Robust Multimodal Deep Learning for Lymphoma Subtype Classification Using ¹⁸F-FDG PET Maximum Intensity Projection...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Classifying lymphoma subtypes using 18F-FDG PET imaging is challenging due to imaging inconsistencies and variability.
• Previous methods lacked standardization in protocols, scanner types, and inter-institutional data.
• A multimodal deep learning framework is proposed to address these limitations.

Purpose of the Study:

• To develop and validate a deep learning framework for hierarchical classification of lymphoma subtypes.
• To integrate harmonized PET imaging features with structured clinical information.
• To overcome inter-institutional variability and scanner inconsistencies in lymphoma classification.

Main Methods:

• Collected multi-center data including 18F-FDG PET images and clinical variables.
• Implemented a Scanner-Conditioned Normalization (SCN) module to harmonize feature distributions.
• Utilized a multimodal deep learning model integrating anterior and lateral MIP images with clinical data.

Main Results:

• Achieved AUC of 0.89 (internal) and 0.84 (external) for Hodgkin vs. non-Hodgkin lymphoma classification.
• Achieved AUC of 0.84 (internal) and 0.76 (external) for DLBCL vs. FL classification.
• Demonstrated robust performance across internal and external validation cohorts.

Conclusions:

• Deep learning with 18F-FDG PET and clinical data shows potential for lymphoma subtype classification.
• The proposed framework, LymphoMAP, offers a step towards assistive tools for early clinical decision-making.
• Automated pipelines can support, not replace, conventional diagnostic workflows in personalized lymphoma management.