# Robust Multimodal Deep Learning for Lymphoma Subtype Classification Using 18F-FDG PET Maximum Intensity Projection Images and Clinical Data: A Multi-Center Study

## TL;DR

A deep learning model combining PET imaging and clinical data accurately classifies lymphoma subtypes, improving diagnostic consistency and early assessment.

## Contribution

A novel multimodal deep learning framework with scanner harmonization for lymphoma subtype classification using PET and clinical data.

## Key findings

- The model achieved 89% internal and 84% external accuracy in distinguishing Hodgkin from non-Hodgkin lymphoma.
- It showed 0.84 internal and 0.76 external AUC for diffuse large B-cell versus follicular lymphoma classification.
- The model integrates PET images and clinical data to improve diagnostic consistency and early subtype assessment.

## Abstract

Lymphoma subtypes require different therapeutic strategies. However, accurate classification is challenging due to variable imaging phenotypes and patient characteristics. Although histopathology is the gold standard, noninvasive fluorine-18 fluorodeoxyglucose (18F-FDG) positron emission tomography (PET) provides complementary subtype-specific information. In this study, we developed a deep learning model integrating 18F-FDG PET images with structured clinical data. We applied harmonization techniques to multi-institutional datasets from six centers. The model achieved 89% (internal) and 84% (external) accuracy in distinguishing Hodgkin from non-Hodgkin lymphoma. This approach improves diagnostic consistency and reproducibility, facilitates early subtype assessment prior to histopathological confirmation, and holds promise for broader applications in imaging-based disease classification.

Background: Previous attempts to classify lymphoma subtypes based on metabolic features extracted from 18F-FDG PET imaging have been hindered by inconsistencies in imaging protocols, scanner types, and inter-institutional variability. To overcome these limitations, we propose a multimodal deep learning framework that integrates harmonized PET imaging features with structured clinical information. The proposed framework is designed to perform hierarchical classification of clinically meaningful lymphoma subtypes through two sequential binary classification tasks. Methods: We collected multi-center data comprising 18F-FDG PET images and structured clinical variables of patients with lymphoma. To mitigate domain shifts caused by different scanner manufacturers, we integrated a Scanner-Conditioned Normalization (SCN) module, which adaptively harmonizes feature distributions using manufacturer-specific parameters. Performance was validated using internal and external cohorts, with the statistical significance of performance gains assessed via DeLong’s test and bootstrap-based CI analysis. Results: The proposed model achieved an area under the curve (AUC) of 0.89 (internal) and 0.84 (external) for Hodgkin lymphoma versus non-Hodgkin lymphoma classification and 0.84 (internal) and 0.76 (external) for diffuse large B-cell lymphoma versus follicular lymphoma classification (p > 0.05). These results were obtained using a multimodal model that integrated anterior and lateral maximum intensity projection (MIP) images with clinical data. Conclusions: This study demonstrates the potential of a deep learning-based approach for lymphoma subtype classification using non-invasive 18F-FDG PET imaging combined with clinical data. While further validation in larger, more diverse cohorts is necessary to address the challenges of rare subtypes and biological heterogeneity, LymphoMAP serves as a meaningful step toward developing assistive tools for early clinical decision-making. These findings underscore the feasibility of using automated pipelines to support, rather than replace, conventional diagnostic workflows in personalized lymphoma management.