Lymphomas are the most common blood malignancies in the developed world and the ninth most common malignant neoplasm globally, with 627,439 new diagnoses and 283,169 deaths recorded in 2020. The two major categories are Hodgkin lymphoma (HL) and non-Hodgkin lymphoma (NHL). For both, accurate staging and response assessment are clinically essential, not only for guiding individual patient management but also for ensuring comparability across clinical trials evaluating treatment efficacy.
Limitations of current staging criteria: The Lugano classification, introduced in 2011 and the most widely used staging system in clinical practice, relies on two-dimensional CT measurements of up to six target lesions and on qualitative FDG-avidity scoring using the five-point Deauville scale for PET/CT. The CT component multiplies the two longest perpendicular dimensions of each lesion and sums across selected targets, producing a value proportional to cross-sectional area rather than true three-dimensional volume. This approach introduces high inter- and intra-observer variability, with studies documenting meaningful discordance in Lugano restaging classifications between readers. The Deauville-based PET criteria, meanwhile, are considered unreliable in lymphomas with variable FDG avidity such as marginal zone lymphomas, and prior work has shown that FDG-avid areas sometimes represent only approximately 25% of the total tumor mass.
The case for volumetric assessment: PET/CT-derived metabolic tumor volume (MTV), defined as the total tumor volume demonstrating high glucose metabolism above a standardized uptake value (SUV) threshold, offers several advantages over two-dimensional measurements. It eliminates the need for subjective lesion selection, removes the tedium of identifying and comparing previously selected lesions at follow-up, and has been validated as an independent prognostic factor in certain lymphoma subtypes including diffuse large B-cell lymphoma (DLBCL). However, PET/CT carries higher radiation exposure than standard CT, is substantially more expensive, and is less widely available, particularly for routine follow-up in a pediatric population that requires multiple examinations over time.
The proposed solution: This 2023 paper from a group at Poznan University of Medical Sciences presents a fully automated approach to segmenting thoracic lymphoma on contrast-enhanced CT scans using the nnU-Net deep learning framework. The goal is to enable three-dimensional volume calculation from routine CT alone, avoiding the radiation, cost, and availability barriers of PET/CT while reducing the observer variability inherent in manual two-dimensional Lugano measurements. Critically, the authors also release their annotated pediatric dataset publicly to facilitate further research in this underserved area.
The dataset consisted of 30 contrast-enhanced CT scans from 30 distinct pediatric patients diagnosed with any type of lymphoma, all hospitalized at Karol Jonscher University Hospital in Poznan, Poland between 2013 and 2020. The cohort included 17 boys and 13 girls with a mean age of 12.8 years (standard deviation 4.1, range 2 to 17 years). Only one CT scan per patient was included to maximize variance in the training data and improve potential model generalizability. Scans were acquired on two Siemens SOMATOM scanner models (Definition AS+ and Force) using Visipaque as the contrast agent, administered at 1 mL/kg for children under 35 kg or 40 to 50 mL for those over 35 kg, at injection rates of 1.5 to 2.5 mL/s, all in the portal venous phase. Median scan size was 296 by 512 by 512 voxels.
Data partitioning: Five patients were randomly selected as a held-out test set, and the remaining 25 were used for training and cross-validation. This 5-patient test set is small, which the authors acknowledge as a limitation, but it is consistent with the overall dataset size constraint. The dataset was deliberately designed to represent real-world variance in pediatric lymphoma presentations rather than being curated to contain only clear-cut cases.
Manual segmentation process and challenges: Two radiology residents performed the manual segmentations using 3D Slicer software, which served as the ground truth for model training. The segmentation process surfaced several practically significant challenges that the team documented systematically. Cervical lymph nodes were included whenever distinguishable from surrounding soft tissues. When the border between lymphoma and thymic tissue was unclear, the thymus was included in the segmentation rather than excluded. Pericardial and pleural effusions were excluded where possible, while liquefactive necrosis within the tumor was included, though the authors note this distinction was difficult in some cases. Abdominal lymph nodes were excluded entirely from all segmentations.
Ground truth quality and inter-observer variability: The authors raise an important epistemological point about the reliability of human-made segmentations as ground truth. Drawing on work by Tingelhoff et al., they note that 21 participants segmenting the same CT scan produced total volumes ranging from 30.9 to 47.1 cm3 for well-defined sinus structures, illustrating that even relatively unambiguous anatomical boundaries produce substantial inter-observer variability. For lymphoma borders, which are often poorly defined, this variability is likely even larger. The final model output consists of voxel-wise segmentations that can be edited by a reviewing radiologist, which partly addresses this limitation.
