Malignant lymphomas are a heterogeneous family of hematologic cancers arising from clonal lymphocyte proliferation, broadly divided into Hodgkin lymphoma (HL) and non-Hodgkin lymphoma (NHL). 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) has become the standard imaging modality for staging, treatment response assessment, and prognostication across both groups. The modality's functional advantage is that it measures glucose metabolism in lesions, producing quantitative metrics that carry significant prognostic weight.
The segmentation bottleneck: Accurate quantification of lymphoma disease burden on PET/CT depends on segmenting all metabolically active lesions throughout the body. This process yields critical parameters: metabolically active tumor volume (MTV), SUVmean, total metabolic tumor volume (TMTV, the sum of MTV across all lesions), total lesion glycolysis (TLG), whole-body metabolic burden (WBMB), metabolic heterogeneity (MH), and lesion dissemination (Dmax). Manual segmentation of these lesions, however, requires 30 to 45 minutes per patient depending on disease burden, and carries high inter- and intra-observer variability across different operators and segmentation methods. The challenge is amplified by the normal biodistribution of 18F-FDG, which creates physiologically intense activity in the brain, renal collecting ducts, and bladder, confounding intensity-based detection approaches.
Scope of this review: Published in 2022 and authored by investigators at the National Institutes of Health Clinical Center and the BC Cancer Research Institute, this scoping review systematically maps the evidence base for AI applications in lymphoma PET imaging. The review covers detection, classification, segmentation, radiophenomics, and prediction and prognostication tasks, drawing on studies from 2009 through September 2021. The paper was conducted following PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analysis extension for Scoping Reviews) guidelines, ensuring methodological rigor in evidence mapping.
The authors frame three critical clinical objectives that AI must serve in lymphoma PET: (1) quantification of disease burden, (2) evaluation of therapy response, and (3) extraction of additional prognostic and diagnostic information from imaging features. Each of these objectives maps directly onto specific AI tasks examined throughout the review.
The literature search was conducted across four databases: PubMed/MEDLINE, EMBASE, Cochrane Library, and Google Scholar, with a search cutoff of September 1, 2021. In PubMed/MEDLINE, the query combined Medical Subject Headings (MeSH) terms for artificial intelligence (or deep learning, machine learning, support vector machine, convolutional neural network, artificial neural network) with terms for positron emission tomography (PET, PET-CT, or PET-MR) and "lymphoma." Additional records were identified through manual bibliography and citation searches until no further relevant references were found. A single investigator independently screened all titles and abstracts before selecting articles for full-text review.
Eligibility criteria: Studies were included if they reported the diagnostic measurement of an AI, machine learning, or deep learning algorithm applied to any lymphoma subtype using PET imaging in human populations. Articles had to be published in English and in peer-reviewed journals. Non-peer-reviewed conference papers, pre-print archives, and studies without direct relevance to AI in lymphoma PET imaging were excluded. The search covered all primary articles published since 2009, corresponding to the period when deep learning approaches began entering medical imaging research.
Screening results: The initial search retrieved 1,122 documents. After removing ineligible records at the title and abstract stage, 1,089 met criteria for abstract review. Of these, 75 met criteria for full-text review. Following full-text screening, 20 articles were included in the final analysis. These 20 papers spanned both algorithm development studies and clinical assessment studies, covering AI tasks including detection, classification, segmentation, characterization, prediction, prognostication, and combinations thereof on PET/CT or PET-only images.
Data extraction: For each included study, the authors recorded the AI task, model architecture, input data type, output, ground truth (GT) definition, sample size, training and test split sizes, and figures of merit (FoM) relevant to the task. FoMs included sensitivity, specificity, Dice Similarity Coefficient (DSC), Hausdorff distance (HD), and Area Under the Curve (AUC) as appropriate for each task. Notably, the authors emphasize that non-uniformity in GT definitions across studies introduces a degree of uncertainty when comparing performance metrics, underscoring the need for standardized GT approaches in future work.
Detection in the context of PET imaging refers to localizing an area within the image that contains an object of interest with a stated level of certainty. This task combines localization with a degree of classification: the input is image data (pixels or voxels), and the output is a set of locations containing suspicious lesions. A critical complication unique to PET imaging is the presence of High Normal Activity (HiNA) regions, where physiologic 18F-FDG uptake (for example in the brain, heart, kidneys, renal collecting ducts, and bladder) can mimic malignant lesions or obscure adjacent pathology. Effective detection AI must therefore either jointly model HiNA regions or suppress them as a preprocessing step.
