Introduction

Lymphoma is a malignancy of the hematopoietic system. Over the past two decades, there has been a significant global increase in lymphoma incidence. According to the Global Cancer Statistics approximately 20.0 million new cancer cases were diagnosed worldwide in 2022, including 0.08 million cases of HL and 0.5 million cases of NHL. Among the 9.73 million cancer-related deaths, lymphoma accounted for 0.27 million (2.8%), establishing it as a principal cause of cancer mortality.¹ Lymphoma biology and clinical manifestations are heterogeneous.² Lymphoma usually presents as painless lymph node enlargement, but it can also occur without obvious symptoms, leading to misdiagnosis. Although survival rates for lymphoma patients have improved in recent decades, there are still wide variations in five-year survival rates. The high heterogeneity and misdiagnosis rates of lymphoma highlight the urgency of developing non-invasive, effective diagnostic tools. Medical imaging research is essential for optimizing lymphoma diagnosis and treatment.

Lymphoma is a highly heterogeneous tumor.^3–5 Tumor heterogeneity reflects the characteristics of tumor growth and biology.^6,7 It serves as a crucial prognostic marker for assessing cancer progression, relapse potential, and treatment resistance.^8–10 Lymphoma is not limited to the lymphatic system, but extends widely to various parts of the body, including solid organs (e.g., liver and lungs), soft tissues, and bone marrow.^11,12 In clinical practice, biopsy or lesion resection is usually performed on a single lesion site, meaning that insufficient heterogeneity reflects the whole tumor and information can only be acquired from a scant few sites, perhaps even just one.¹³ Additionally, this approach is highly invasive, non-repeatable,^14–16 and may cause complications. Recently, several studies have clearly shown that radiomics shows substantial potential for non-invasive assessment of tumor heterogeneity.^17–20 This methodology employs sophisticated high-dimensional analysis of imaging data to effectively characterize tumor heterogeneity,^17,18 thereby offering robust reinforcement for the prognosis of clinical outcomes.²¹ Its advantage is that it can realize non-invasive, real-time, dynamic, and reproducible monitoring of tumor, which opens up a new way for tumor diagnosis and treatment.²² Artificial intelligence (AI) is developing rapidly. Studies increasingly prove that integrating AI with MRI,²³ pathology,²⁴ and genetic data²⁵ is important for lymphoma diagnosis and treatment. Particularly, PET/CT imaging—a crucial molecular imaging technology—offers important advantages when combined with AI. Based on existing lesion morphological features, PET/CT radiomics incorporates first-order, second-order, and higher-order features, alongside the characteristic metabolic features of PET, in order to more accurately depict tumor heterogeneity.^26–28 This paper aims to investigate the role of AI-assisted PET/CT radiomics in lymphoma management, including diagnosis, subtyping, staging, risk stratification, treatment selection, and prognosis prediction. Additionally, we discuss key methodological aspects such as radiomic feature extraction, model comparative analysis, and advanced deep learning techniques.

Radiomics and Artificial Intelligence

PET/CT radiomics can extract detailed tissue and lesion features from medical images.^29–31 These features can be used either independently or in conjunction with demographic, histological, genomics, or proteomic data to predict the underlying biological properties and tumors’ biological behavior, thereby elucidating the correlation between patients’ clinical characteristics and prognosis.^32–34 The basic workflow of PET/CT imaging radiomics comprises: acquiring and standardizing images, region labeling, segmenting the images, calculating and selecting features, reducing dimensionality, constructing the model, and finally, evaluating the model (Figure 1).

Figure 1 The picture depicts the workflow of radiomics. (1) Image acquisition and standardization, (2) Labeling and segmentation, (3) Feature extraction, (4) Feature selection, (5) Model construction, (6) Model evaluation.

Radiomics features (RF) are generally classified into statistical features, encompassing histogram-based feature and texture-based features; model-based feature; transform-based feature; and shape-based feature.³⁵ Histogram features (also called first-order features) represent the most basic statistical descriptors. Derived from individual pixel or voxel analysis without incorporating spatial information, these features mainly include mean, maximum, minimum, variance, and percentiles of the gray levels.³⁶ In PET imaging, the routinely used SUV measurements—SUV max, SUV mean, and SUV peak—all represent first-order features. Texture features, also known as second-order features, characterize the spatial organization and intensity hierarchy of voxels, serving as one of the conventional methods for quantifying tumor heterogeneity.³⁷ The most commonly utilized second-order features include GLCM, GLRLM, GLSZM, GLDZM, NGTDM, and the NGLDM.³⁸ Model-based features are engineered to decode spatial grayscale data, enabling the characterization of objects or shapes.³⁹ Transform-based features analyze gray-level patterns in different mathematical spaces through various transformations, including Fourier, Gabor, and Haar wavelet transformation.³⁵ Artificial intelligence (AI) has demonstrated rapid advancement in medical imaging applications, showing remarkable capabilities in Image segmentation;^40–42 Feature extraction;^31,43 Feature selection and optimization;^44–46 Predictive model development.^47–49 Machine learning (ML), a branch of artificial intelligence, emerges as a promising auxiliary tool.⁵⁰ It specializes in identifying the intricate relationships between high-dimensional input radiomics features and target variables, utilizing training examples to develop predictive models with the utmost accuracy.⁵¹ Four commonly used ML methods: reinforcement learning, supervised learning, unsupervised learning, semi-supervised learning, can be used to solve different tasks. In practical applications, common ML methods include support vector machine (SVM), random forest (RF), decision tree (DT) and others. Three common terms in the application of ML are training set, validation set, and test set. The training set is used for model training, the validation set for optimizing hyperparameters, and the test set for evaluating the model’s generalization ability. Model performance is often assessed using the receiver operating characteristic (ROC) curve and its area under the curve (AUC). Deep learning (DL), a pivotal subset of ML,⁵² forms the foundation of modern AI systems. It harnesses intricate structures or multiple layers of processing—composed of a series of nonlinear transformations—to discern data algorithms and delve into the inherent patterns and representational depths of sample data.⁵³ Convolutional neural network (CNN) are the most commonly used method in DL. Recent advances demonstrate that ML approaches, particularly DL techniques combined with 18F-FDG PET/CT radiomic features, provide valuable decision-support capabilities for managing patients. Figure 2 depicts the workflow of AI in lymphoma management.

