weighted probability maps. On the held-out validation set (n=405), mean Dice was 0.835, essentially identical to training performance, demonstrating excellent generalization without overfitting. On the fully independent test set (n=407), mean Dice was 0.76 with a median of 0.81, representing a modest reduction consistent with the greater heterogeneity of unseen cases.
Jaccard coefficient and voxel-level metrics: On the test set, the mean Jaccard coefficient (also called intersection over union, IoU) was 0.64 +/- 0.17, with a median of 0.68 and interquartile range of 0.55 to 0.76. Jaccard is a stricter metric than Dice and these values confirm meaningful but imperfect overlap, consistent with the difficulty of whole-body lymphoma segmentation across heterogeneous lesion sizes and locations. Voxel-level sensitivity was 0.76 +/- 0.17, specificity 0.99 +/- 0.00, positive predictive value (PPV) 0.81 +/- 0.18, and negative predictive value (NPV) 0.99 +/- 0.00. The near-perfect specificity and NPV confirm that the model is highly conservative about labeling non-tumor tissue as tumor, with false positives remaining rare - an important property for clinical utility.
Raw prediction before post-processing: Before any threshold-specific post-processing, the raw binary predicted segmentation (applying a 0.5 probability cutoff to the CNN output) showed median Dice scores of 0.71 for the 41% SUVmax ground truth, 0.70 for SUV 2.5, and 0.81 for SUV 4.0 against the respective manual binary masks. The higher performance against the SUV 4.0 reference is mechanistically sensible: voxels above SUV 4.0 are those most likely to receive high probability scores from all four training thresholds, making them the most consistently labeled part of the TMTVprob map and thus the easiest for the CNN to predict with confidence.
After applying the three threshold-specific post-processing steps to the raw CNN output, the study compared automated TMTV predictions against corresponding manually computed TMTV values across 407 test patients. For the 41% SUVmax methodology, median TMTV values were 240 +/- 498 mL (manual) versus 208 +/- 420 mL (predicted), a difference that was not statistically significant (NS). For the SUV 4.0 cutoff, medians were 212 +/- 457 mL (manual) versus 207 +/- 430 mL (predicted), also not significant. The SUV 2.5 cutoff showed a larger and statistically significant discrepancy: 400 +/- 621 mL (manual) versus 253 +/- 472 mL (predicted), with p less than 0.001.
Spearman correlation analysis: Despite absolute volume differences, the rank-order correlation between manual and predicted TMTV was very strong across all three methods. Spearman correlation coefficients were r=0.92 (41% SUVmax), r=0.95 (SUV 2.5), and r=0.98 (SUV 4.0), all with p less than 0.001. These correlations indicate that even where absolute TMTV values differ, the relative ranking of patients by tumor burden is well preserved - a critical property for prognostic scoring, where patients are typically stratified by quartile or median TMTV rather than absolute volume thresholds. The high r=0.98 for the SUV 4.0 method reflects this cutoff's alignment with the high-probability core voxels that the CNN models most accurately.
Dice scores per methodology after post-processing: Following threshold-specific post-processing, median Dice scores improved compared to the raw prediction for most methods. Final Dice scores were 0.77 (41% SUVmax), 0.70 (SUV 2.5), and 0.90 (SUV 4.0) across all patients. The median differences in TMTV volume between manual and predicted were 32 mL (41% SUVmax), 147 mL (SUV 2.5), and 5 mL (SUV 4.0). These differences must be interpreted in context of the high standard deviations of TMTV in this cohort - over 400 mL - reflecting the extreme range of disease burden from patients with minimal tumor load to those with bulky disseminated disease. A difference of 5 mL for the SUV 4.0 method is clinically negligible given this variance.
The 41% SUVmax post-processing challenge: The 41% SUVmax methodology proved the hardest to reproduce because the threshold is computed per lesion rather than applied globally across the body. The CNN's raw output is a single whole-body mask without lesion-level identity. A watershed sub-segmentation algorithm was applied to tumor areas exceeding 30 mL (likely coalescent tumor masses) to isolate sub-components, allowing per-lesion SUVmax values to be computed. The resulting automated ROI count (median 25, mean 33 +/- 30) was comparable to but higher than the manual physician-drawn ROI count (median 15, mean 21 +/- 20), with the excess driven by watershed over-splitting of large contiguous tumor masses.
One of the primary goals of this study was to build a model generic enough to perform across three clinically distinct lymphoma subtypes rather than requiring subtype-specific models. The test set contained all three subtypes - HL, FL, and DLBCL - and the post-processed Dice scores were reported separately for each. For HL, median Dice values were 0.70 (41% SUVmax), 0.68 (SUV 2.5), and 0.93 (SUV 4.0). For FL, they were 0.76 (41%), 0.68 (2.5), and 0.90 (4.0). For DLBCL, they were 0.85 (41%), 0.75 (2.5), and 0.87 (4.0).
Cross-subtype interpretation: DLBCL showed the highest Dice for the 41% SUVmax method (0.85) compared to HL (0.70) and FL (0.76). This likely reflects the characteristically higher and more homogeneous FDG uptake in DLBCL lesions, which produces more consistent threshold boundaries. HL and FL lesions can have more variable uptake intensity and more complex morphologies, including bulky mediastinal masses with heterogeneous internal uptake in HL and often small scattered nodes in FL. The strong performance across all three subtypes with a single shared model is clinically meaningful because it means a single deployed tool can serve the full spectrum of FDG-avid lymphoma patients seen in practice, without requiring users to select a subtype-appropriate model.
