Diffuse large B-cell lymphoma (DLBCL) is the most common non-Hodgkin lymphoma subtype and a disease where 18F-FDG PET/CT imaging is standard of care for both initial staging and treatment response assessment. Two imaging-derived biomarkers have gained substantial prognostic traction in DLBCL: total metabolic tumor volume (TMTV), which measures the total burden of metabolically active lymphoma across the body, and Dmax, which captures the spatial dissemination of disease by measuring the largest distance between any two lesions. Both have been validated in multiple prospective and retrospective cohorts as independent predictors of progression-free survival (PFS) and overall survival (OS).
The calculation problem: Despite their prognostic value, TMTV and Dmax require manual or semi-automated delineation of lymphoma lesions from full 3D whole-body PET/CT volumes, which typically contain more than 200 transaxial slices per patient. This process is time-consuming, prone to interobserver variability, and practically difficult to implement outside of specialized academic centers. Even when AI-assisted 3D segmentation methods using convolutional neural networks (CNNs) are employed, their outputs require expert review and correction across all 3D slices before the biomarkers can be trusted for clinical decision-making.
The MIP hypothesis: Nuclear medicine physicians routinely generate 2D maximum-intensity projection (MIP) images as a compressed, whole-body summary of the 3D PET distribution. These coronal and sagittal MIP views collapse all the axial slices into a single projection image that is far faster to visually assess than the full 3D volume. This study hypothesized that surrogate versions of TMTV and Dmax could be automatically estimated from just two MIP images using a deep learning algorithm, and that these surrogates would retain clinically meaningful prognostic power comparable to the 3D-derived gold-standard biomarkers.
The study drew on two independent DLBCL trial cohorts - REMARC (NCT01122472) and LNH073B (NCT00498043) - with expert-validated 3D lesion delineations, enabling a rigorous comparison between MIP-derived surrogate biomarkers and the established 3D measurements across both training and external test populations.
The study enrolled 475 patients total from two independent clinical trials, of whom 93 were excluded because their baseline PET/CT images lacked expert segmentation or contained fewer than two lymphoma lesions (the minimum needed to compute a Dmax distance). This left 382 evaluable patients: 287 from the REMARC cohort and 95 from the LNH073B cohort. The REMARC trial was an international double-blind, randomized phase III study initiated in 2010, while LNH073B was a prospective multicenter randomized phase II study that began enrollment in 2007. The demographic differences between the two cohorts are notable: REMARC patients had a median age of 68 years (IQR 64-73), while LNH073B patients had a median age of 46 years (IQR 33-55), reflecting their different eligibility criteria and treatment contexts. The vast majority of patients in both cohorts had Ann Arbor stage II or higher disease (99.6% in REMARC, 100% in LNH073B).
Reference biomarker computation: For each patient, physicians manually identified lymphoma regions in the 3D PET images, removing areas of physiologic FDG uptake (such as the brain, heart, kidneys, and bladder) and adding pathologic lesions missed by automated methods. These expert-validated 3D tumor masks were then used to calculate reference TMTV (the total volume of voxels above a predefined SUV threshold belonging to lymphoma regions) and reference Dmax (the largest centroid-to-centroid distance between any two distinct lesion clusters in 3D space). These 3D measurements served as the gold standard against which all AI-derived surrogate biomarkers were benchmarked.
MIP generation: From each patient's 3D PET image, two 2D MIP images were generated: one in the coronal plane and one in the sagittal plane, 90 degrees apart. For each pixel in the MIP, the value was set to the maximum intensity observed along the ray perpendicular to the projection plane. Corresponding 2D binary mask images (MIP_masks) were derived by projecting the expert-validated 3D lymphoma regions into the same coronal and sagittal planes. These MIP_masks, which mark which MIP pixels correspond to lymphoma, served as the training labels for the deep learning segmentation model.
The AI model was built on a previously published supervised 2D deep learning architecture incorporating residual learning principles (adapted from ResNet) with a context feedback loop designed for robust medical image segmentation. The coronal and sagittal PET MIP images were treated as independent single-channel inputs. The network was trained to transform each MIP image into a binary output mask: pixels belonging to lymphoma regions were labeled 1, all others 0. Using both views as independent inputs (rather than combining them) allowed the model to process each projection separately and then combine the resulting segmentation information at the biomarker calculation stage.
