PET-based lesion graphs meet clinical data: An interpretable cross-attention framework for DLBCL treatment response...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Diffuse Large B-cell Lymphoma (DLBCL) is a growing concern in lymphatic cancers.
• Current diagnosis and follow-up rely on clinical biomarkers and 18F-Fluorodeoxyglucose (FDG)-PET/CT imaging.
• Early identification of high-risk DLBCL patients is crucial for effective treatment strategies.

Purpose of the Study:

• To develop a novel framework for early identification of high-risk DLBCL patients.
• To integrate information from both PET/CT images and tabular clinical data.
• To improve prognostic prediction of 2-year progression-free survival (PFS).

Main Methods:

• A graph neural network (GNN) model was developed to represent PET/CT images as attributed lesion graphs.
• A cross-attention module was designed to fuse image-derived attributes with clinical indicators.
• The framework integrates information across multiple lesions and data modalities for prediction.

Main Results:

• The proposed GNN framework effectively integrates multi-lesion image information, outperforming models using only clinical data.
• Experimental validation on a 545-patient multicentric dataset demonstrated the framework's efficacy.
• The graph-based design offers interpretability, allowing tracing predictions to key lesions and features.

Conclusions:

• The developed GNN model provides an effective approach for integrating diverse data sources in DLBCL risk stratification.
• This method enhances early identification of high-risk patients, potentially improving treatment outcomes.
• The interpretability of the model facilitates a deeper understanding of prognostic factors in DLBCL.