Seven studies on detection: The review identified 7 studies performing lymphoma lesion detection on PET/CT. Bi and colleagues (2017) used a multi-scale superpixel encoding CNN that took 3D whole-body coronal 18F-FDG PET with CT slices as two input channels, outputting regions classified as High Normal Activity or suspicious lymphoma lesions. Sibille and colleagues (2020) proposed a model that simultaneously detected HiNA and lesions suspicious for lung cancer and lymphoma. Wiseman and colleagues (2021) deployed a 3D DeepMedic model that implicitly learned HiNA regions during training and achieved 85% true positive rate (TPR) on average for lesion detection, though performance was notably better above the diaphragm than below, because of greater HiNA heterogeneity in the lower abdomen (bladder, kidney, ureter).
HiNA suppression strategies: To improve model performance, HiNA regions can be excluded from the training pipeline either manually or automatically as a pre- or post-processing step. For instance, Yu and colleagues used a semi-automatic approach where HiNA regions were first identified and removed, followed by AI-based lymphoma lesion detection. An ideal end-to-end detection algorithm would combine HiNA identification and lymphoma lesion detection into a single unified pipeline, removing the need for separate HiNA processing steps. Additional detection studies from the review include Yuan and colleagues (2021), Zhou and colleagues (2021), and Hu and colleagues (2019).
Zhou and colleagues specifically developed a 3D convolutional neural network for computer-aided detection of mantle cell lymphoma on 18F-FDG PET/CT, one of the few studies focusing on this less common but clinically aggressive NHL subtype. The heterogeneity of lesion size, shape, uptake pattern, and anatomical distribution in lymphoma makes detection considerably harder than in organ-confined solid tumors, and model performance metrics must be interpreted with this clinical context in mind.
Segmentation refers to delineating the boundary of a lesion given its location, producing a voxel-level membership map for each lesion. Accurate segmentation is essential because it enables extraction of both lesion-level metrics (SUVmax, SUVmean) and whole-body quantitative metrics (TMTV, TLG, Dmax) that carry prognostic significance. The review identifies two broad approaches to lesion segmentation: cascaded models that separate detection and segmentation into sequential neural network stages, and end-to-end models that receive raw image data and output segmentation maps without an intermediate detection step.
Cascaded segmentation (6 studies): In cascaded approaches, the detection stage produces a coarse probability map or single-pixel lesion locations, which are then fed as a second input into a dedicated segmentation network. This architecture explicitly divides the problem into two manageable steps, each with its own evaluation metrics. Human involvement can occur at either stage: clinicians may manually identify lesions before automated segmentation (pre-segmentation human detection), or alternatively, a fully automatic detection step generates candidate lesions and clinicians then verify and discard false positives. The latter was used by Yu and colleagues, who ran automatic detection first and then had clinicians confirm true lymphoma lesions from the candidate list. Six studies in this review used cascaded segmentation: Pinochet et al. (2021), Yuan et al. (2021), Sadik (2019), Hu et al. (2019), Yu et al. (2018), and Grossiord et al. (2017).
End-to-end segmentation (4 studies): End-to-end approaches eliminate the explicit detection stage and learn directly from raw image inputs to produce segmentation maps, offering advantages in memory efficiency and applicability in data-limited settings. Weisman and colleagues (2020) proposed an end-to-end model receiving whole-body 18F-FDG PET and CT as dual input channels, producing masked segmented lymphoma lesion maps with a Dice Similarity Coefficient (DSC) of 0.86. This DSC value represents strong overlap between AI and ground truth segmentations and is among the highest reported for whole-body lymphoma segmentation at the time. The remaining end-to-end studies were Blanc-Durand et al. (2020), Li et al. (2019), and Desbordes et al. (2016).
Fully automated disease burden assessment: A subset of studies performed fully automated end-to-end disease burden quantification without any physician intervention in lesion identification. These included Pinochet et al., Bi et al., Li et al., and Blanc-Durand et al. For instance, Grossiord et al. used a random forest (RF) classifier combined with morphologic hierarchy to extract PET and CT image features and classify lesions into three categories - organ, tumor, and non-relevant - before automatically segmenting the tumor-class lesions. Fully automated models remove the bottleneck of individual lesion identification by clinicians, which is impractical in routine workflows given the extent of lymphomatous involvement and typical clinical time constraints.