Training configuration: Training was performed exclusively on the REMARC cohort (298 patients after exclusions). A 5-fold cross-validation scheme was used: patients were randomly split into five groups, and five separate models were trained on 80% of the REMARC data with the remaining 20% used for validation in each fold. This approach maximizes data utilization when the training set is of limited size. Training followed the CLAIM guidelines for AI in medical imaging and the Society of Nuclear Medicine and Molecular Imaging AI Task Force recommendations. The trained model is publicly available at https://github.com/KibromBerihu/ai4elife.
External validation: After training on REMARC, the model was applied without any retraining or fine-tuning to the independent LNH073B cohort (174 patients), which was acquired from a different trial with different patient demographics, acquisition protocols, and time period. This external validation is a critical test of generalizability, since a model that performs well only on the data it was trained on provides limited clinical value.
Segmentation performance metrics: The CNN's segmentation accuracy was evaluated using the Dice similarity coefficient (which measures overlap between predicted and reference masks, ranging from 0 to 1), sensitivity (fraction of true lymphoma pixels correctly identified), and specificity (fraction of non-lymphoma pixels correctly excluded). Wilcoxon signed-rank tests assessed whether performance differed significantly between coronal and sagittal views.
Once the CNN produces binary MIP segmentation masks for both the coronal and sagittal views, two surrogate biomarkers are calculated from these 2D masks. The surrogate total metabolic tumor volume (sTMTV) is defined as the total area in square centimeters occupied by lymphoma pixels in the MIP_masks. For each patient, sTMTV is computed as the sum of the pixel areas belonging to tumor regions in the coronal MIP mask (sTMTV_coronal) plus those in the sagittal MIP mask (sTMTV_sagittal). This gives a 2D area measurement in cm^2 rather than the 3D volume in cm^3 obtained from full PET/CT delineation, but it serves as a proxy for the total metabolic burden.
Surrogate Dmax (sDmax) calculation: Measuring the spatial spread of disease from a 2D projection requires a robust distance estimation approach. Simply finding the two most distant lymphoma pixels is unreliable because MIP projections can produce isolated noisy pixels at the image periphery. Instead, the method computes the sum of pixel counts along each column (x-profile) and each row (y-profile) of the MIP_mask. In each profile, the distance between the 2nd percentile and 98th percentile positions is calculated, yielding measurements of disease extent along two orthogonal axes. The sDmax for a single view is the sum of these two percentile-based extent measurements. For a patient, the final sDmax is the sum of sDmax from the coronal and sagittal views. The use of 2nd-98th percentile bounds, rather than absolute extremes, intentionally trims outlier pixels and improves robustness to segmentation noise.
Rationale for 2D surrogates: While sTMTV (in cm^2) and sDmax (in cm, combining information from two 2D projections) are dimensionally different from their 3D counterparts TMTV (cm^3) and Dmax (cm in 3D), the study's hypothesis was that the rank ordering of patients by these surrogate values would be sufficiently correlated with the rank ordering by the reference 3D values to preserve prognostic stratification ability. The Spearman rank correlation coefficient was used to test this assumption, since survival analysis using median cutoffs depends primarily on correctly classifying patients as above or below the median rather than on absolute value agreement.
The CNN achieved strong segmentation performance on both cohorts. On the REMARC cross-validation set, the median Dice score was 0.80 (IQR 0.63-0.89), with sensitivity of 80.7% (IQR 64.5%-91.3%) and specificity of 99.7% (IQR 99.4%-99.9%). The high specificity reflects the fact that lymphoma lesions occupy a small fraction of the total MIP image area, so correctly classifying the large background as non-tumor is relatively straightforward. Sensitivity at 80.7% indicates that roughly one-fifth of true lymphoma pixels are missed, which may correspond to small or low-uptake lesions that are visually difficult even for human readers.