Figure 2 The picture depicts the workflow of AI in lymphoma management.

Recent Innovations in Radiomics for Lymphoma

Diagnosis, and Differential Diagnosis

Achieving effective management and accurate prognosis of lymphoma depends fundamentally on timely and early diagnosis.^54,55 Over the past decades, PET/CT radiomics have improved the accuracy of lymphoma diagnosis. The impact of PET/CT radiomics on lymphoma diagnosis has dramatically expanded, evolving from the examination of isolated features to a comparative analysis of models and the employment of advanced deep learning methodologies. Kong et al⁵⁶ investigated the utilization of shape, first-order, and texture features derived from 18F-FDG-PET imaging to distinguish central nervous system lymphoma from glioblastoma (GBM). They pinpointed 13 radiomics features capable of effectively discerning lymphoma from GBM, suggesting that lymphoma generally displays higher SUV values across most interval segments and is quantitatively more heterogeneous compared to GBM. This study demonstrated that radiomics features can effectively distinguish GBM from central nervous system lymphoma (Table 1). A retrospective study evaluated the ability of 18F-FDG PET/CT radiomics combined with machine learning to distinguish breast cancer from breast lymphoma.⁵⁷ Six unique classification models were created, each containing a distinct combination of clinical and imaging data. The findings indicated that both the PETa model (which merges clinical details with SUV measurements and radiomic features) and CTa model (which combines clinical information and radiomic features derived from CT scans) exhibited exceptional proficiency in distinguishing between the training and validation groups. Notably, the PETa model achieved an AUC of 0.867 for the training set and 0.806 for the validation set, whereas the CTa model achieved an AUC of 0.891 for the training set and 0.759 for the validation set, showcasing their respective discriminative strengths (Table 1). In a large-scale retrospective case study,⁵⁸ 629 patients were enrolled to undergo 18F-FDG PET/CT uptake classification utilizing a deep convolutional neural network (CNN) to discriminate between lymphoma and lung cancer. In the evaluation of CNN performance for 18F-FDG PET/CT imaging, utilizing five input features, distinct results emerged. 18F-FDG PET alone achieved a remarkable AUC of 0.97, highlighting its superior performance. The incorporation of 18F-FDG PET and CT, further enhanced the diagnostic capability with an AUC of 0.98 (Table 1). Significant to highlight that PET/CT radiomics also plays a crucial role in: discerning lymphoma from metastatic tumors;⁵⁹ differentiating renal lymphoma from renal cell carcinoma;⁶⁰ and distinguishing pancreatic cancer from pancreatic lymphoma (Table 1).⁶¹

Table 1 Application of 18F-FDG PET/CT Radiomic Features in Lymphoma: Diagnosis, Subtyping, Staging, and Risk Stratification

Distinguishing Between the Lymphoma Subtypes

The significant variations in therapeutic interventions and predicted outcomes for individual lymphoma subtypes emphasize the importance of timely pathological subtype classification to guide accurate and tailored treatment plans. Notably, lymphoma subtype and patient heterogeneity are major drivers of patient outcome.⁶⁶ At present, lymphoma classification relies on invasive biopsy and pathological analysis. Consequently, the utilization of noninvasive techniques for the automated classification of lymphoma subtypes holds significant clinical implications. An increasing number of studies demonstrate that PET/CT radiomics plays a pivotal role in distinguishing among diverse lymphoma subtypes. Abenavoli EM et al conducted a retrospective review of histopathological diagnoses for mediastinal volume disease. The application of machine learning approaches was utilized to assess the diagnostic importance of 18F-FDG PET/CT volumetric and textural features for identifying distinct histological subtypes of lymphoma. The findings revealed substantial variations in SUV max, SUV mean, MTV, TLG, and various texture features at both first- and second-order gray levels across the lymphoma groups. Among machine learning classifiers, tree-based ensemble techniques demonstrated the highest efficacy for distinguishing lymphoma histological subtypes.⁶⁷ Researchers classified patients into four major lymphoma subtypes: DLBCL, FL, HL, and MCL. By integrating multiple-instance learning with support vector machines (SVMs) and random forest classifiers, they achieved 97.0% sensitivity and 94.1% positive predictive value for HL diagnosis, evaluated at both volume-of-interest (VOI) and patient levels.⁶⁸ Lovinfosse et al integrated clinical data, including age, sex, and weight, with radiomic features from both original images and tumor-liver radiomics (TLR). They employed seven distinct feature selection methods with four machine learning (ML) classifiers to differentiate HL and DLBCL. The TLR lesion-based approach achieved an AUC of 0.95, while the patient-based approach (incorporating original radiomics and age) yielded an AUC of 0.86 (Table 1).⁶² Regarding lymphoma subtype classification, PET/CT radiomics has undertaken multiple endeavors. These endeavors consist of the application of the Radiomics-Deep Learning (RA-DL) approach for tumor subtype classification;⁶⁹ as well as the implementation of classification methods that are derived from pseudo-spatiotemporal (PST) radiomics features, and the utilization of structural recurrent convolutional neural networks for distinguishing different subtypes.⁷⁰