The review draws a careful distinction between two meanings of "classification" in the AI literature. Image classification, in the computer vision sense, refers to assigning a whole image or image patch to a category (for example, normal versus abnormal, or lymphoma subtype A versus subtype B). Statistical classification, by contrast, refers to using quantitative input features (radiomic features such as SUV, TMTV, TLG, entropy) to assign patients to diagnostic or prognostic groups. Both types are represented in the 20 included studies.
Image classification (5 studies): Lippi and colleagues (2019) used a machine learning algorithm to classify PET lesions into four malignant lymphoma subtypes: DLBCL, Hodgkin lymphoma, follicular lymphoma, and mantle cell lymphoma. This multi-class subtype classification from PET imaging alone, without histopathology, is a clinically ambitious target. Lartizien and colleagues (2014) took a different approach, using an SVM and random forest model trained on 12 radiomic features extracted from PET and CT scans to distinguish benign from malignant 18F-FDG-avid lesions, achieving an AUC of 0.91. Additional image classification studies included Pinochet et al. (2021), Sadik et al. (2021), Guo et al. (2021), and Grossiord et al. (2017).
Prediction and prognosis (3 studies): The review identified 3 studies using statistical classification for lymphoma prediction and prognostication. Baseline TMTV is an established prognostic factor in DLBCL, primary mediastinal B-cell lymphoma (PMBCL), and HL. Vercellino and colleagues analyzed baseline TMTV in patients aged 60 to 80 with DLBCL and found that a high baseline TMTV was associated with significantly poorer progression-free survival (PFS) and overall survival (OS). Mayerhoefer and colleagues combined entropy (a measure of glucose metabolism heterogeneity within the TMTV) with TMTV and SUVmax to predict PFS in mantle cell lymphoma, achieving an AUC of 0.72. Annunziata and colleagues used end-of-treatment 18F-FDG PET/CT features including Deauville Score, quantitative PET (qPET), and MTV, combined with CT features, and applied a neural network to predict time-to-event outcomes (relapse or progression) as a statistical classification task.
These prognostic studies illustrate a key theme: combining multiple radiomic features consistently outperforms single-feature models. Entropy as a measure of intratumoral metabolic heterogeneity, TMTV as a measure of overall disease burden, and SUVmax as a measure of peak metabolic activity are complementary dimensions of lymphoma biology that together provide stronger predictive signal than any single parameter alone.
The authors dedicate a substantial discussion section to practical considerations for the clinical transition of AI in lymphoma PET management. The dominant bottleneck today is the detection and segmentation of lesions required for TMTV and TLG calculation - tasks that are slow, variable, and currently impractical at scale in routine clinical settings. AI removes this bottleneck and opens the door to high-throughput, reproducible metabolic biomarker extraction for every patient.
Fully automated versus semi-automated pipelines: A fully automated AI model can process whole-body PET/CT without physician lesion identification, enabling streamlined integration into clinical workflows. This is especially valuable in lymphoma, where patients may have dozens of lesions distributed throughout the body and exhaustive manual identification is impractical within a typical clinical reading session. Semi-automated approaches, where the clinician either pre-identifies lesions for AI-based refinement or reviews AI-generated candidate lesions and removes false positives, offer a middle ground with potentially higher precision at the cost of retained physician time. The right approach likely varies by clinical context and the downstream application of the segmentation output.
Performance evaluation framework: The authors argue that conventional sensitivity and specificity metrics for lesion detection are insufficient for clinical decision-making, because not all lesions carry equal clinical significance. A "critical miss" of a large, metabolically active lesion has far greater consequences than a missed small peripheral node. Lumped detection sensitivity therefore does not reflect clinical acceptability. The review recommends task-specific composite figures of merit incorporating metrics such as Dmax (spatial dissemination of lesions) and MH (metabolic heterogeneity), in addition to TMTV. Weisman and colleagues' approach of characterizing CNN performance using a comprehensive panel (SUVmax, MTV, TLG, surface area/MTV ratio, Dmax) is cited as a best-practice model for multi-metric evaluation.
Ground truth standardization: A critical limitation across the 20 included studies is the non-uniformity of ground truth (GT) definitions. PET images are visually analyzed, and GT labels used for AI training and validation vary substantially in how they were generated (manual expert annotation, semi-automatic thresholding, consensus panels). This heterogeneity makes it difficult to compare model performance across studies and raises questions about whether models trained on one GT definition will generalize to clinical workflows using different annotation standards. Standardized GT protocols, ideally generated by multi-expert consensus or calibrated automated methods, are an unmet need.
