Lymphoma, a cancer that arises in the immune system’s own cells, now accounts for half or more of all malignant blood disorders worldwide, and catching it early can mean the difference between a curable disease and a fatal one. A comprehensive new review published in Neural Computing and Applications maps out how artificial intelligence is transforming every stage of lymphoma diagnosis, from the first suspicious scan to the final subtype classification under the microscope. The review, led by Nesreen M. Ezz-Eldeen of Delta University for Science and Technology together with colleagues at Mansoura University, sifts through years of research on machine learning and deep learning applied to lymphoma detection, organizing a sprawling field into a coherent picture of what works, what fails, and where the next breakthroughs are likely to come from.
The stakes are considerable. Each year, close to half a million people develop cancer globally, and lymphoma’s dozens of subtypes, split broadly into Hodgkin lymphoma and non-Hodgkin lymphoma, demand precise diagnosis because treatment strategies differ radically between them. Determining the stage of the disease depends primarily on segmenting the glandular structure of affected tissue, a task that is currently performed manually by pathologists and radiologists. Manual segmentation is not only slow and labor-intensive; it also carries risks for patients and introduces variability between observers. The review argues that automating this critical pre-treatment step with computational tools could standardize diagnosis, reduce human error, and ultimately improve survival outcomes by detecting cancer in its earliest stages.
One of the review’s central contributions is its systematic tour of the imaging modalities that feed AI systems. Positron emission tomography, particularly fluorodeoxyglucose PET combined with computed tomography, has become a workhorse for staging lymphoma and assessing treatment response, because malignant cells avidly absorb the radioactive glucose tracer. Magnetic resonance imaging, including diffusion-weighted sequences that measure how freely water molecules move through tissue, offers a radiation-free alternative that is especially valuable for children and adolescents who need repeated scans. The review also highlights hybrid PET/MRI systems, emerging tracers such as gallium-68 pentixafor that target chemokine receptors on lymphoma cells, and whole-body MRI protocols that can stage patients without ionizing radiation. Each modality presents distinct computational challenges, from the low resolution of PET to the complex signal physics of MRI.
On the analysis side, the review draws a sharp line between conventional machine learning and modern deep learning. The older paradigm relies on handcrafted features: engineers design mathematical descriptors of texture, shape, color, and statistical properties, such as Haralick texture measures, sample entropy signatures, Fisher vectors, and scale-invariant feature transform descriptors, and then feed them into classifiers like support vector machines. These approaches dominated early lymphoma studies and remain useful when training data is scarce, but they depend heavily on expert intuition about which features matter. Studies applying such methods to histopathology images of non-Hodgkin lymphoma achieved respectable classification results using percolation theory features, color histograms, and morphological descriptors, yet the pipeline required painstaking manual tuning at every step.
Deep learning has largely inverted that workflow. Convolutional neural networks learn their own features directly from pixels, and architectures such as VGG16, residual networks, and U-shaped segmentation networks have swept through the lymphoma literature. The review documents striking examples: deep learning models that accurately diagnose lymphoma on whole-slide histopathology images, platforms that distinguish Burkitt lymphoma from diffuse large B-cell lymphoma with high accuracy across multiple hospitals, and networks that segment lymphoma lesions in PET images using multi-view three-dimensional fusion strategies. Hybrid systems that fuse deep features with handcrafted ones have pushed performance further, and ensemble transfer learning frameworks have advanced malignant lymphoma diagnosis by combining multiple pretrained networks. More recently, vision transformers, which process images through attention mechanisms rather than convolutions, have entered the field, with hybrid convolutional-transformer networks and transformer-based knowledge transfer frameworks showing promise for lymphoma subtyping.
The review also emphasizes the data infrastructure that makes these models possible. Public benchmark datasets, annotated whole-body PET/CT collections with manually labeled tumor lesions, and grand challenges for whole-slide image analysis have lowered the barrier to entry for researchers. Performance evaluation relies on a standard battery of metrics, including the Dice similarity coefficient, Jaccard index, Hausdorff distance, and receiver operating characteristic analysis, which quantify how closely automated segmentations and classifications match expert ground truth. Yet the authors are candid about a persistent weakness: many studies train and test on small, single-center datasets, making it unclear whether reported accuracies will survive contact with the messy diversity of real clinical populations. Label noise, the mislabeling of training examples by tired or inconsistent annotators, compounds the problem and has become an active research area in its own right.
