Lymphoma is a biologically and clinically heterogeneous group of cancers that arises in the lymphatic system. It is broadly divided into Hodgkin lymphoma (HL) and non-Hodgkin lymphoma (NHL). NHL itself includes several major subtypes: diffuse large B-cell lymphoma (DLBCL), the most common aggressive form; follicular lymphoma (FL), the most common indolent form; mantle cell lymphoma; and marginal zone lymphoma. Although survival has improved with standardized treatments like R-CHOP and targeted therapies, the disease burden of NHL remains substantially higher than that of HL.
This 2024 review by Yuan, Zhang, and Wang (Shandong Provincial Hospital) provides a comprehensive survey of how machine learning (ML) and artificial intelligence (AI) are being applied across the full spectrum of lymphoma care. The authors cover ML applications in diagnosis (digital pathology, radiomics, genomics, and laboratory examination), prognosis (survival prediction, risk stratification), and treatment (chemotherapy response, drug discovery). They also describe the hierarchy of AI methods, noting that deep learning (DL) is a subset of ML, with architectures including convolutional neural networks (CNNs), recurrent neural networks, and deep belief networks.
The review highlights how the digitization of medical records, the growth of multiomics data (genomics, proteomics, transcriptomics), and the availability of whole slide images have created the large-scale datasets that ML algorithms need. The general supervised ML workflow involves training on labeled image datasets through data cleaning and feature extraction, then applying the resulting model to newly acquired medical images for classification and diagnosis.
The diagnostic challenge: Pathological examination of lymph node tissue is essential for lymphoma diagnosis, but it requires pathologists to survey images to determine the specific subtype, a process that demands significant time, experience, and expertise. The global shortage of pathologists, especially hematopathologists with access to molecular analysis equipment, compounds this problem. AI offers a way to build fast, accurate classifiers that help pathologists analyze and quantify pathological images.
CNN-based NHL classification: Steinbuss et al. (2021) applied convolutional neural networks to histopathological images for the classification of non-Hodgkin lymphoma subtypes. A separate study by Zhang et al. used a deep residual neural network to preprocess lymphoma pathological images with data augmentation techniques such as image flipping and color transformation, achieving a classification accuracy of 98.63%. This network model provides objective evidence for clinicians diagnosing the specific type of NHL.
SVM-based cell recognition: Alferez et al. trained an SVM (support vector machine) recognition system for automatic image-based analysis of lymphocyte morphology. When tested on images from 21 new patients and compared against confirmed diagnoses, the system achieved an overall recognition accuracy of 97.67% for classifying cells into normal lymphocytes, abnormal lymphocytes, and reactive lymphocytes. Tagami et al. (2023) also used an SVM with texture analysis for automatic classification of histopathological images of ocular adnexal mucosa-associated lymphoid tissue (MALT) lymphoma, working with 990 pathological images from 99 patients.
Follicular lymphoma pixel classification: A symbol-based ML method applied to FL image pixel classification demonstrated that the classification ability is successful at least at the image preprocessing stage, with accuracy as high as 95%. Whole slide image analysis using deep neural networks has also become an essential tool for lymphoma diagnosis, representing an interdisciplinary approach that merges pathology with computational science.
Radiomics and CNN for B-cell lymphoma: The combination of radiomics and ML has accelerated the analysis of lymphoma imaging data. CNNs serve as a promising tool for classifying large volumes of B-cell lymphoma information obtained through multiparameter flow cytometry (MFC). However, the authors note that significant work remains in optimizing CNN performance across images with different magnifications, image attributes, and network structures.
DLBCL genomic subtyping: Zhuang et al. (2023) used ML to establish a predictive model based on seven key genes enriched in immune infiltration and cell cycle pathways. They used an SVM trained on the integrated genome of 1,143 DLBCL samples to differentiate DLBCL subtypes based on mRNA features. Gene expression profiling divides DLBCL into germinal center B-cell-like (GCB), activated B-cell-like (ABC), and unclassified subtypes, each with different prognoses and treatment implications.
Eight-marker classification: Zhao et al. demonstrated that ML can detect eight specific molecular markers (including BCL6, MYBL1, LMO2, MME, NFKBIZ, IRF4, PDE4B, and SLA) to hierarchically manage DLBCL patients and guide personalized treatment choices. Another study used a Bayesian ML approach to analyze targeted RNA expression data from sequencing of 1,408 genes in DLBCL, helping determine clinical course. These genomic ML applications represent a shift from qualitative to quantitative pathological analysis.