External validation on LNH073B: When applied to the independent LNH073B cohort without any retraining, the model performed as well as on the training set or better: median Dice score 0.86 (IQR 0.77-0.92), sensitivity 87.9% (IQR 74.9%-94.4%), specificity 99.7% (IQR 99.4%-99.8%). The mean Dice scores were 0.80 +/- 0.17 on the coronal view and 0.79 +/- 0.17 on the sagittal view for LNH073B, with no statistically significant difference between views (P > 0.05) in either cohort. This consistency across cohorts from different time periods, institutions, and patient demographics is a meaningful indicator of model robustness.
Correlation between surrogate and reference biomarkers: After computing sTMTV and sDmax from the CNN-generated masks, Spearman rank correlations with the expert-derived 3D values were assessed. For the REMARC cohort, TMTV correlated with sTMTV at r = 0.878 (P < 0.001), and Dmax correlated with sDmax at r = 0.709 (P < 0.001). The lower correlation for dissemination versus volume suggests that the 2D projection compresses some 3D spatial information about disease spread that the 1D percentile-based approach partially recovers but cannot fully recapitulate. Of patients with TMTV above the median, 84% also had sTMTV above the median sTMTV; for Dmax, 78.5% concordance was observed.
On the external LNH073B cohort, TMTV-sTMTV correlation was r = 0.752 and Dmax-sDmax correlation was r = 0.714 (both P < 0.001). Among patients with above-median TMTV, 87.5% also had above-median sTMTV; for Dmax, 81.3% concordance. These cross-cohort figures are nearly identical to the REMARC results, confirming that the AI-derived surrogates reliably track the 3D reference values despite being computed from simplified 2D projections.
The central clinical question was whether sTMTV and sDmax carry prognostic value for PFS and OS that is comparable to the established 3D biomarkers TMTV and Dmax. Univariate Cox proportional hazard models and time-dependent AUC analyses were performed separately for each biomarker in both cohorts. In the REMARC cohort, all four biomarkers were statistically significant prognosticators of PFS: TMTV had an HR of 11.24 (95% CI 2.10-46.20) and AUC of 0.67 (0.60-0.73); sTMTV had an HR of 11.81 (95% CI 3.29-31.77) and AUC of 0.65 (0.58-0.72); Dmax had an HR of 9.0 (95% CI 2.53-23.63) and AUC of 0.65 (0.58-0.72); sDmax had an HR of 12.49 (95% CI 3.42-34.50) and AUC of 0.68 (0.62-0.75). These nearly overlapping hazard ratios and AUCs confirm that the AI-derived surrogates perform as well as the 3D reference measurements for PFS stratification.
Overall survival in REMARC: For OS, the 3D TMTV HR was 16.43 (95% CI 2.42-77.29) and AUC 0.67 (0.58-0.76); sTMTV HR was 22.14 (95% CI 4.73-69.06) and AUC 0.67 (0.58-0.76). For dissemination, Dmax HR was 8.60 (95% CI 1.47-28.33) and sDmax HR was 22.79 (95% CI 3.80-79.21). The surrogate biomarkers again produced hazard ratios and AUCs within the confidence intervals of the 3D reference values, with overlapping uncertainty ranges across all estimates.
Three-category risk stratification: When TMTV and Dmax were combined into a three-group risk model (low risk: both below median; intermediate risk: one above median; high risk: both above median), the REMARC cohort was divided into 108 low-risk, 112 intermediate-risk, and 67 high-risk patients. When the same stratification was performed using sTMTV and sDmax from the AI, the three-category classification matched the 3D-based classification with 71.4% accuracy. In LNH073B, the 3D-based classification identified 45 low-risk, 37 intermediate-risk, and 13 high-risk patients; the AI-based classification achieved 64.2% categorical agreement. Among the 13 patients classified as 3D-based high risk, 9 (69.2%) had OS under 4 years and 10 (76.9%) had PFS under 4 years.
When a simplified two-group classification based on a single biomarker was used (high vs. low tumor burden or high vs. low dissemination), the AI-based classification matched the 3D-based classification with 79% accuracy. This higher binary concordance suggests that the AI surrogates more reliably classify patients at the extremes of the risk spectrum, with disagreements concentrated near the median cutoff values.
