Diffuse large B-cell lymphoma (DLBCL) is the most common histologic subtype of non-Hodgkin lymphoma, with approximately 150,000 new cases diagnosed worldwide each year. Roughly 60% of patients are cured with standard anthracycline-containing immunochemotherapy regimens such as R-CHOP (rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone) or dose-adjusted EPOCH-R. However, the remaining 40% who are refractory or relapse early face a very poor prognosis: median overall survival in refractory DLBCL is only 6.3 months. Identifying which patients are at high risk before or early in treatment is therefore a pressing clinical need.
The role of PET/CT and MTV: 18F-fluorodeoxyglucose (FDG) PET/CT is the standard imaging modality for staging and treatment response assessment in DLBCL, guided by the Lugano classification. Metabolic tumor volume (MTV), calculated from FDG-PET/CT by summing the volumes of all FDG-avid lesions above a defined uptake threshold, has emerged as a robust prognostic biomarker. High baseline MTV is consistently associated with inferior progression-free and overall survival across frontline, relapsed, and CAR-T cell therapy settings. The International Metabolic Prognostic Index, developed by Mikhaeel and colleagues, integrates MTV with age and stage from the IPI score to generate individualized survival estimates, signaling that MTV is on the verge of routine clinical adoption.
The measurement problem: Despite its prognostic value, MTV has not been widely integrated into clinical practice largely because measuring it is laborious. Currently available semiautomatic software requires a nuclear medicine reader to manually review and adjust region-of-interest (ROI) contours in three planes, identify and include all lesions across the whole body, and exclude physiologic structures with naturally high FDG uptake such as the brain, heart, kidneys, and bladder. For a patient with widespread DLBCL, this process can take 20 minutes or more per scan. The process also introduces inter- and intra-observer variability, limiting reproducibility across centers and readers.
This paper from the University of Miami Sylvester Comprehensive Cancer Center addresses this problem directly by developing and validating a fully automated method (AM) for MTV calculation in DLBCL using deep learning. The study tests whether the automated pipeline can replicate the results of two experienced, blinded nuclear medicine readers on a cohort of 100 patients from a major multicenter clinical trial.
The study cohort was drawn from the Alliance/CALGB 50,303 trial (NCT00118209), a phase III intergroup randomized study comparing dose-adjusted EPOCH-R versus standard R-CHOP as frontline therapy for newly diagnosed DLBCL. This trial enrolled 491 eligible patients across 17 participating centers in the United States. The final published trial results showed that EPOCH-R was more toxic and did not improve progression-free survival or overall survival compared with R-CHOP, making R-CHOP the retained standard. Of the 491 enrolled patients, 155 had whole-body FDG-PET/CT scans at study enrollment publicly available through The Cancer Imaging Archive (TCIA). The authors randomly selected 100 of these 155 patients for the present study.
Scanner diversity: A key strength of this cohort is its imaging heterogeneity, reflecting real-world multi-institutional practice. PET/CT images were acquired on three different scanner platforms: Siemens Biograph (n=53), GE Discovery (n=30), and Philips GEMINI (n=17). Reconstruction algorithms varied across sites, ranging from ordered-subset expectation maximization (OSEM) to blob-based iterative time-of-flight (BLOB-OS-TF) and point-spread-function (PSF) modeling with and without time-of-flight. PET slice thickness ranged from 2 mm to 4.25 mm, with the most common being 3.25 mm or 4.25 mm (83% of cases). This variability in acquisition and reconstruction parameters is a meaningful real-world challenge for automated analysis methods.
PET acquisition protocol: Per the trial protocol, patients fasted for at least 4 hours and confirmed plasma glucose below 200 mg/dL before receiving an intravenous injection of 8-20 mCi of FDG. PET/CT scans were acquired approximately 60 to 80 minutes post-injection. Concomitant low-dose CT scans for attenuation correction and anatomic localization extended from skull base to thighs, acquired at 110-140 kVp with a reference dose of 200 mAs and reconstructed with slice thickness ranging from 2 mm to 4 mm.
Additional training data: To fine-tune the deep learning segmentation model, the authors used a supplementary dataset of 50 whole-body CT scans from the TCIA whole-body FDG-PET/CT dataset (Gatidis et al.), for which two expert radiologists had provided consensus contours of the brain, heart, kidneys, and bladder. This IRB-waived study used only publicly available, de-identified patient data.