Limitations of genomic ML: Despite great potential, the application of MFC-based ML remains limited. Whole-genome and whole-transcriptome sequencing methods have not been fully explored for lymphoma. The sheer number of genes to be detected and interpreted means that genomic prognostic classification has not yet been widely adopted in clinical practice, though DL exploration of these high-dimensional, multimodal datasets holds significant promise.
Morphogo for abnormal lymphocyte detection: Tang et al. developed Morphogo, an ML system based on digital imaging analysis algorithms, to distinguish abnormal lymphocytes from normal lymphocytes in peripheral blood and bone marrow. This tool combines peripheral blood analysis, bone marrow morphology, and ML to provide more accurate preliminary lymphoma diagnosis. Separately, Matek et al. demonstrated that training CNNs on large image datasets enables highly accurate differentiation of bone marrow cell morphologies.
Flow cytometry automation: Multiparameter flow cytometry (MFC) is a cornerstone of lymphoma clinical decision-making, capable of quickly quantifying cell surface markers. However, manual gating of cell populations is time-consuming and subjective. Mallesh et al. used transfer learning to classify seven B-cell tumor subtypes and nine-color panel healthy controls, extending DL models to MFC panels for high-precision multi-class classification. Gaidano et al. combined ML with flow cytometry to generate clinically applicable prediction systems for every biomaterial involved in lymphoproliferative disorders, from blood to lymph nodes and pleural effusion.
Random forest for chronic lymphoproliferative diseases: Gross et al. (2024) used a random forest classifier to distinguish between tumor and non-tumor B cells using flow cytometry data annotated with over 300 annotations. Their model proposed the correct classification of chronic lymphoproliferative diseases in over 90% of cases. This demonstrates that ML-assisted flow cytometry can reliably separate malignant from benign B-cell populations across diverse clinical specimens.
Pediatric lymphoma scoring: Zijtregtop et al. used an ML logistic regression model to create a diagnostic scoring system for pediatric lymphoma referral. The model facilitates prompt referral of suspected lymphoma patients to oncologists while reducing unnecessary referrals and invasive procedures for children with benign lymph node enlargement. Can et al. (2024) also developed a deep learning model for diagnosing cervical lymph node enlargement, trained on data from 400 patients who underwent surgery between 2010 and 2022.
Total metabolic tumor volume (TMTV): Many indicators can predict lymphoma prognosis, but TMTV, measured by whole-body FDG-PET/CT, has emerged as an independent prognostic factor showing strong value in both HL and NHL. Capobianco et al. demonstrated that a DL method can fully automatically estimate TMTV for DLBCL, producing results consistent with expert measurements. Their model showed significant prognostic value for both progression-free survival and overall survival. Blanc-Durand et al. used a CNN for fully automatic segmentation of DLBCL lesions on 3D FDG-PET/CT to predict total metabolic tumor volume.
Outperforming the International Prognostic Index: Biccler et al. integrated clinical data into a new ML-based prognostic model for DLBCL, validated through the Nordic Lymphoma Group registry. This model was shown to be remarkably better than existing DLBCL prognostic indexes, including the widely used International Prognostic Index (IPI). For mantle cell lymphoma, however, the IPI for mantle cell lymphoma and Ki-67 remain the strongest predictors, even though TMTV adds value for risk stratification.
EHR-based and gene expression models: ML applied to structured electronic health records (EHR) can forecast individual survival at the start of first-line treatment. For follicular lymphoma, the critical prognostic benchmark is progression within 24 months of first-line therapy, and ML provides more accurate EHR-based prediction tools. Carreras et al. used ML to analyze FL gene expression and identified overall survival genes related to patient prognosis and the immune microenvironment. Irshaid et al. showed that ML and deep learning criteria can predict the probability of large cell transformation in FL patients undergoing bone marrow biopsy.
Circulating tumor DNA monitoring: Zhao et al. (2024) used a decision tree model to forecast DLBCL progression by monitoring the rearrangement of immunoglobulin heavy chain (IgH) genes in circulating tumor DNA (ctDNA), working with 55 DLBCL sequencing cases. Sun et al. developed prognostic models for extranodal NK/T-cell lymphoma using Cox regression and ML on clinical data from 250 patients treated with non-anthracycline regimens. The probability-calibrated ensemble method was also applied to predict mortality in DLBCL patients, with 30-50% of DLBCL patients at risk of relapse after standard chemotherapy.
Chemotherapy response prediction: ML algorithms are increasingly used to improve treatment decisions for lymphoma patients. PET/CT imaging has high sensitivity for primary and metastatic lymphoma and can assess treatment effects and guide follow-up therapy. The probability-corrected ensemble method can predict DLBCL patient mortality, helping clinicians judge treatment efficacy and improve chemotherapy regimens. In HL, AI-driven radiomics enables more effective risk stratification, directly influencing treatment decision-making. The combination of next-generation sequencing (NGS) and AI is used to predict chemotherapy response.
