Automatic Segmentation of Primary Central Nervous System Lymphoma at Clinical Routine Postcontrast T1-weighted MRI.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Primary central nervous system lymphoma (PCNSL) is a rare brain tumor requiring accurate segmentation for treatment planning.
• Manual segmentation of PCNSL on MRI is time-consuming and subject to inter-observer variability.
• Deep learning offers potential for automated and standardized segmentation of brain tumors.

Purpose of the Study:

• To develop and validate a deep learning model for automatic segmentation of PCNSL using postcontrast T1-weighted MRI.
• To assess the model's performance and robustness across internal and external datasets from multiple clinical centers.

Main Methods:

• Retrospective collection of data from patients with pathologically proven immunocompetent PCNSL.
• Training and validation of a deep learning model using the nnU-Net framework on postcontrast T1-weighted MRI scans.
• Performance evaluation using Dice score, mean average surface distance, and F1 score, with comparisons using Mann-Whitney U test and bootstrap resampling.

Main Results:

• The deep learning model achieved high performance with a mean Dice score of 0.84 (internal) and 0.88 (external).
• Strong volumetric correlation (r=0.99 internal, r=0.98 external) was observed between automatic and manual segmentations.
• Model performance was consistent across multiple centers, despite variations in MRI acquisition parameters.

Conclusions:

• A deep learning model provides accurate and robust automatic segmentation of PCNSL.
• This AI-driven approach has the potential to standardize PCNSL segmentation and improve diagnostic workflows.
• The model demonstrates generalizability across different clinical settings for brain tumor analysis.