The automated MTV pipeline begins by solving a prerequisite problem: automatically identifying and masking the anatomic structures with intense physiologic FDG uptake that would otherwise be falsely included in the MTV measurement. These structures are the brain, heart, kidneys, and bladder. To segment these four organs on the CT component of the PET/CT scan, the authors deployed a 2D dilated residual U-Net architecture developed by Manteia Medical Technologies (FDA-cleared under K191928).
Network architecture details: The U-Net follows an encoder-decoder structure where both branches are composed of five cascades of residual blocks. Residual connections within each block were chosen specifically to address the vanishing gradient problem that afflicts very deep networks during training. Each residual block contains two 3x3 convolution layers. The encoder performs spatial downsampling via maximum pooling, while the decoder upsamples via bilinear interpolation. Shortcut connections between corresponding encoder and decoder feature maps, a hallmark of U-Net architecture, help preserve spatial detail lost during downsampling. Batch normalization is applied throughout to stabilize training by reducing internal covariate shift.
Fine-tuning process: Rather than training from scratch, the authors fine-tuned a pre-trained version of the model by resetting only the final output layer weights to random values, leaving all other weights from pre-training intact. This transfer learning approach resulted in only 165 trainable parameters, dramatically reducing the data requirements. The 50 annotated whole-body CT scans were split 5:1:4 for training, validation, and testing. Data augmentation used affine transforms including rotation, translation, scaling, and flipping. Training ran for up to 100 epochs, with a learning rate initialized at 3x10^-4 and decayed to 3x10^-6 after approximately 60 epochs. The objective function combined cross-entropy loss and Dice loss, optimized using the Adam optimizer with a weight decay of 1x10^-4. Training loss converged dramatically for all four organs, from initial values around 1.3-1.6 down to 0.02-0.08.
Segmentation accuracy: The Dice coefficient, which measures overlap between predicted and ground-truth contours, reached 0.9885, 0.9441, 0.9145, and 0.9045 for brain, heart, kidneys, and bladder, respectively, on the training/validation set. On the held-out test set, Dice scores were 0.9524 (brain), 0.9023 (heart), 0.9107 (kidneys), and 0.8809 (bladder), all of which are clinically acceptable for this application. The framework was implemented in PyTorch v1.10.
Once the deep learning model has segmented the physiologically avid structures on the CT, those contours must be transferred onto the corresponding PET scan before MTV can be computed. This transfer is not straightforward because the spatial characteristics of CT and PET differ significantly (resolution, slice thickness, voxel size). The pipeline addresses this using a suite of image-processing algorithms including region-growing, active contours, and fast marching methods to adapt the CT-derived contours to their PET representations, accounting for the blurring and spatial extent differences inherent to PET imaging.
Handling the ureter problem: The ureters present a specific challenge: they carry radioactive urine from the kidneys to the bladder and can show focal FDG activity that is difficult to distinguish from tumor. The deep learning segmentation model underperforms on ureters due to limited annotated training data and the wide anatomical variability of ureteral course. To handle this, the authors defined a trapezoid-shaped exclusion zone extending cranio-caudally from the superior poles of the kidneys to the mid-plane of the bladder, and laterally between the midlines of the two kidneys (on the top base) and the lateral borders of the bladder (on the bottom base). Any candidate ROI falling within this zone is subject to additional scrutiny, and scattered or isolated uptake in this region resulting from the kidneys or bladder is excluded.
MTV thresholding and lesion identification: Within the whole-body PET volume after masking the excluded organs and zones, MTV is determined using a threshold of 41% of the SUVmax, a method that has been validated and widely adopted in lymphoma PET research. All contiguous voxels exceeding this threshold are clustered into isolated regions of interest. A minimum volume constraint of 1 cm3 is applied, retaining only ROIs above this size. ROIs smaller than 2 cm3 that fall within the trapezoid exclusion zone are discarded. If a retained candidate ROI has its own SUVmax screened out by the algorithm, its volume is removed from the MTV tally and the process iterates until all retained ROIs satisfy the criteria. The entire process requires no manual intervention at any step.
