MR-based glioblastoma tumour detection and segmentation (EUCAIM-SW-021_T-01-02-004)

MR-based glioblastoma tumour detection and segmentation (EUCAIM-SW-021_T-01-02-004) performs automatic segmentation of glioblastoma tumors and subregions in MRI using the nnU-Net convolutional neural network from the PRIMAGE project to enable precise tumor delineation for research and clinical assessment.

Key Features:

• Automatic Segmentation: Automates segmentation of glioblastoma tumors and subregions on MRI images.
• Algorithm: Implements the nnU-Net convolutional neural network (CNN) architecture developed within the PRIMAGE project.
• Accuracy: Reported median Dice Similarity Coefficient (DSC) values include 0.965 and 0.997 in separate studies, indicating high overlap with manual segmentations.
• Robustness Across Imaging Variability: Maintains performance across variations in magnetic field strength, type of T2 sequence, tumor location, and in MR images acquired post-chemotherapy.
• Efficiency: Demonstrated a 92.8% reduction in segmentation time compared to manual methods in one study.
• Inter-Observer Consistency: Performance is comparable to expert radiologists, showing similar variability and reliability in segmentation.
• Multicenter Training Data: Trained on a diverse dataset of MR images collected from multiple centers.

Scientific Applications:

• Clinical Diagnosis and Treatment Planning: Provides precise tumor delineation to support diagnosis and treatment planning for glioblastoma.
• Research and Development: Enables large-scale processing of MRI datasets for studies of tumor characteristics, treatment response, and outcome prediction.

Methodology:

Training of the nnU-Net model on a diverse multicenter MR image dataset with cross-validation to ensure robustness and generalizability; the trained model performs automatic segmentation with the option for minor manual edits by radiologists.