Abstract
Background
Manual segmentation of brain tumors requires expertise, is time-consuming, and is subject to inter-rater variability. Fully automatic brain tumor segmentation is possible for glioma and meningioma when volumetric T1, T1 contrast-enhanced (T1c), T2, and Fluid-attenuated inversion recovery (FLAIR) MRIs are available. In clinical care of meningiomas, however, often only volumetric T1c scans are available. In this work, we trained a deep learning network to segment meningiomas using only T1c scans for use in clinical research.
Material and Methods
NnU-Net, a deep learning model that is optimized for medical image segmentation, was trained to segment meningiomas from T1c images. This was performed on a large clinically collected meningioma dataset (n=374) of T1c scans with semi-automatically generated enhancing tumor masks and additional data from the BraTS2020 glioma dataset. Model performance was compared against inter-rater reliability, between different models, between anatomical tumor locations, and against models using multiple MRI modalities.
Results
The best performing model obtained a Dice score of 0.90. This performance was 0.03 points lower when compared to inter-rater reliability (Dice=0.93) and almost equal to models using multiple MRI modalities. Model performance split over anatomical tumor locations was between 0.90 and 0.97 (Dice).
Conclusion
Fully automatic meningioma segmentation using only T1c images is possible with an accuracy that is similar to inter-rater reliability and models using multiple imaging modalities.
Manual segmentation of brain tumors requires expertise, is time-consuming, and is subject to inter-rater variability. Fully automatic brain tumor segmentation is possible for glioma and meningioma when volumetric T1, T1 contrast-enhanced (T1c), T2, and Fluid-attenuated inversion recovery (FLAIR) MRIs are available. In clinical care of meningiomas, however, often only volumetric T1c scans are available. In this work, we trained a deep learning network to segment meningiomas using only T1c scans for use in clinical research.
Material and Methods
NnU-Net, a deep learning model that is optimized for medical image segmentation, was trained to segment meningiomas from T1c images. This was performed on a large clinically collected meningioma dataset (n=374) of T1c scans with semi-automatically generated enhancing tumor masks and additional data from the BraTS2020 glioma dataset. Model performance was compared against inter-rater reliability, between different models, between anatomical tumor locations, and against models using multiple MRI modalities.
Results
The best performing model obtained a Dice score of 0.90. This performance was 0.03 points lower when compared to inter-rater reliability (Dice=0.93) and almost equal to models using multiple MRI modalities. Model performance split over anatomical tumor locations was between 0.90 and 0.97 (Dice).
Conclusion
Fully automatic meningioma segmentation using only T1c images is possible with an accuracy that is similar to inter-rater reliability and models using multiple imaging modalities.
| Original language | English |
|---|---|
| Article number | P18.08.B |
| Pages (from-to) | ii95–ii96 |
| Journal | Neuro-Oncology |
| Volume | 24 |
| Issue number | Suppl 2 |
| DOIs | |
| Publication status | Published - 2022 |