Comparison against semiautomatic readers: Two experienced nuclear medicine readers independently reviewed all 100 FDG-PET/CT scans using the Hermes Affinity Viewer software. Each reader manually adjusted ROI contours in three planes, excluded adjacent physiologic FDG-avid structures, and applied the same 41% SUVmax threshold and 1 cm3 minimum volume constraint. Bone marrow involvement was included only if focal; splenic involvement was included if there was focal uptake or diffuse uptake greater than 150% of liver background. Readers were blinded to each other's results and to the automated method output.
The primary analysis compared MTV and SUVmax values calculated by the automated method (AM) against those from two blinded nuclear medicine readers across all 100 DLBCL patients. The mean MTV across the cohort was 226.470 mL (SD 260.066, CV 114.834%) for reader 1 and 226.799 mL (SD 261.965, CV 115.505%) for reader 2. The AM produced a mean MTV of 205.704 mL (SD 245.825, CV 119.504%), slightly lower than both readers but with comparable variance. No patients were excluded due to failure of the automated method, confirming 100% pipeline completion.
Correlation and concordance statistics: As a benchmark, reader 1 versus reader 2 showed near-perfect agreement: Pearson's r = 0.9997 (p less than 0.0001) and ICC = 1.0 for both MTV and SUVmax. Comparing reader 1 to AM, the Pearson's correlation coefficient was 0.9814 (p less than 0.0001) and the ICC was 0.98 (p less than 0.001; 95% CI 0.96-0.99) for MTV; for SUVmax, r = 0.9868 and ICC = 1.0 (95% CI 0.99-1.0). Comparing reader 2 to AM produced essentially identical results: r = 0.9818 and ICC = 0.98 (95% CI 0.96-0.99) for MTV; r = 0.9868 and ICC = 1.0 for SUVmax. These correlation levels are considered excellent by standard clinical agreement thresholds.
Bland-Altman analysis: Bland-Altman plots, which assess systematic bias across the range of measurements rather than just linear correlation, showed only relatively small systematic errors between the AM and both readers for both MTV and SUVmax. For MTV, the RMSE between readers (averaged) and AM was 54.7 mL, with a positive bias of 28.4 and negative bias of 0.27. The mean difference between readers was 20.92 mL (95% limits of agreement: -49.77 to 91.63). The AM consistently produced slightly smaller MTV values than the expert readers, suggesting a small but systematic underestimation tendency rather than random noise. For SUVmax, the RMSE was 1.93, with a mean difference between readers of -0.03 (95% limits of agreement: -3.34 to 3.26).
Performance by scanner type: When stratified by PET/CT scanner platform, the only notable discrepancy emerged for SUVmax on Philips GEMINI scanners: the ICC between readers and AM was 0.81 (p less than 0.0001; 95% CI 0.57-0.93), which, while still statistically strong, was lower than for Siemens and GE systems. No scanner-type differences were observed for MTV volumes across any of the three platforms. The somewhat lower SUVmax agreement on Philips scanners likely reflects their distinct reconstruction algorithm (BLOB-OS-TF) producing different image characteristics compared to the other two platforms.
Several automated and semi-automated methods for MTV calculation in DLBCL have been proposed before this paper, and the authors position their approach against a field that is rapidly advancing. Capobianco and colleagues tested the PARS (PET-Assisted Reporting System) prototype from Siemens in 301 DLBCL patients from the REMARC trial, finding a correlation of rho = 0.76 between automated and reader-derived MTV. Jiang and colleagues trained a 3D U-Net on patches from 414 DLBCL patients, reporting R2-linear = 0.882 (training) and 0.939 (validation). Most recently, Revailler and colleagues used deep learning to complete a training dataset of 407 patients in only 93 hours. By contrast, this study's AM achieved Pearson r of 0.981 and 0.982 against its two readers, numerically outperforming or matching prior approaches while using a smaller primary cohort of 100 patients.
Interpretability and clinical emulation: A key differentiating argument is that most prior automated methods function as "black box" models that provide little insight into how decisions are made. The proposed pipeline, by contrast, explicitly emulates the reasoning process of nuclear medicine physicians. It identifies and excludes physiologically avid normal structures by name (brain, heart, kidneys, bladder), handles the ureter problem through a clinically motivated anatomical zone, and applies the same 41% SUVmax threshold and 1 cm3 volume cutoff that expert readers use. This step-by-step parallel to clinical practice makes the outputs more interpretable and the failure modes more predictable.