The study's key finding - that surrogate biomarkers extracted from two 2D MIP images by an AI algorithm are prognostically equivalent to the 3D gold-standard measurements in two independent DLBCL cohorts - has direct implications for clinical implementation. The primary advantage of the MIP-based approach is the dramatic reduction in the time and effort required for expert quality control. When clinicians need to verify or correct AI segmentation results, reviewing and adjusting two 2D projection images is far faster than examining and correcting across 200+ axial PET slices. This alone could make TMTV and Dmax calculation feasible in high-volume clinical settings and at centers lacking the specialized image analysis infrastructure currently required for 3D biomarker computation.
Computational efficiency: Beyond expert review time, the 2D MIP approach also reduces the computational demands of training the AI itself. A deep learning model that segments 2D images has far fewer parameters than one designed to segment 3D volumes, requires less training data, and can be trained and deployed with less powerful hardware. This democratization of AI-based biomarker computation could extend access to TMTV and Dmax quantification to community oncology centers and lower-resource settings where GPU-intensive 3D deep learning pipelines are not practical.
Where MIP-based biomarkers perform best: Visual assessment of segmentation results in this study revealed that 2D MIP-based biomarkers performed better in patients with widely disseminated disease spread across multiple body regions, where the projection captures lesion extent effectively. Performance was comparatively weaker in patients with large, bulky, localized masses, where the MIP projection may not accurately represent the true 3D tumor volume because the projection collapses depth information. This limitation should be considered when selecting which patients or clinical scenarios benefit most from the MIP-based approach versus full 3D segmentation.
Consistency of Dmax cutoff values: An interesting secondary finding was that the optimal cutoff values for sDmax were nearly identical between the REMARC and LNH073B cohorts (approximately 122 cm and 122 cm for sDmax, matching the biological consistency of the 3D Dmax cutoff near 44-60 cm across cohorts). This robustness across demographically different patient populations suggests that sDmax as a measure of disease dissemination has inherent biological stability that transcends the specific trial population, supporting its potential utility as a generalizable prognostic threshold.
Retrospective multicohort validation but limited scale: Both REMARC and LNH073B are retrospective clinical trial datasets with expert-curated segmentations, representing a controlled research environment that may not reflect the variability encountered in real-world clinical practice. The LNH073B external validation used only 95 patients after exclusions, limiting statistical power for subgroup analyses. The authors acknowledge that training the model on larger and more diverse datasets - ideally from multiple centers with varying scanners, acquisition protocols, and reconstruction algorithms - would likely improve segmentation accuracy and prognostic concordance further.
Information loss from projection: Maximum-intensity projection is a lossy compression: multiple lesions at different depths along the projection axis can overlap, making it impossible to distinguish them in 2D. For patients with superimposed lesions (for example, separate mediastinal and thoracic spine lesions that overlap in the coronal view), the MIP may appear as a single continuous region when in reality they are distinct lesion clusters. This projection ambiguity particularly affects Dmax-type measurements, as it can make the distance between lesions appear larger or smaller than the true 3D value. The 64-71% three-category concordance (compared to 79% for binary classification) reflects this information loss at the boundaries between risk categories.
Absence of confidence scores: The current implementation produces binary segmentation masks without an associated confidence measure. In practice, clinical adoption would benefit from the model outputting a per-pixel or per-case confidence score, which could then be propagated into a confidence interval around the sTMTV and sDmax estimates and ultimately translated into a risk classification confidence. Patients whose biomarker values fall near the median cutoff would benefit most from knowing the segmentation uncertainty, as small errors in those cases can flip the risk category.
Future directions: The authors note that prospective validation in multicenter cohorts using standardized PET/CT acquisition protocols will be required before the MIP-based AI pipeline can be adopted as a clinical tool. Integration with existing lymphoma staging workflows, where MIP images are already routinely generated and reviewed, represents a natural entry point. Combining sTMTV and sDmax with clinical variables (IPI score, cell-of-origin subtype, ctDNA levels) in multivariate models could further improve prognostic discrimination beyond what either biomarker class achieves alone. Extension to other lymphoma subtypes beyond DLBCL (follicular lymphoma, Hodgkin lymphoma, mantle cell lymphoma) where TMTV has also shown prognostic relevance would broaden the applicability of the approach.