One challenge in treating lymphoma is accurately identifying cases that have undergone histological transformation (HT). HT refers to the process whereby indolent lymphoma transforms into a clinically aggressive form. Despite declining conversion rates, the disease remains a substantial threat with poor prognosis.^71–73 Richter transformation (RT), initially delineated by pathologist Moritz Richter in 1928,⁷⁴ describes the conversion of chronic lymphocytic leukemia/small lymphocytic lymphoma (CLL/SLL) into aggressive B-cell lymphoma or HL.^75,76 This transformation occurs in roughly 2–10% of cases.⁷⁴ Several genetic, clinical, and biological factors are linked to elevated RT risk, yet a unified consensus remains elusive.^77–79 Prognosis is generally poor limited effective treatments.^80–82 Previous studies indicate 18F-FDG can identify transformation from indolent to aggressive lymphoma.^83–85 Among semi-quantitative studies, SUV max has been the most extensively researched.^86–88 A meta-analysis of 1,593 CLL patients established an optimal SUV max threshold of 5 for detecting RT, demonstrating 86.8% sensitivity and 90.5% negative predictive value.⁸⁹ In addition to SUV max, studies also incorporated other semi-quantitative parameters. Significant differences between RT and non-RT patients were observed in PET metrics including: maximum standardized uptake value body weight (SUV bw), lean body mass (SUV lbm), body surface area (SUV bsa), lesion-to-liver SUV ratio (L-L SUVR), and lesion-to-blood-pool SUV ratio (L-BP SUVR).^87,90 Large-cell transformation represents another common form of HT. The predictive value of SUV max for identifying large-cell transformation remains controversial. Previous smaller series suggested an association between baseline SUV max and HT,^91,92 FDG activity was overall higher in HT patients than in histological untransformed patients.^93,94 Nonetheless, some research endeavors have revealed that SUV max does not function as a reliable indicator for predicting HT.⁹⁵

Staging, Especially in the Bone Marrow Infiltration (BMI)

Accurate staging plays a crucial role in devising effective therapeutic strategies for both HL and NHL. According to the National Comprehensive Cancer Network Working Center (NCCN) and Society of Nuclear Medicine (SNM), FDG-PET/CT is indispensable for lymphoma staging.⁹⁶ PET/CT implementation has markedly enhanced identification of nodal and extranodal manifestations, allowing for either an upgrade or downgrade in staging.^97–99 Research has indicated that it alters the staging in 18% to 45% of FL patients, 3% to 45% of HL patients, and around approximately 5% of DLBCL diagnoses.^100–104 Bone marrow involvement (BMI) occurs in approximately 50% of patients with NHL and can be seen in up to 15% of HL.^105–107 While bone marrow biopsy (BMB) remains the gold standard for assessing marrow aggressiveness, its reliance on random sampling contributes to high false-negative rates.^108,109 Whether 18F-FDG PET/CT radiomics can replace BMB is a current research focus. The outcomes of numerous visual18F-FDG-PET investigations evaluating bone marrow infiltration proved unsatisfactory, exhibiting either low sensitivity or specificity, particularly with a sensitivity ranging from 12% to 52% in MCL cases.^110–113 Numerous studies have attempted to differentiate involved and uninvolved bone marrow using quantitative PET/CT metrics, yet the findings remain contentious.^89,114–116 Researchers sought to pinpoint threshold values capable of distinctly separating the SUV max, SUV mean, and SUV peak of affected and unaffected bone marrow. They arrived at the optimal cutoff points for cSUV mean (3D partial volume corrected mean standardized uptake value), SUV max, and SUV peak as 1.3, 2.1, and 1.7, correspondingly, resulting in unique pairings of sensitivity and specificity 75.0% and 85.7%, 87.5% and 85.7%, and 87.5% and 85.7%, separately.¹¹⁶ However, Adams HJ et al included 40 newly diagnosed DLBCL patients in their study and conducted a head-to-head comparison against BMB, revealing a substantial intersection in the cSUV mean, SUV max, and SUV peak values between the groups.¹¹⁷ Aide and colleagues extracted a range of features from PET scans of 82 DLBCL patients, including histogram, co-occurrence matrix, and size region matrix features. Among the histogram features, SkewnessH was identified as the most accurate measure for determining BMI, with a sensitivity of 81.8% and a specificity of 81.7%.¹¹⁸ Other studies have incorporated texture analysis and radiomics alongside standard PET parameters. The analysis identified two specific radiomic metrics—code similarity and long-run emphasis—as having significant predictive value for BMI identification.¹¹⁹ In a separate retrospective study, researchers extracted standard uptake value parameters (SUVmax, SUVmean, SUVpeak) and 16 co-occurring matrix texture features. Utilizing a multilayer perceptron neural network, compared three distinct combinations to predict BMI. SUVs, the radiomic features, and the radiomic features when integrated with laboratory data, showed AUCs reaching up to 0.66, 0.73, and 0.81 for the assessment of involved versus uninvolved bone marrow. These results indicate that FDG-PET texture features augment the predictive power of SUV-based methods for identifying BMI in MCL (Table 1).⁶³ Therefore, PET/CT-based radiomics plays a pivotal role in precise lymphoma staging and significantly contributes to clinical decision-making.