Drug discovery and compound synthesis: ML algorithms have been widely adopted in computer-aided drug discovery. Traditional computing cannot reliably predict the impact of candidate drug compounds on their targets, leading to high failure rates in clinical trials. ML prediction methods can save substantial resources by identifying toxic compounds and potential side effects before drugs enter trials. In a recent study, researchers devised a pipeline integrating ML with molecular dynamics simulation to uncover potential dual inhibitors of B-cell lymphoma targeting BTK and JAK3 from natural product databases.
Combination therapy frameworks: Julkunen et al. proposed ComboFM, an ML framework for predicting drug combination response in preclinical research, including combined medications based on cell lines or patient-derived cells. DL neural networks can also clarify pharmacological features across different biological systems using transcriptome data, enabling drug repositioning. DL gives priority to new chemical targets in the early phase of drug discovery, and the authors note that the continuous efforts of ML may eventually influence the design and development of lymphoma drugs.
Personalized medicine trajectory: The growing body of evidence shows that patients benefit from the paradigm shift toward precision medicine enabled by genomics and AI. NGS technology and electronic tools are opening new avenues for evaluating drug efficacy, safety, pharmacokinetics, and pharmacodynamics. Surface-enhanced Raman spectroscopy (SERS) combined with ML has been shown to work as a noninvasive, label-free tool for monitoring hematological malignancies, establishing disease burden scores for standard treatment patients.
Emerging AI technologies in lymphoma: The review identifies several cutting-edge ML technologies transforming lymphoma clinical practice. Large language models (LLMs) offer sophisticated means to analyze patient imaging datasets, enabling physicians to rapidly discern lymphoma lesions and guide treatment strategies. Natural language processing (NLP) augments physicians' capacity to process complex patient medical histories and clinical records, facilitating nuanced evaluation of disease progression and treatment efficacy.
Explainable AI (XAI) for clinical trust: In the complexities of treatment selection, XAI technology serves as a critical ally, furnishing clinicians with decision support mechanisms and explaining the predictive outcomes of ML models. The authors emphasize that XAI is essential because the opacity of modern DL systems makes it difficult for clinicians with limited technical backgrounds to understand how models produce their recommendations. Without explainability, it becomes challenging to justify clinical decisions based on AI output.
Multi-technology integration: The authors envision bespoke treatment paradigms crafted through the full spectrum of ML advancements. By harnessing LLMs for imaging analysis, NLP for clinical record processing, and XAI for decision transparency, clinicians can achieve heightened precision and efficacy in treatment outcomes. The convergence of these technologies represents a shift from isolated AI tools to integrated clinical intelligence systems for lymphoma management.
Data quality and overfitting: The medical industry has remarkable "big data" characteristics, but data quality remains a critical barrier. Due to the lack of unified standards across medical institutions, collected data is often scattered, nonstandard, incomplete, or missing. For modest sample sizes, virtually any ML model is predisposed to overfitting, which can lead to artificially inflated accuracy and even clinical implications for decision-making. Feature selection and regularization have shown some effectiveness in preventing overfitting, but only a few studies externally validate their models.
Domain-specific challenges: The review identifies four major technical domains with distinct limitations. Immunophenotyping via multiparameter flow cytometry struggles with accurately representing original data. Cytomorphology via microscopy faces challenges in correctly identifying cell boundaries. Molecular genetics via genomic analysis contends with sparse data matrices where signals are easily lost. Radiology via 18F-FDG PET imaging must learn to distinguish recurrence from post-treatment inflammation, infection, or active fibrosis.
Data security and privacy: Healthcare data security is a persistent concern. Recent data breaches highlight the need to protect patient privacy. Khan et al. explored an information entropy l-diversity model that minimizes implementation time and reduces the chance of critical data leakage while improving data accuracy. The authors stress that data access agreements and cybersecurity considerations aligned with patient privacy laws are foundational to safe AI development in lymphoma research.
Future directions: The convergence of ML with advances in genomics, imaging, and therapeutic modalities holds great potential for improving precision medicine in lymphoma treatment. Key priorities include developing standardized data collection protocols, implementing interpretability frameworks for ML algorithm transparency, fostering interdisciplinary collaboration between clinicians and technologists, and conducting prospective multi-institutional validation studies. The authors call for a comprehensive strategy that addresses technological, ethical, and societal factors to fully realize ML's potential in transforming lymphoma management.