The U-Net architecture, first described by Ronneberger et al. in 2015, has become the dominant framework for segmentation tasks in medical imaging. Its encoder-decoder structure with skip connections enables the network to combine high-resolution spatial information from the encoder path with semantic context from the decoder path, producing accurate pixel-level (and voxel-level) segmentation maps. Numerous modifications and extensions have been proposed since 2015, but parameter selection across the many possible configurations (2D vs. 3D, patch size, batch size, learning rate, data augmentation strategy) is typically dataset-specific and resource-intensive to optimize manually.
nnU-Net as a self-configuring framework: Isensee et al. developed nnU-Net (no-new-UNet) in 2018 as a self-adapting framework that automatically derives a well-performing configuration from the statistical properties of the input dataset. The system divides its parameters into three categories: fixed parameters (learning rate, loss function, architecture template, optimizer, data augmentation, training and inference procedures), rule-based parameters derived from dataset properties (intensity normalization, image target spacing, network topology, patch size, batch size, and 3D U-Net Cascade trigger), and empirical parameters tested during cross-validation (postprocessing configuration and ensemble selection). This automation removes a major barrier for researchers without extensive ML engineering expertise.
Three U-Net configurations explored: nnU-Net evaluates three architectural variants. The 2D U-Net processes each CT slice independently. The 3D U-Net operates on full volumetric 3D images by cropping 3D patches that span voxel space. The 3D U-Net Cascade uses a two-stage approach: first generating segmentations on a downsampled version of the image, then refining them using both the initial segmentation and the full-resolution image as inputs. The Cascade configuration is skipped if the standard 3D U-Net patch size can already accommodate a large proportion of the input image. Performance of all three configurations is evaluated through 5-fold cross-validation before the best model or ensemble is selected for inference.
Key training parameters: For this dataset, the 2D configuration used a patch size of 512 by 512 with a batch size of 12, while the 3D configuration used a patch size of 96 by 160 by 160 with a batch size of 2. Both configurations used 16-bit float precision, a maximum of 1,000 epochs, 250 batches per epoch, an initial learning rate of 0.01, stochastic gradient descent (SGD) with momentum of 0.99, and weight decay of 0.00003. Training required approximately 2 days per fold per model configuration, running on 10 CPUs and a single NVIDIA Tesla V100 GPU. Version 1.7.0 of the nnU-Net package was used. The Dice coefficient was used as the primary evaluation metric, selected because it is more sensitive to total segmented volume than to precise border reproduction, which aligns with the clinical goal of volumetric comparison over time.
Across the five model configurations evaluated by nnU-Net during cross-validation on the 25-patient training set, results were tightly clustered. The standard 3D U-Net achieved the highest average Dice coefficient at 0.7262, followed by the ensemble of 2D and 3D U-Net (0.7221), the ensemble of 2D U-Net with 3D Cascade (0.7203), the 2D U-Net alone (0.7065), the ensemble of 3D U-Net with 3D Cascade (0.7148), and the 3D U-Net Cascade alone (0.7024). The consistently similar performance across configurations suggests that the nnU-Net framework is robust to architectural choice for this dataset, and the 3D U-Net was selected as the best single configuration for test set evaluation.
Test set results: On the five held-out test patients, the 3D U-Net achieved a mean Dice coefficient of 0.81 (standard deviation 0.17). The individual results were: Patient 1, Dice 0.88, manual volume 288.79 cm3, predicted volume 257.68 cm3 (difference 31.11 cm3); Patient 2, Dice 0.73, manual volume 631.34 cm3, predicted volume 865.01 cm3 (difference 233.67 cm3); Patient 3, Dice 0.92, manual volume 776.99 cm3, predicted volume 686.14 cm3 (difference 90.85 cm3); Patient 4, Dice 0.55, manual volume 146.19 cm3, predicted volume 331.21 cm3 (difference 185.02 cm3); Patient 5, Dice 0.95, manual volume 354.63 cm3, predicted volume 352.09 cm3 (difference 2.54 cm3).
Analysis of the outlier (Patient 4): The notably low Dice score for Patient 4 (0.55) was caused by the model incorrectly segmenting brain tissue as tumor. This patient had an unusually broad CT scanning range that extended further cranially than the other training cases, placing the brain within the model's field of view. The model had not encountered this configuration during training and mistakenly identified a portion of the brain as lymphoma, inflating the predicted volume from 146.19 to 331.21 cm3. The authors note that limiting the scanning range for this case with simple postprocessing would increase the Dice score to 0.92, but deliberately refrained from such modification to preserve the integrity of the test set.