Several forward-looking techniques receive particular attention. Self-supervised learning allows models to pretrain on vast quantities of unlabeled medical images, learning general visual representations before fine-tuning on small labeled lymphoma datasets, an approach already proving effective in related cancer imaging tasks. Federated learning offers a privacy-preserving alternative to centralizing patient data: hospitals train shared models locally and exchange only model updates, keeping sensitive records behind institutional firewalls in compliance with regulations such as HIPAA and the GDPR. Generative adversarial networks, including cycle-consistent architectures, can normalize the staining variations that plague histopathology and even synthesize realistic training images. Foundation models, enormous networks pretrained on gigapixel pathology slides, are beginning to enable generalizable cancer diagnosis and survival prediction, and the review points to their aggregation for predicting treatment response in diffuse large B-cell lymphoma as a sign of things to come.
Interpretability emerges as another frontier. Clinicians are understandably reluctant to trust a black-box verdict on a cancer diagnosis, and the review highlights explanation techniques such as SHAP and LIME, which reveal which image regions and features drove a model’s decision. Radiomics, the extraction of hundreds of quantitative features from medical images, has produced machine learning models that can differentiate glioblastoma from primary central nervous system lymphoma on MRI, predict primary treatment failure in diffuse large B-cell lymphoma from CT scans, and distinguish follicular lymphoma from diffuse large B-cell lymphoma using PET features. These models do more than classify; they quantify tumor burden, automate total metabolic tumor volume calculations, and forecast relapse in mantle cell lymphoma, extending AI’s role from diagnosis into prognosis and treatment planning.
The review does not shy away from the field’s unresolved problems. Computational cost remains high, with deep models demanding graphics processing units that many hospitals lack. Domain shift, where models trained on one scanner or staining protocol underperform on another, threatens clinical deployment. Ethical questions around algorithmic bias, data governance, and the double opacity of federated systems remain open. The authors call for larger, multi-center, publicly available datasets, rigorous external validation, standardized reporting, and closer collaboration between computer scientists and clinicians. They also note that early detection through screening, including the removal of pre-malignant lesions, remains a cornerstone of prevention that AI can support but not replace.
What emerges from this sweeping survey is a field in rapid maturation. In barely a decade, lymphoma diagnosis has moved from hand-tuned feature extractors to self-supervised vision transformers that can digest whole-slide images and whole-body scans alike. The trajectory suggests a future in which computer-aided diagnosis systems flag suspicious lesions on a radiologist’s workstation within seconds, pathologists receive AI-generated second opinions on every slide, and prognosis is computed from quantitative image signatures before treatment even begins. If the challenges of data sharing, validation, and interpretability can be solved, the review concludes, artificial intelligence could substantially reduce lymphoma mortality by catching the disease earlier and tailoring treatment more precisely, turning one of medicine’s most labor-intensive diagnostic puzzles into a computable problem.
Subject of Research: Artificial intelligence techniques for lymphoma cancer diagnosis across imaging modalities and histopathology
Article Title: Lymphoma cancer diagnosis based on artificial intelligence techniques: a comprehensive literature review of different modalities and analysis techniques
Article References: Ezz-Eldeen, N. M., Sakr, N. A., Elmogy, M., & Ali, H. A. (2026). Lymphoma cancer diagnosis based on artificial intelligence techniques: a comprehensive literature review of different modalities and analysis techniques. Neural Computing and Applications, 38(17), Article 721. https://doi.org/10.1007/s00521-026-12417-0
Image Credits: AI Generated
DOI: 10.1007/s00521-026-12417-0
Keywords: lymphoma, artificial intelligence, deep learning, machine learning, medical imaging, PET, MRI, histopathology, computer-aided diagnosis, vision transformers, segmentation, digital pathology
News Source: Nathaniel Bowman. (October 5, 2026). AI Takes on Lymphoma: How Machine Learning Is Rewriting Blood Cancer Diagnosis. Scienmag.