Treatment Options Based on Risk Stratification

PET/CT radiomics plays a crucial role in personalized lymphoma treatment management. Risk-adapted treatment strategies are essential for optimizing cure rates while minimizing treatment-related toxicity. PET/CT radiomics provides clinically significant quantitative data, facilitating both pretreatment risk stratification and treatment response monitoring through phenotypic characterization. The capacity to identify patients at elevated relapse risk before treatment enables early stratification and potentially improves outcomes.

For the most common type of NHL, DLBCL, current prognostic systems include the International Prognostic Index (IPI),¹²⁰ the revised IPI,¹²¹ or the National Comprehensive Cancer Network IPI,¹²² do not effectively discern patients at high risk.^123,124 Consequently, recent studies have shown growing interest in using PET/CT-derived biomarkers for initial risk stratification in lymphoma patients to optimize treatment outcomes. Researchers analyzed 152 DLBCL patients, extracting 1245 radiomics features from pretreatment PET/CT images, with primary focus on total metabolic tumor volume (TMTV) and metabolic tumor volume (MTV). These features, alongside clinical variables, were employed to develop a mixed nomogram. The AUC values for TMTV and MTV were significantly higher than those of the IPI. Decision curve analysis (DCA) demonstrated that the composite nomogram provided greater net benefit than the IPI alone. By integrating RF and IPI, the nomogram incorporating MTV and TMTV achieved more pronounced risk stratification compared to IPI alone, with more significant differences among subgroups and an enhanced risk ratio.¹²⁵ 95 individuals with advanced DLBCL were included in the study. Baseline MTV and radiomics features were calculated. Patients exhibiting elevated MTV and Dmax (maximum tumor dissemination) were identified as having risk factors for reduced progression-free survival (PFS). Furthermore, these factors correlated with reduced overall survival (OS). Combining MTV (> 384 cm³) and Dmax (> 58 cm) established three distinct risk categories. They suggest that integrating MTV with parameters reflecting tumor burden spread improves DLBCL staging risk stratification.¹²⁶ A model integrating metabolic heterogeneity (MH) with MTV can early identify patients with refractory DLBCL who could benefit from intensive therapy. High-risk patients with a diminished PFS (HR=5.6) and shortened OS (HR=7.6).¹²⁷ To identify DLBCL patients at high risk of disease progression or relapse, baseline lesions were delineated using semi-automated segmentation (SUV≥4.0) and 490 radiomics features were extracted.⁶⁴ The fusion of radiomics and clinical features, particularly the blend of tumor-centric indicators (MTV, SUV peak, and D maxbulk) with patient-centric parameters (WHO performance status and age exceeding 60 years), yielded the highest performance with an AUC of 0.79 (Table 1). The incorporation of radiomics features into clinical features resulted in a 15% elevation in PPV (positive predictive value), thereby increasing the exactness in pinpointing high-risk patients.

Most HL patients have a favorable prognosis.^128,129 However, early identification of the minority with refractory or recurrent disease, which can be life-threatening, would be clinically invaluable.^130,131 A retrospective analysis of 267 stage I–II HL patients found that in multivariate Cox regression, a 100-unit increase in total MTVt (HR = 1.14) and a 500-unit increase in total TLGt (HR = 1.096) significantly correlated with freedom from progression. This suggests MTV and TLG provide quantifiable tumor burden indicators for HL risk stratification.¹³² In order to precisely detect patients who are at a high risk for HL, the authors incorporated a cohort consisting of 258 patients diagnosed with stage I to II, who had undergone baseline PET scans and interim PET scans (iPET 2) following the completion of two treatment sessions (doxycycline, bleomycin, vinblastine, and dacarbazine).¹³³ Four high-risk categories were defined by TMTV (with a threshold of 147) and interim PET 2 (DS score), displaying a 5-year PFS of 95%, 81.6%, 50% and 25%, respectively. Compared to existing staging frameworks (EORTC,¹³⁴ GHSG,¹³⁵ NCCN)¹³⁶, TMTV improves initial risk assessment, enabling earlier treatment intensification and timelier therapeutic decisions.^137,138 For predicting HL radiotherapy need, 65 radiomics features and clinical parameters were analyzed using binary logistic regression to identify predictors and calculate odds ratios (OR).⁶⁵ The findings revealed that the first-order PET parameter sphericity (OR = 1.9), the CT parameter gray level zone length matrix high gray level zone emphasis (GLZLM SZHGE mean, OR = 2), PARAMS spatial resampling (OR = 2.1), and abnormal hemoglobin levels served as predictors for the final model, which achieved an AUC of 0.79 (Table 1).

