Brain tumor segmentation using 3D Mask R-CNN for dynamic susceptibility contrast enhanced perfusion imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Neoplasm segmentation is crucial for radiotherapy, disease monitoring, and outcome prediction.
• Functional MRI, including DSCE and DCE perfusion imaging, aids in brain tumor diagnosis, histology assessment, grading, and biopsy guidance.
• Manual tumor contouring is time-consuming, costly, and prone to inter-observer variability.

Purpose of the Study:

• To develop and validate an automated 3D Mask R-CNN method for segmenting brain tumors in DSCE MRI perfusion images.
• To simultaneously achieve region-of-interest (ROI) localization and voxel-wise segmentation of brain tumors.
• To compare the performance of the proposed method against state-of-the-art techniques and physician-delineated ground truth.

Main Methods:

• A 3D Mask R-CNN deep learning model was employed for automated tumor segmentation.
• The training process incorporated ROI localization, regression, and voxel-wise segmentation.
• A combined loss function, including classification, ROI regression, and segmentation losses, was used for network supervision.

Main Results:

• The 3D Mask R-CNN method achieved high accuracy in segmenting brain tumors from DSCE perfusion MRI.
• Quantitative metrics demonstrated strong agreement with ground truth contours, including an average Dice Similarity Coefficient (DSC) of 0.90 ± 0.04.
• The method showed promising results for Hausdorff distance, mean surface distance, and center of mass distance.

Conclusions:

• The proposed 3D Mask R-CNN method is feasible for accurate brain tumor localization and segmentation in DSCE perfusion MRI.
• This automated approach offers a promising alternative to manual contouring, reducing time and variability.
• The method holds significant potential for future clinical applications in neuro-oncology treatment planning and monitoring.