Visual quality of segmentations: Three-dimensional renderings comparing human and nnU-Net segmentations revealed that automated segmentations had notably smoother boundaries than manual ones. The authors interpret this as reflecting the inherent subjectivity and inconsistency of manual border delineation, and suggest that the smoother automated outputs may in some respects be more reproducible than the ground truth itself. This is a recurring observation in deep learning segmentation literature: networks trained on human annotations tend to produce cleaner outputs than their training targets.
The clinical motivation for replacing or supplementing Lugano two-dimensional measurements with automated volumetric computation extends beyond technical convenience. Three-dimensional tumor volume captures the full extent of lymphoma burden in a way that the sum of products of diameters cannot, particularly for irregularly shaped mediastinal masses that may be elongated or bilobed in ways that distort two-dimensional cross-sectional approximations. As the authors note, the value produced by Lugano CT criteria is proportional to axial surface area rather than volume, making it a less intuitive and less precise surrogate for true tumor burden change over time.
Comparison with MTV from PET/CT: The relationship between automated CT volume and PET-derived MTV is an important area of prior investigation. Studies in Hodgkin lymphoma have shown that the evolution of MTV and total tumor volume during treatment is similar, and that both FDG-avid and non-avid areas within the tumor tend to shrink at comparable rates under treatment. This supports the premise that CT-based volume tracking, even without metabolic information, may be a valid surrogate for monitoring treatment response in pediatric lymphoma populations where repeated PET/CT is impractical or undesirable.
Comparison with prior segmentation work: To the best of the authors' knowledge, this study represents the first published application of automated machine learning-based volume calculation specifically for thoracic lymphoma. Prior automated segmentation work on CT has focused on anatomical structures (colon, lung, multi-organ segmentation covering 104 structures) or on lymph node detection tasks where the primary endpoint was detection rate rather than volume accuracy. The CNN-based approach by Iuga et al. for thoracic lymph node detection achieved a detection rate of 69.9% on their validation set but targeted small non-pathological nodes rather than large lymphoma masses. The nnU-Net approach presented here is the first to target volumetric quantification of pathological thoracic lymphoma on CT.
Practical implications for radiologist workflow: A fully automated segmentation tool that produces editable three-dimensional output could fundamentally change how radiologists report lymphoma CT studies. Rather than manually measuring up to six lesion diameters per scan at each staging and restaging evaluation, a radiologist could review and correct an automatically generated 3D segmentation and receive a precise volume calculation in return. This would reduce reading time, eliminate inter-observer measurement variability, and generate a volume metric that is directly comparable across time points without requiring consistency in which specific lesions are selected as targets.
Small dataset and limited diversity: The most significant limitation of this study is the dataset size. With only 30 CT scans from 30 patients, of which five were withheld for testing, the 3D U-Net was trained and cross-validated on 25 cases. This is at the lower end of what is typically considered sufficient for training robust deep learning segmentation models, even for a relatively constrained anatomical region (thorax) and target structure (lymphoma). The authors acknowledge that this limitation was responsible for at least two observed failure modes: cases where the algorithm mistook solid organs (brain in Patient 4, iliopsoas muscle in other instances) for tumor tissue due to atypical scanning scope, and inconsistency across cross-validation folds in how the model handled ambiguous structures like pleural effusion and thymic tissue.
Ground truth reliability: Manual segmentation by two radiology residents introduces inter-observer variability directly into the training labels. The authors reference Tingelhoff et al.'s illustration of this problem: 21 trained and untrained participants segmenting the same clearly bounded anatomical structure produced volume estimates spanning 30.9 to 47.1 cm3, a range of more than 50%. For lymphoma with poorly defined, infiltrative borders, the variability in the ground truth segmentations is likely substantially larger. Depending on the cross-validation fold and therefore the training subset composition, the model showed inconsistent behavior toward borderline structures including pleural effusion, liquefactive necrosis, and normal thymus, suggesting that these label inconsistencies propagated into the learned representations.
Lack of external validation: All evaluation was performed on cases from a single institution (Karol Jonscher University Hospital, Poznan) using a consistent imaging protocol across two scanner models from the same manufacturer (Siemens). Performance on CT scans from different institutions using different scanners, contrast agents, injection protocols, or reconstruction kernels is unknown. External validation is a standard prerequisite before clinical deployment of any medical AI tool, and the absence of a geographically or institutionally distinct test set means the reported Dice scores may overestimate the model's generalizability.