Predicting Prognosis

Studies have identified multiple prognostic factors associated with lymphoma outcomes. IPI is regarded as a prognostic framework that is grounded in the characteristics of various clinical factors, encompassing age, Ann Arbor stage, extranodal involvement, serum lactate dehydrogenase (LDH) levels, and performance status.¹²⁰ However, these conventional parameters primarily function as indirect proxies for tumor burden and fail to fully capture functional and metabolic tumor characteristics.^139,140 Recent advances in PET/CT imaging have revealed various quantitative biomarkers with potential prognostic value. These indicators encompass a spectrum from semi-quantitative metrics to tumor volume, and extend to metabolic characteristics (eg, shape and texture).

SUV Indicators Used for Predicting Outcomes

Owing to its stability and reproducibility, SUV max is the most widely used parameter in lymphoma prognostic studies. Multiple studies have demonstrated correlations between SUV max and survival outcomes in lymphoma patients.^141–145 Researchers have proposed SUV max as a key predictor of disease progression in DLBCL. Using a cutoff value of 15, they found that patients with SUV max <15 had a 3-year overall survival (OS) of 90%, while those with SUV max ≥15 showed 72% survival. PFS was 90% for the low SUV max group and 39% for the high SUV max group.¹⁴¹ A retrospective analysis included 123 elderly patients with HL who underwent baseline18F-FDG-PET/CT scans at the onset of treatment and subsequent PET/CT scans at the conclusion of therapy. Through semi-quantitative index analysis, they explored whether baseline PET/CT metabolic parameters could be used as prognostic indicators for HL. The results showed that both L-BP SUVR (lesion to blood-pool SUV max ratio) and L-L SUV R (lesion to liver SUV max ratio) were significantly associated with PFS and OS.¹⁴³ In patients diagnosed with extranodal natural killer/T-cell lymphoma (NKTCL), a baseline SUV max of 9.65 emerged as an independent predictor of poor prognosis for distant relapse-free survival (HR=4.58) and PFS (HR=2.6). The authors suggest that the baseline SUV max could be a useful tool for identifying NKTCL patients at heightened risk of disease progression.¹⁴⁶ Nevertheless, some studies indicate that the prognostic predictive ability of SUV max in lymphoma is limited. One large-scale study indicated that SUV max has negligible predictive value for PFS and OS in 258 patients with stage I and II HL.¹³³

SUV max represents the most widely studied PET/CT biomarker in lymphoma prognosis, though its prognostic reliability remains debated. This ambiguity may stem from various factors influencing the SUV, including the spatial resolution of the scanner, the image acquisition protocols, and the PET reconstruction parameters. Furthermore, clinical study limitations such as heterogeneous cohort sizes across disease stages, along with differences in therapeutic regimens and progression-free survival (PFS) criteria, collectively contribute to the persistent uncertainty regarding its prognostic utility.^147,148

Tumor Burden Metrics for Prognostic Prediction

While SUV max primarily characterizes the most metabolically active tumor regions, it does not fully capture spatial heterogeneity within the lesion. Moreover, it overlooks critical prognostic elements, particularly tumor burden information. Recently, innovative metrics reflecting aggregate tumor burden, such as MTV and TLG, have been employed to predict PFS and OS in lymphoma patients. These indicators have been identified as potential baseline predictors for prognosis in different types of lymphoma.^149–151

The determination of the SUV threshold plays a pivotal role in delineating the boundary of lymphoma lesions, which in turn influences the calculation of the MTV. Commonly used thresholds include 41% of SUV max,^152,153 SUV 4,^154,155 and SUV max set at 2.5,¹⁵⁶ among others. As depicted in Figure 3, the contour range varies for an example DLBCL patient at different thresholds. To a certain degree, the diverse cut-offs utilized could have played a role in the survival differences observed between the groups. Determining the suitable SUV threshold for MTV computation is dependent on the specific patient cohort under evaluation.

Figure 3 FDG PET/CT images in a DLBCL patient show three distinct contouring thresholds (red arrows): Green= 41% SUV max; Yellow= 2.5 SUV; Blue= 1.5× liver SUV mean (3D Slicer [https://www.slicer.org/]). (A(a–d)) Left supraclavicular lymph node; (B(a–d)) Mediastinal lymph node; (Ca–d) Retroperitoneal lymph node ((A) axial CT (B) axial PET; (C) coronal PET; (D) sagittal PET). Abbreviation: MIP, Maximum intensity projection.