One of the most forward-looking sections of the review addresses the transition from spatial (single-time-point) to spatiotemporal AI in lymphoma PET. Currently, nearly all AI methods in the field analyze a single PET scan in isolation. This is counterintuitive to clinical practice, where radiologists and oncologists routinely compare prior imaging to current scans as one of the most informative dimensions of disease assessment. Temporal changes in lesion characteristics, such as changes in SUVmax between baseline and interim PET, carry substantially more prognostic information than any single time-point measurement in isolation.
Delta radiomics: Delta radiomics formalizes the measurement of change in radiomic features across time points during or after treatment. In principle, changes in SUVmax, TMTV, MTV, TLG, and apparent diffusion coefficient (ADC) between baseline and interim assessments (DELTASUV, DELTAMTV, DELTATMTV, DELTAADC) can be quantified and fed into predictive models. CT-based delta radiomics has already demonstrated utility for predicting treatment response and side effects in lung cancer, gastric cancer, and radiation therapy toxicity. The authors argue that PET-based delta radiomics is underexplored in lymphoma and represents a high-priority direction for future AI development.
Delta imaging for heterogeneous response: AI-generated delta images, visualizing the voxel-by-voxel difference between baseline and interim PET scans, can reveal intralesional heterogeneity of treatment response. For example, when one portion of a bulky tumor shrinks while another portion grows (a heterogeneous response), a delta image can spatially map this divergence in a way that aggregate metrics like overall TMTV change cannot. This capability could identify non-responding lesion regions earlier in the treatment course, enabling targeted biopsy of the non-responding portion for mutational analysis, or the earlier application of external beam radiation or percutaneous ablation to refractory sub-volumes. Integrating temporal delta features with static radiomic features - combining MH, Dmax, and TMTV change over time - represents a more complete model of lymphoma biology than any single dimension alone.
The authors also envision that AI-based PET imaging surveillance protocols, using delta image analysis to detect early signs of recurrence in patients in remission, could eventually be embedded into lymphoma follow-up guidelines. Early and accurate detection of recurrence using quantitative delta image biomarkers could prompt earlier intervention and improve outcomes in the significant proportion of patients who relapse after initial therapy.
Beyond methodology, the review identifies data availability as a foundational barrier to progress in AI for lymphoma PET imaging. High-quality, large, and diverse imaging datasets are essential for training robust models and for enabling the generalizability needed for real-world clinical deployment. Lymphoma's biological heterogeneity across subtypes (HL vs. NHL, DLBCL vs. follicular vs. mantle cell vs. extranodal NK/T cell, among many others) means that a dataset curated at a single institution captures only a narrow slice of the disease's clinical spectrum. Models trained on such narrow datasets routinely perform well internally but degrade when tested at external centers with different patient demographics, imaging equipment, acquisition protocols, and annotation practices.
Current data resources: At the time of publication, there was no centralized, publicly available medical imaging data repository specific to lymphoma. The Cancer Imaging Archive (TCIA), one of the largest publicly accessible de-identified cancer imaging datasets, contained approximately 31 million cancer images at that time, including PET/CT studies from 155 DLBCL patients. The NIH DeepLesion dataset made available over 10,500 labeled CT imaging studies from 4,400 patients covering lung nodules, liver tumors, enlarged lymph nodes, and other critical findings - but this CT-only resource does not include the FDG-PET metabolic data needed for lymphoma quantification. Both resources, while valuable, fall well short of the scale and subtype diversity needed for generalizable lymphoma PET AI models.
Multi-centric ground truth and federated data collection: The authors specifically emphasize the need for datasets that include multi-centric GT data generated by experts with varying levels of experience, to recreate the heterogeneity that exists in real clinical practice rather than the controlled environment of a biomedical research study. This is an important nuance: datasets annotated exclusively by expert radiologists at academic centers may not reflect the annotation variability that community-based AI tools would encounter in deployment. Institutional collaboration and federated data curation frameworks are identified as necessary for assembling the large, diverse, expertly annotated datasets required.
AI across the imaging lifecycle: The review closes by situating lymphoma PET AI within a broader transformation affecting the entire PET imaging lifecycle. AI is being applied not only to detection, segmentation, and prognostication, but also to scheduling and operational workflow optimization, image acquisition parameter tuning, image reconstruction quality enhancement, image harmonization across scanners, and integration of PET imaging with multi-omics data for high-throughput biomarker discovery. The convergence of deep learning with PET-based radiophenomics and multi-omics integration represents the next major frontier for personalized lymphoma management.