Reduced annotation dependency: Prior deep learning methods that learn directly from annotated PET images require large training datasets of labeled lymphoma lesions, which are expensive and time-consuming to create and may embed reader bias into the model itself. This pipeline avoids this problem: the deep learning component only learns to segment normal organs on CT (a comparatively well-defined task), while the MTV calculation itself is performed by explicit rule-based logic rather than a learned black-box mapping. This design means the model does not need annotated tumor data for training and is less vulnerable to compounding the inter-observer variability it is designed to eliminate.
Speed advantage: The median processing time for the automated method was 5 minutes per patient, compared to approximately 20 minutes for expert semiautomatic analysis. For a clinical trial involving hundreds of patients, this represents a 4-fold improvement in throughput that could make routine MTV analysis in clinical research settings operationally feasible for the first time.
Single-software validation standard: The automated method was validated exclusively against readings generated with the Hermes Affinity Viewer, a single semiautomatic software package. While Hermes is widely used and generally accepted, MTV values can vary systematically across different semiautomatic platforms (e.g., MIM Software, LIFEx, PET-Edge) due to differences in segmentation algorithms and user interface design. The absolute MTV values reported in this study should be interpreted in the context of the specific software used as the reference standard, and performance against readers using different platforms may differ.
Single-institution readers: Both nuclear medicine readers were from the same institution, which introduces the possibility of shared reader tendencies that may not represent the full spectrum of inter-reader variability encountered in true multicenter practice. The near-perfect reader-to-reader correlation (r = 0.9997) observed here is higher than what is typically reported across institutions with different training backgrounds, equipment, and protocols, and may reflect this shared institutional context rather than universal agreement.
Sample size and generalizability: Although the 100-patient cohort was randomly selected from a homogeneous clinical trial dataset and results were consistent across the cohort, it remains relatively small for a deep learning validation study. More importantly, the study does not include patients with lymphoma subtypes other than DLBCL, limiting its immediate applicability to follicular lymphoma, mantle cell lymphoma, T-cell lymphomas, or Hodgkin lymphoma, all of which differ in FDG avidity patterns and lesion distribution.
Edge cases and anatomical complexity: The authors acknowledge that the pipeline may perform less well in scenarios where tumor lesions lie in close proximity to normal physiologically avid structures such as the bladder or kidneys, or when normal anatomy is distorted by disease or imaging artifacts including misregistration or patient motion. These are relatively uncommon clinical scenarios but are precisely the cases where automated tools are most likely to fail and where physician oversight remains most important. The method was also not tested on post-treatment follow-up scans, where physiologic uptake patterns may differ from baseline.
The central conclusion of this paper is that a deep learning-aided, fully automated method can calculate metabolic tumor volume in DLBCL with a high degree of concordance to expert nuclear medicine readers, as quantified by ICC values of 0.98 and Pearson correlations above 0.98 against both readers. The authors argue that this level of performance is sufficient to enable the integration of MTV as a standard PET biomarker in clinical research settings, where it has historically been underutilized precisely because of the measurement burden it imposes.
Enabling large-scale biomarker studies: The 4-fold reduction in processing time from 20 minutes (manual) to 5 minutes (automated) means that a retrospective analysis of 500 PET/CT scans - the scale of many major lymphoma clinical trials - could be completed in approximately 42 hours of compute time rather than 167 hours of expert reading time. This unlocks the feasibility of using MTV as a secondary endpoint or stratification variable in trials where manual MTV calculation would be logistically impractical.
Next steps identified by the authors: Future work should extend validation to other lymphoma subtypes and cancer groups beyond DLBCL. The method should be tested prospectively on scans from centers not represented in the training pipeline to assess cross-institutional generalizability. Performance should be evaluated on post-treatment and interim-treatment PET/CT scans in addition to baseline staging scans. Additionally, the clinical utility of the automated MTV values should be assessed directly by linking them to clinical outcomes such as progression-free survival and overall survival to confirm that the automated measurements retain the prognostic signal established for manually derived MTV in the literature.
Broader implications: The pipeline described in this paper represents a practical model for how deep learning can be deployed in clinical oncology imaging: not by replacing physician judgment wholesale with a black box, but by automating the most time-consuming and variable components of a measurement process while preserving the interpretable, rule-based logic that physicians already apply. As MTV moves toward routine integration in DLBCL prognostic models and treatment algorithms - including the newly developed International Metabolic Prognostic Index - fully automated tools like this one will be essential for making those models operationally feasible at the scale required for routine clinical use.