MR-based neuroblastoma tumour detection and segmentation (EUCAIM-SW-019_T-01-02-002)

MR-based neuroblastoma tumour detection and segmentation (EUCAIM-SW-019_T-01-02-002) performs automated detection and segmentation of primary neuroblastoma tumours in magnetic resonance (MR) images using the nnU-Net convolutional neural network to support diagnostic assessment and treatment planning.

Key Features:

• Automatic Segmentation: Employs a fully automatic nnU-Net convolutional neural network (CNN) for segmentation of primary neuroblastoma tumours in MR images, developed within the PRIMAGE context.
• High Accuracy: Achieves a median Dice Similarity Coefficient (DSC) of 0.997 across a cohort of 300 children with neuroblastic tumours, indicating near-perfect agreement with expert-edited masks.
• Inter-Observer Variability: Demonstrates performance comparable to manual segmentations by radiologists, showing similar inter-observer variability.
• Efficiency: Reduces segmentation-related time from an average of 124 seconds for manual editing to approximately 7.9 seconds for automatic visual inspection, representing about 92.8% time savings.
• Robustness Across Variability: Maintains high accuracy across different MR magnetic fields, T2 sequence types, tumour locations, and images acquired before or after chemotherapy.
• Minimal Manual Intervention: Allows minor manual adjustments when necessary while operating primarily in an automatic mode.
• Evaluation Metrics: Uses Dice Similarity Coefficient (DSC), False Positive Rate (FPRm), and False Negative Rate (FNR) for performance assessment.

Scientific Applications:

• Clinical Diagnostics: Provides precise tumour segmentations to aid diagnosis, treatment planning, and monitoring of neuroblastoma.
• Research and Development: Enables large-scale, consistent segmentation outputs for studies of neuroblastic tumours.
• Multicenter Studies: Supports evaluation and comparison across heterogeneous datasets from multiple centers due to demonstrated cross-site robustness.

Methodology:

The nnU-Net model is trained on a diverse dataset of MR images with cross-validation, then validated on an independent cohort, and performance is evaluated using DSC, FPRm, and FNR.