An artificial intelligence method using FDG PET to predict treatment outcome in diffuse large B cell lymphoma...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Diffuse large B-cell lymphoma (DLBCL) treatment response prediction is crucial for patient outcomes.
• The International Prognostic Index (IPI) is a standard clinical tool, but its predictive accuracy can be limited.
• ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) positron emission tomography (PET) scans provide valuable metabolic information.

Purpose of the Study:

• To investigate the feasibility of using Convolutional Neural Networks (CNNs) with ¹⁸F-FDG PET maximum intensity projection (MIP) images for predicting time-to-progression (TTP) in DLBCL.
• To compare the predictive performance of the CNN model against the established International Prognostic Index (IPI).

Main Methods:

• A CNN model was developed using baseline ¹⁸F-FDG PET/CT MIP images from 296 DLBCL patients in the HOVON-84 trial.
• The model's performance was validated on an external dataset of 340 DLBCL patients.
• Cross-validation was performed using coronal and sagittal MIPs; associations with tumor volume and probabilities after tumor removal were assessed.

Main Results:

• The CNN model achieved an area under the curve (AUC) of 0.74 for predicting 2-year TTP, outperforming the IPI-based model (AUC = 0.68).
• High predicted probabilities from the CNN significantly decreased after synthetically removing tumors from the PET images.
• The model demonstrated the ability to predict treatment outcome in DLBCL.

Conclusions:

• Maximum intensity projection (MIP)-based CNNs show promise as a tool for predicting treatment outcomes in DLBCL.
• This AI approach may offer improved prognostic accuracy compared to conventional methods like the IPI.
• Further validation and integration into clinical practice could enhance personalized treatment strategies for DLBCL patients.