Researchers conducted a retrospective analysis of 169 patients with stage II–III DLBCL. The boundary of the target lesion was delineated employing a SUV threshold of 2.5, and MTV was subsequently quantified. ROC analysis identified 220 cm³ as the optimal MTV cutoff value. Patients with a lower MTV (<220 cm³) demonstrated significantly prolonged PFS and OS in comparison to those with a higher MTV (≥220 cm³)¹⁵⁷ (Table 2). In another study, margin values were determined using lesion SUV max thresholds of 25%, 50%, and 75%, which correspondingly yielded TLG values of TLG25, TLG50, and TLG75. High TLG50 values (> 415.5) were linked to decreased survival rates in comparison to low TLG 50 values (≤415.5) (73% versus 92% for 2-year PFS; 2-year OS 81% versus 93%). A high IPI score significantly diminished the OS (2-year OS: 79% vs 90%). Ann Arbor Phase III/IV had an adverse effect on the PFS (P=0.013). However, neither higher IPI scores nor stage III/V alone significantly impacted PFS. High TLG50 values independently predicted survival outcomes, with hazard ratios (HR) of 4.4 for PFS and 3.1 for OS¹⁵⁸ (Table 2). To assess baseline metabolic tumor volume (TMTV 0) in HL patients, TMTV 0 was quantified was quantified using a semi-automated technique with an SUVmax threshold of 41%. Both TMTV0 (225mL) and tumor volume (10 cm) were found to be indicators of PFS at the 4-year mark, with survival rates of 42% versus 85% and 44% versus 79%, respectively. Multivariate analysis confirmed TMTV0 (relative risk [RR] = 4.4) as an independent PFS predictor, demonstrating stronger prognostic value than tumor volume in HL patients¹⁵⁹ (Table 2). Investigator findings from Phase III of the GOYA study indicated that elevated TMTV (366cm³) and TLG (3004g) were associated with inferior PFS in patients with DLBCL¹⁶⁰ (Table 2). This study is based on the III GOYA study, and the prospective, standardized PET procedure and a large number of PET-CT scan samples at baseline enhance the confidence in the data conclusions and further confirm the prognostic value of baseline quantitative PET indicators in lymphoma. To date, no consensus exists on standard techniques for contouring MTV and TLG, despite numerous proposed limitations to define metabolically active tumors.^158,161,162 In PET/CT scans of 301 DLBCL patients, three-dimensional (3D) regions of interest (ROIs) were identified using an automated whole-body high-uptake segmentation algorithm and processed with a convolutional neural network (CNN) to calculate TMTVPARS. These results were compared with TMTVREF (measured by two experienced experts utilizing independent semi-automated software). The analysis showed strong agreement between TMTVPARS and TMTVREF, with both demonstrating significant predictive value for PFS and OS in DLBCL patients¹⁶³ (Table 2). In relapsed/refractory (R/R) cHL, high MTV at relapse and prior to AutoSCT (Autologous Stem Cell Transplantation) predicted more severe PFS.^164,165 Researchers are developing CD30-targeted chimeric antigen receptor T (CAR-T) cells for R/R HL treatment. Preliminary results from patients treated with CAR T cells suggest that high MTV (60mL) before lymphocyte depletion is associated with PFS at one year. The 1-year PFS rates were 14% versus 58% for high and low MTV groups, respectively¹⁶⁶ (Table 2). Similar to DLBCL and HL, MTV independently predicts outcomes in FL patients^167,168 (Table 2). In 185 FL patients receiving chemoimmunotherapy, the optimal MTV cutoff was 510 cm³. Patients with elevated MTV had significantly reduced survival rates (5-year PFS: 33%, HR=2.90; 5-year OS: 85%, HR=3.45).¹⁶⁷ A meta-analysis of 27 studies demonstrated MTV’s prognostic value, with a pooled HR of 3.05 for PFS.¹⁶⁹ In summary, MTV and TLG cut-off values were customized for particular patient groups, and the SUV thresholds applied require careful consideration when contrasting findings from distinct studies.

Table 2 Summary of Included Prognostic Studies by Lymphoma Type

Parameters That Quantify Tumor Spread Predict Prognosis

Although metabolic tumor volume (MTV) provides volumetric assessment, it fails to account for the spatial complexity of multifocal lymphoma manifestations. Typically, lymphoma involves numerous scattered nodal areas that can be connected to sites outside the nodes, often resulting in mutation-related heterogeneity that influences clinical outcome.¹²⁶ Maximum tumor dissemination (D max) serves as a novel radiological indicator, representing the maximum distance between the two farthest lesions.¹⁷⁸ Some studies have employed standardized D max (SD max), which adjusts the peak dose according to body surface area.^170,179 This geometric parameter offers unique advantages: (1) Simplified computation requiring minimal contour dependency; (2) Enhanced reproducibility across imaging platforms and operators; (3) Strong prognostic correlation in lymphoid malignancies. Clinical validation studies confirm Dmax’s independent predictive value for PFS and OS in both HL and DLBCL^170,175 (Table 2). In HL patients treated with ABVD, Dmax correlated with 3-year PFS. The most precise threshold for forecasting treatment outcomes was determined to be 16.2cm, with an AUC of 0.62¹⁷⁵ (Table 2). Another retrospective study found Dmax significantly predicted PFS (HR=4.3) and OS (HR=3.7) in advanced HL¹²⁶ (Table 2).

