Predicting 2-year time to progression in diffuse large B cell lymphoma using 3D CNNs on whole-body PET/CT scans

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiomics

Background:

• Diffuse large B-cell lymphoma (DLBCL) patient outcomes require improved predictive models.
• Current prognostic tools like the International Prognostic Index (IPI) have limitations.
• Positron Emission Tomography/Computed Tomography (PET/CT) scans offer rich data for prediction.

Purpose of the Study:

• To develop and evaluate 3D convolutional neural networks (CNNs) for predicting 2-year time to progression in DLBCL patients.
• To compare the predictive performance of 3D CNNs against the IPI and a 2D CNN (MIP-CNN).

Main Methods:

• Development of two 3D CNN models (L-PET3D-CNN, LW-PET3D-CNN) using baseline PET/CT scans from DLBCL patients.
• Training on a dataset of 636 patient scans and independent testing on 496 scans from five external trials.
• Performance evaluation using Area Under the Curve (AUC) and comparison with IPI and MIP-CNN via DeLong test; occlusion maps used for interpretability.

Main Results:

• The 3D CNN models (L-PET3D-CNN and LW-PET3D-CNN) achieved AUCs of 0.65 and 0.64, significantly outperforming the IPI (AUC 0.53).
• Both 3D CNN models demonstrated consistent superiority over IPI across individual external clinical trials.
• The performance of the 3D CNNs was comparable to the 2D MIP-CNN model (AUC 0.65).

Conclusions:

• 3D CNN models are predictive of outcome in DLBCL patients across external datasets, surpassing the IPI.
• While performance is similar to 2D CNNs, 3D CNNs offer enhanced interpretability through 3D occlusion maps.
• These findings support the potential of 3D CNNs for improving prognostic accuracy and understanding in DLBCL.