Clinical applications of artificial intelligence in the histopathology of lymphoma: diagnosis, treatment and prognosis

Mengyao Kang, Zibo Yang, Yu Tian, Dongyu Li, Zhiqiong Wang, Liting ChenDiscover 2025 · DOI

Abstract

Artificial intelligence (AI) is an important branch of computer science. With the rapid development of AI, the application in oncology has become increasingly widespread. As a hematologic malignancy characterized by remarkable heterogeneity, lymphoma has long posed significant challenges in both diagnosis and treatment, particularly with regard to its complex classification and difficulties in prognostic evaluation. Breakthroughs in AI technology have provided a new paradigm for precision treatment of tumors. AI can integrate and analyze HE pathology slides and PET/CT images to enhance diagnostic efficiency; meanwhile, in terms of treatment prognosis, AI can identify biomarkers to accurately classify lymphoma subtypes for therapeutic guidance, simultaneously quantify biomarkers to minimize the influence of subjective variability, and predict the patients' prognosis based on the extracted features to assist in the precise treatment of lymphoma. This review aims to provide an overview of AI related to various fields of lymphoma, introducing the core technologies and principles of AI, including deep learning, decision trees, regression models, and so on. We also discuss the clinical applications of AI in lymphoma pathology slides and PET/CT images, systematically analyze the clinical applications of AI in diagnosis, treatment, and prognosis, as well as innovatively summarize the cutting-edge applications of AI in 3D pathology of lymphomas, and finally, we emphasize the development potentials and current challenges of AI in the field of lymphomas to promote precision lymphoma diagnosis and treatment by providing a theoretical foundation.

Introduction: The Challenge of Lymphoma Heterogeneity

Lymphoma is a highly heterogeneous blood cancer with complex sub-classifications that dictate vastly different treatment paths. Currently, diagnosing and staging lymphoma relies on a multi-dimensional, manual integration of tissue biopsies, genetic tests, and imaging. This traditional workflow is often plagued by subjective variability in human interpretation, delays in molecular testing, and limitations in overall accuracy. Accurately subtyping lymphomas and predicting which patients will resist standard chemotherapy remains incredibly difficult. This paper explores how AI can address these gaps by automating feature extraction, standardizing diagnoses, and uncovering hidden prognostic patterns that human clinicians might miss.

Methodology: The AI Toolkit in Lymphoma Research

To address the complexities of lymphoma, researchers are deploying a variety of AI techniques tailored to different data modalities. The core methodologies driving the field include:

  • Deep Learning (DL) & Convolutional Neural Networks (CNNs): These are the primary tools for image-based tasks, used to analyze digitized Whole Slide Images (WSIs) and PET/CT scans for tumor classification and cell segmentation.
  • Decision Trees & Random Forests: These classical machine learning algorithms classify lymphoma subtypes by integrating clinical features and high-dimensional genetic data (e.g., RNA-seq), offering high interpretability.
  • Regression Models (e.g., LASSO): Employed for prognostic prediction and risk factor analysis, these models select the most relevant clinicopathological variables to evaluate treatment efficacy.
  • Emerging 3D Pathology: A cutting-edge approach combining AI with light-sheet microscopy to analyze 3D tumor microenvironments, moving beyond the limitations of flat, 2D tissue slices.
The machine learning workflow for lymphoma analysis, integrating multi-modal data from training to clinical application.

Key Insights: Multi-modal Fusion and Hidden Biomarkers

The real power of AI in oncology lies in cross-modal data fusion. By combining morphological features from pathology slides with metabolic data from PET/CT and molecular signatures from RNA-seq, AI creates a holistic patient profile far more predictive than any single modality. Furthermore, AI is identifying new prognostic indicators, such as spatial nuances in the tumor microenvironment that correlate with survival but are easily missed by the human eye. This allows for risk re-stratification beyond traditional tools like the International Prognostic Index (IPI), identifying high-risk patients even when standard metrics suggest a favorable prognosis.

Key Results: Quantitative Benchmarks

Recent literature highlights several impressive quantitative benchmarks achieved by AI models:

  • Diagnostic Accuracy: CNN models routinely achieve >95% accuracy in classifying complex lymphoma subtypes, significantly outperforming traditional pathological assessments which averaged around 83.3% in comparative studies.
  • Non-Invasive Triage: Deep learning models analyzing PET/CT images achieved up to 92.5% accuracy in diagnosing cervical lymphadenopathies, potentially reducing the need for invasive biopsies.
  • Prognostic Superiority: Machine learning models predicting primary treatment failure in DLBCL achieved an Area Under the Curve (AUC) of 0.83, outperforming existing standard clinical criteria.
Performance metrics and advantages of AI in lymphoma research, highlighting high accuracy and prognostic superiority in specific studies.

Limitations and Future Directions

Despite its promise, several hurdles remain. Training robust AI models requires massive, well-annotated datasets, which are difficult to collect for rare lymphoma subtypes. Additionally, the "black box" nature of deep learning models can lead to clinician hesitancy. Future work should pivot toward multimodal transformer models and graph neural networks (GNNs), which are better equipped to capture complex clinical relationships and generalize across different patient populations.

Summary updated

Abstract

Artificial intelligence (AI) is an important branch of computer science. With the rapid development of AI, the application in oncology has become increasingly widespread. As a hematologic malignancy characterized by remarkable heterogeneity, lymphoma has long posed significant challenges in both diagnosis and treatment, particularly with regard to its complex classification and difficulties in prognostic evaluation. Breakthroughs in AI technology have provided a new paradigm for precision treatment of tumors. AI can integrate and analyze HE pathology slides and PET/CT images to enhance diagnostic efficiency; meanwhile, in terms of treatment prognosis, AI can identify biomarkers to accurately classify lymphoma subtypes for therapeutic guidance, simultaneously quantify biomarkers to minimize the influence of subjective variability, and predict the patients' prognosis based on the extracted features to assist in the precise treatment of lymphoma. This review aims to provide an overview of AI related to various fields of lymphoma, introducing the core technologies and principles of AI, including deep learning, decision trees, regression models, and so on. We also discuss the clinical applications of AI in lymphoma pathology slides and PET/CT images, systematically analyze the clinical applications of AI in diagnosis, treatment, and prognosis, as well as innovatively summarize the cutting-edge applications of AI in 3D pathology of lymphomas, and finally, we emphasize the development potentials and current challenges of AI in the field of lymphomas to promote precision lymphoma diagnosis and treatment by providing a theoretical foundation.
BibTeX
@inproceedings{kang2025clinical,
  author = {Mengyao Kang and Zibo Yang and Yu Tian and Dongyu Li and Zhiqiong Wang and Liting Chen},
  title = {{Clinical applications of artificial intelligence in the histopathology of lymphoma: diagnosis, treatment and prognosis}},
  booktitle = {Discover},
  year = {2025},
  doi = {10.1007/s12672-025-03970-6},
}
Figure