No external dataset for validation: The authors note that, to their knowledge, no publicly available large dataset of segmented CT scans from lymphoma patients existed at the time of publication. The largest comparable dataset in the lymph node segmentation literature (Roth et al.) contained 176 abdominal CT scans, but manual inspection showed that many lymph nodes in the region of interest were not segmented, and the focus was on abdominal rather than thoracic pathological masses. This absence of benchmarking data also made it impossible to compare the nnU-Net performance against competing approaches on a common reference set.
A central contribution of this paper beyond the segmentation model itself is the public release of the annotated dataset. The 30 pediatric CT scans with corresponding manual segmentations are deposited on Zenodo, alongside the trained model weights from the best-performing nnU-Net configuration and the GitHub-hosted inference code. The authors describe this as the first openly available labeled dataset for pediatric lymphoma segmentation on CT, a claim that highlights both the significance of the contribution and the underdevelopment of this specific research area.
Why public datasets matter in medical AI: As Varoquaux and Cheplygina have argued in their analysis of machine learning methodological failures in medical imaging, dataset availability is one of the strongest determinants of where research attention focuses within a field. The absence of shared, annotated datasets for pediatric lymphoma CT has meant that the limited research effort on this problem has been fragmented across institutions, each starting from scratch with its own small private dataset and making external validation or comparison essentially impossible. By releasing both data and model, the Poznan group enables other researchers to use the dataset for alternative approaches, train models on the combined pool if additional annotated cases are contributed elsewhere, and reproduce the published results precisely.
Prospective value for model improvement: The authors explicitly encourage international scientific cooperation to grow the dataset. They hypothesize that increasing dataset size would primarily reduce performance variance on the test set (pointing to Patient 4's outlier Dice of 0.55 as a variance-driven failure) and reduce dependence on the specific characteristics of the training subset. Incorporating segmented whole-body CT scans from healthy individuals, or standardizing the cranial extent of the scanning range across cases, would also address the misclassification of non-thoracic organs seen in this initial dataset.
Ethical and regulatory considerations: The study and data release were approved by the local Bioethics Committee and the head of the Radiology Department at Poznan University of Medical Sciences. Patients and their legal guardians consented to retrospective research use of the imaging data, and all data were fully anonymized before release. The study was funded by a Polish National Science Centre grant (2020/36/T/NZ5/00515).
Dataset expansion and external validation: The most immediate priority identified by the authors is expanding the training dataset through multi-institutional collaboration. A larger and more diverse dataset would address the core limitation of this work: the 25-case training set is too small to achieve robust generalization across the full spectrum of pediatric lymphoma presentations, scanner hardware, and institutional protocols. Federated learning approaches, which allow model training across multiple institutions without requiring raw data sharing, represent a technically feasible path for building larger effective datasets while respecting patient privacy and institutional data governance requirements.
Whole-body and extended-range CT integration: The failure mode observed in Patient 4, where the model misclassified brain tissue due to an atypically wide scan range, could be addressed by including whole-body CT data from healthy controls in training or by incorporating a scan range normalization step in preprocessing. The nnU-Net framework supports multi-task and multi-modal configurations that could in principle combine thoracic-focused lymphoma segmentation with anatomical structure recognition to constrain the output to biologically plausible locations.
Prospective clinical validation and outcome correlation: The authors call for prospective studies examining whether automated volumetric measurements correlate with clinical progression and remission in a way that justifies substituting volume for the current Lugano criteria. The relationship between total CT volume at diagnosis and end-of-treatment volume change, and their association with event-free survival and overall survival, needs to be established in a prospectively designed cohort with sufficient follow-up before volume-based response criteria can be proposed for clinical or regulatory use. This is analogous to the validation pathway that established PET/CT-derived MTV as a prognostic biomarker in adult DLBCL.
Integration with PET/CT and multimodal fusion: Given the complementary information in anatomical (CT) and metabolic (PET) imaging, future work could explore joint segmentation models trained on PET/CT data to produce both volumetric and metabolic tumor volume estimates simultaneously. Such a combined model could retain the prognostic advantages of MTV while reducing dependence on any single imaging modality and enabling cross-validation between the two volume estimates. For pediatric patients specifically, where reducing cumulative radiation exposure is a priority, hybrid low-dose PET/CT protocols that the authors mention as a potential future direction may eventually make combined volumetric PET/CT segmentation both feasible and justified for routine follow-up.