Overlap with prior literature: The authors note that reported Dice scores from prior CNN-based TMTV studies using single-subtype training datasets and single-methodology ground truth generally ranged from 0.73 to 0.88. The present model's performance of median 0.81 on the test set (against TMTVprob) sits within this range despite the substantially more difficult task of generalizing across three subtypes simultaneously. Previous models trained on single DLBCL datasets by Blanc-Durand et al. achieved Dice of 0.76 and Spearman r=0.92 for 41% TMTV - comparable to the current study's 41% results while the current study achieves this across three subtypes with a methodology-agnostic approach.
The authors made a deliberate architectural choice to use a single whole-body 3D V-NET operating on the full resampled volume, rather than the cascaded or patched approaches used in several previous studies. Cascaded architectures divide the body into head, chest, and abdomen segments and apply separate 2D or 3D CNNs to each, with a final aggregation step. Patched approaches split the volume into overlapping 3D sub-volumes to fit GPU memory constraints. The whole-body approach used here, enabled by resampling to 4 mm isotropic voxels (fitting the 128x128x256 input grid), preserves spatial context across body regions without introducing stitching artifacts or requiring region-specific models. This simplicity makes deployment more straightforward.
Where the model struggles: The SUV 2.5 absolute cutoff proved the hardest methodology for the model to reproduce, yielding Dice of 0.70 and Spearman r=0.95 versus 0.98 for SUV 4.0. The authors attribute this to the low 2.5 threshold capturing background voxels that are inconsistently labeled across the other three training methods, pulling the TMTVprob values for these voxels toward intermediate (0.25 or 0.50) probability rather than firmly toward 1.0. The CNN is then asked to reproduce a very fuzzy boundary condition that poorly represents the actual lesion biology. A larger absolute volume difference of 147 mL was observed for the SUV 2.5 method, driven by systematic underestimation of large, diffuse uptake zones that 2.5 SUV captures but the CNN's probability threshold misses.
False positives and the need for manual review: The model's near-perfect specificity (0.99) does not eliminate clinically meaningful false positives. The paper presents a case in which tracer extravasation at the injection site (left arm) produced a high-uptake region that was classified as tumor, yielding a Dice score of 0.32 for that patient. Ectopic kidney positions represent another known failure mode. The authors explicitly state that manual review and correction will remain necessary even with automated TMTV, because visual inspection by a physician can quickly identify these obvious false positives. What changes is the time burden - reviewing and correcting an automated segmentation is far faster than starting from scratch.
Toward clinical deployment: The authors developed and released the open-source Dicom-To-CNN library (MIT License) to handle the full preprocessing and output pipeline, converting DICOM inputs to NIfTI for model inference and exporting predicted segmentations back to interoperable DICOM formats (RTSTRUCT and/or DICOM SEG). Collaborations with major open-source medical image processing projects are underway to integrate the inference, visualization, and segmentation editing workflow into viewer tools that can be practically deployed in hospital PACS environments.
The study authors frame the next phase of work around full clinical integration rather than further model refinement. While the CNN's segmentation accuracy is strong, the practical bottleneck is now workflow embedding - connecting the model to clinical PACS and PET/CT workstations in a way that is reliable, fast, and fits radiologist or nuclear medicine physician review processes. The Dicom-To-CNN library represents a first step, but integration with established open-source medical imaging platforms (such as 3D Slicer, OHIF Viewer, or Orthanc) is needed to reach a deployable state at institutions that lack dedicated computational infrastructure teams.
Prospective validation in decisional trials: The authors call for embedding the automated TMTV tool into prospective LYSA clinical trials where TMTV is used as a treatment stratification or response monitoring variable. This would allow head-to-head comparison of automated versus manual TMTV for prognostic accuracy, inter-reader agreement improvement, and determination time in a real workflow context rather than a retrospective research setting. Critically, it would establish whether automated TMTV is equivalent to or better than manual TMTV for the specific clinical decision (e.g., treatment intensification in high-TMTV patients) that it would be used to support.
Expanding subtype coverage: The current model covers the three major FDG-avid lymphoma subtypes (HL, DLBCL, FL), but several other FDG-avid NHL entities - including mantle cell lymphoma, peripheral T-cell lymphoma, and primary mediastinal B-cell lymphoma - were not represented in the training data. Future model versions incorporating cases from these rarer subtypes could broaden the applicability of a single deployed model to cover the full spectrum of aggressive lymphomas encountered in a hematology center. Given that these subtypes share overlapping imaging patterns with DLBCL, transfer learning from the current model may require only modest additional labeled data.
Combined AI workflow: TMTV automation is one component of a larger AI-assisted PET/CT analysis workflow for lymphoma. Future platforms could integrate automated TMTV calculation with automated Deauville scoring of interim and end-of-treatment PET, automated identification of new extranodal lesions, and quantitative texture analysis for prognostic radiomics - all generated from a single inference pass over the baseline and interim PET/CT pair. Such an integrated output would provide clinicians with a comprehensive quantitative report to supplement visual interpretation, potentially standardizing response assessment across institutions in cooperative group trials.