Multimodal deep learning model on interim [¹⁸F]FDG PET/CT for predicting primary treatment failure in diffuse large...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Predicting primary treatment failure (PTF) in diffuse large B-cell lymphoma (DLBCL) is essential for improving patient outcomes.
• Interim 18F-fluoro-2-deoxyglucose (18F-FDG) positron emission tomography/computed tomography (PET/CT) imaging offers potential for early prediction.

Purpose of the Study:

• To develop and evaluate multimodal deep learning (MDL) models for predicting PTF in DLBCL patients using interim PET/CT data.
• To explore different multimodal fusion strategies to optimize prediction performance.

Main Methods:

• Developed five MDL models based on the Conv-LSTM network, incorporating various fusion strategies (pixel intermixing, separate channel, separate branch, quantitative weighting, hybrid learning).
• Utilized a primary dataset of 205 DLBCL patients and an external dataset of 44 patients for validation.
• Optimized the best-performing hybrid learning model using a contrastive training objective.

Main Results:

• The optimized contrastive hybrid learning model achieved 91.22% accuracy and an AUC of 0.926 on the primary dataset.
• The model demonstrated good generalization ability with 88.64% accuracy and an AUC of 0.925 on the external dataset.
• Deep learning analysis validated the predictive value of interim PET/CT, potentially exceeding human interpretation capabilities.

Conclusions:

• The proposed contrastive hybrid learning model accurately predicts primary treatment failure in DLBCL patients.
• The study highlights the significant predictive value of interim PET/CT imaging in DLBCL.
• These findings support the use of MDL models for guiding individualized clinical treatment strategies.