Related Experiment Video
Updated: Jul 11, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Multi-class glioma segmentation on real-world data with missing MRI sequences: comparison of three deep learning
Hugh G Pemberton1,2, Jiaming Wu1, Ivar Kommers3
1Centre for Medical Image Computing (CMIC), University College London, London, UK.
Scientific Reports
|November 3, 2023
Summary
This study evaluated three AI models for brain tumor segmentation, finding that nn-Unet performed best on multi-center MRI data, even with missing sequences. This supports automated glioma segmentation for clinical use.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuro-oncology
Background:
- Brain Tumor Segmentation (BraTS) challenge models are crucial for neuro-oncology.
- Generalizability of these models to diverse, real-world clinical data is essential for adoption.
- Variations in image quality and incomplete datasets pose challenges for automated segmentation.
Purpose of the Study:
- To test the generalizability of three BraTS challenge models (DeepMedic, nn-Unet, NVIDIA-net) on a multi-center dataset.
- To evaluate model performance with varying image quality and incomplete MRI data.
- To assess the clinical feasibility of automated glioma segmentation.
Main Methods:
- Retrospective study using preoperative MRI data for glioblastoma (GBM) and low-grade gliomas (LGG).
- Models trained and tested on BraTS 2021 and clinical data from 12 hospitals.
- Performance evaluated using Dice Similarity Coefficient (DSC) and Hausdorff distance on internal and external test sets.
- Sparsified training applied to assess impact of missing MRI sequences.
Main Results:
- All tested models achieved median DSC within or above human inter-rater agreement (0.74-0.85).
- nn-Unet demonstrated superior performance with the highest DSC (internal=0.86, external=0.93) and lowest Hausdorff distances.
- Sparsified training indicated that missing MRI sequences did not significantly impact performance.
- nn-Unet showed robust segmentation accuracy across different clinical settings and data completeness.
Conclusions:
- nn-Unet achieves accurate glioma segmentation in clinical settings, even with incomplete MRI datasets.
- The model's generalizability supports its potential for clinical adoption.
- Automated segmentation can aid in treatment planning and monitoring of gliomas.