Parameters of Metabolic Heterogeneity in Predicting Prognosis

Metabolic heterogeneity (MH) serves as a marker for the diverse patterns of 18F-FDG uptake seen in tumors, distinct from the traditional evaluation of tumor volume and structure. MH encompasses a complex set of factors—including cellular metabolism, growth kinetics, vascular dynamics, and oxygen deprivation—that collectively capture the clinical and molecular complexities of the disease. MH quantification in the dominant lesion (largest MTV) employed the area under the cumulative SUV—volume histogram (AUC-CSH) curve¹⁸⁰—a method well-documented in solid tumors and lymphoid malignancies.^181–183 A prospective study of 103 primary mediastinal B-cell lymphoma (PMBCL) patients found no significant correlation between MH values and other baseline quantitative PET parameters (SUV max, MTV, and TLG), regardless of methodology. In the stepwise Cox model, elevated MH (HR=12.8) independently predicted substantially reduced PFS¹⁷¹ (Table 2). Similarly, among DLBCL patients receiving standard immunotherapy (R-CHOP regimen), elevated MH predicted a significantly worse prognosis in the patient group with higher MTV¹²⁷ (Table 2).

Structural and Texture Analysis Predicts the Prognosis

Radiomics enables high-throughput extraction of quantitative imaging features through computational analysis of subvisual patterns in medical imaging data.¹⁸⁴ In DLBCL prognosis research, the study demonstrated that integrated assessment of MTV with second-order texture parameters (homogeneity, contrast, correlation, dissimilarity) and higher-order features (LZE, LZLGE, LZHGE, GLNU, ZP) significantly predicts 2-year event-free survival (2y-EFS). Multivariate analysis identified LZHGE as an independent prognostic marker (HR=7.47)¹⁷² (Table 2). Researchers conducted a retrospective analysis of PET/CT-derived radiomic features and clinical parameters to predict 2y-EFS in DLBCL patients. The analysis identified five key predictors: Maximum diameter, NGTDM busyness, TLG, TMTV, and NGTDM coarseness¹⁷³ (Table 2). Another study analyzed baseline PET/CT scans from 229 DLBCL patients treated with R-CHOP¹⁷⁴ (Table 2). Logistic regression models incorporating MTV and six machine learning classifiers were trained and optimized using 4-fold cross-validation. The model that achieved the greatest mean validation AUC combined clinical and radiomic features using a ridge regression approach. It achieved a mean training AUC of 0.77±0.02, and a test AUC of 0.73, outperforming the model that utilized MTV features alone, which yielded an AUC of 0.67. These results demonstrate the clinical utility of radiomic-based predictive modeling for DLBCL outcomes.

Similarly, multiple studies have investigated HL. Researchers evaluated PET/CT radiomic parameters in 251 stage I–II HL patients treated at a tertiary cancer center. The model was crafted and validated through a machine learning algorithm, which resulted in an impressive AUC of 95.2% for the five most predictive features. These features include the first-order feature SUV max and volume, as well as the second-order features: information measure Corr 1, information measure Corr 2, and the mean variance derived from GLCM 2.5. These findings suggest PET/CT radiomic features may predict refractory disease in early-stage HL¹⁷⁷ (Table 2). Another study analyzed 289 cHL patients. The PET/CT data were segmented employing 1.5 × mean liver SUV and a fixed SUV threshold of 4.0, while radiomics features were derived employing PyRadiomics with ComBat harmonization. The ridge regression model (1.5 × mean liver SUV segmentation) showed optimal performance, with validation AUC of 0.79 ± 0.01. The model incorporated age alongside four radiomic features: PET flatness, the length of the PET major axis, GLSZM, GLCM, and the PET lbp-3D-m2. This study confirms that machine learning models based on pre-treatment FDG PET/CT scans can predict outcomes in cHL patients¹⁷⁶ (Table 2).

In conclusion, SUV max represents the most metabolically active region of the tumor and is characterized by high stability and excellent reproducibility. Although widely used in lymphoma prognostic studies, SUV max has limited predictive value for clinical outcomes. MTV and TLG provide better prognostic assessment by quantifying overall tumor burden and heterogeneity. However, clinical application faces a key challenge: standardization of SUV threshold definitions. Dmax—the interlesional distance between two farthest lesions—shows high reproducibility and promising prognostic value in lymphoma. Unfortunately, in actual clinical practice, Dmax tends to be neglected. Metabolic heterogeneity (MH) reflects complex tumor biology (cellular metabolism, proliferation kinetics, vascularity, and hypoxia) and strongly correlates with lymphoma aggressiveness. Furthermore, the integration of structural and texture analysis with AI has significantly enhanced PET/CT’s predictive capability in lymphoma prognosis (Table 3).

Table 3 Application of 18F-FDG PET/CT Radiomic Features in Predicting Lymphoma Prognosis

The application of AI in PET/CT clinical trials for lymphoma has significantly enhanced diagnostic accuracy and treatment efficacy assessment efficiency. Deep learning-based automatic lesion segmentation and quantitative analysis have reduced human bias while improving result reproducibility. Research has shown that AI models, when combined with radiomics features, can predict treatment responses and prognoses more accurately and earlier. However, challenges remain regarding multicenter data standardization, algorithm interpretability, and clinical translation pathways. Future research should focus on large-scale prospective validation and developing AI-driven clinical trial designs to advance precision medicine in lymphoma management.

PET/CT Parameters and CAR-T Therapy

In lymphoma treatment, immunotherapy and chemoimmunotherapy have emerged as increasingly prominent therapeutic approaches, complementing conventional chemotherapy. CAR-T cell therapy, also known as chimeric antigen receptor T cell therapy, is a new approach in the field of immunotherapy for NHL. Studies have shown that 18F-FDG PET/CT radiomics parameters play a key role in monitoring treatment response, evaluating the efficacy of immunotherapy, assessing treatment-related toxicity, and predicting prognosis in CAR-T therapy. Baseline SUV max can predict PFS following CAR-T treatment, and the baseline volume parameters (MTV and TLG) are potential prognostic indicators of CAR-T treatment affecting PFS and OS.¹⁸⁵ The two most prevalent adverse effects of CAR-T cell therapy are cytokine release syndrome (CRS) and neurotoxicity. CRS is characterized by a systemic inflammatory response caused by the rapid release of cytokines from activated CAR-T cells, while neurotoxicity manifests as neurological symptoms such as seizures and disorientation. Multiple PET/CT parameters (SUV max, SUV avg, MTV, TLG) may help predict and assess the severity of these toxicities in CAR-T recipients.¹⁸⁶ Thus, 18F-FDG PET/CT parameters offer a comprehensive strategy for managing CAR-T therapy patients, facilitating timely clinical interventions and serving as crucial prognostic tools for lymphoma patients undergoing this treatment.

Limitations and Challenges

Despite the rapid development of PET/CT radiomics in lymphoma, several challenges remain. The primary challenge involves image quality control, as high-quality imaging data constitute a prerequisite for radiomics analyses. One study examined data heterogeneity arising from variations in acquisition parameters, such as scan duration, iteration and subset numbers, reconstruction type and algorithm, and spatial resolution, as well as image processing methodologies, including segmentation techniques and gray level discretization.¹⁷⁶ Among these factors, spatial resolution variations demonstrated the most pronounced impact, with a coefficient of variation reaching 3.63. A series of guidelines and standards have been progressively implemented including the Radiomics Quality Score (RQS), the Image Biomarker Standardization Initiative (IBSI), the AI in Medical Imaging Checklist (CLAIM), and the Quantitative Imaging Biomarkers Alliance (QIBA), among others.^36,187,188 These aim to establish unified imaging protocols and analytical standardization. Second, most current lymphoma radiomics studies adopt retrospective designs. Given lymphoma’s heterogeneity and subtype transformation potential, obtaining pathological information from a single or limited sites poses challenges in ensuring that each lesion under investigation has corresponding pathological results. This limitation may introduce selection bias in region-of-interest (ROI) delineation and case inclusion. Consequently, large-scale prospective multicenter studies featuring independent external validation through multi-institutional collaborations are required to enhance clinical data granularity, optimize model robustness, and ensure broad applicability.⁵¹

Artificial intelligence (AI) has demonstrated versatile applications in medical imaging, including refined attenuation correction, artifact-free image reconstruction, and automated annotation enhancements. As a core component of AI, deep learning has revolutionized medical imaging workflows, powering key tasks such as automated segmentation, multimodal registration, image fusion, feature engineering, feature extraction, computer-aided diagnosis, and prognostic analysis. These advancements significantly improve the precision and dependability of imaging radiomics classification and predictive analytics. Thus, the incorporation of deep learning into the multifaceted stages of imaging radiomics to offset its inadequacies is expected to shape the forthcoming developmental path for imaging radiomics technology.

In addition, clinicians increasingly aware that single lymphoma biopsy samples inadequately reflect patient genomic profiles. Radiomics can be combined with genomic data (circulating tumor DNA, MYC rearrangement rearrangements) to provide more accurate biological information for lymphoma patients.¹⁸⁹ This aids in monitoring the progression of lymphoma at the molecular level. Research has revealed a correlation between lymphoma imaging phenotypes and gene expression patterns, which is particularly appealing in the management of lymphoma patients.Thus, the authors hold the belief that the integration of PET/CT imaging, radiomics, and genomic data presents significant potential for exploration and application in the clinical management of lymphoma patients.

While our manuscript demonstrates the significant application value of FDG PET/CT imaging radiomics in the comprehensive management of lymphoma, we acknowledge several limitations. These include, but are not limited to, the lack of: lymphoma subtype-specific classification, qualitative properties analysis, histopathological heterogeneity evaluation, and comprehensive overview of treatment and baseline characteristics as cited in the literature.

Conclusions

By correlating imaging features with the biological, pathological, and metabolic profiles of lymphoma, AI-assisted PET/CT radiomics enhances the clinical differential diagnosis of lymphoma, aids in the early risk stratification and prognostic prediction, and offers a foundation for the formulation of personalized lymphoma treatment strategies. This technology establishes a transformative link between medical imaging and precision oncology, driving the evolution of lymphoma management from macroscopic evaluations to microscopic characterizations, and from empirical approaches to biomarker-driven paradigms. We firmly believe that AI-assisted PET/CT radiomics will make accurate diagnosis and treatment of lymphoma possible in the future.