Handling missing MRI sequences in deep learning segmentation of brain metastases: a multicenter study
Endre Grøvik1,2,3, Darvin Yi4, Michael Iv2
1Department of Diagnostic Physics, Oslo University Hospital, Oslo, Norway.
NPJ Digital Medicine
|February 23, 2021
Summary
A novel deep learning model accurately detects and segments brain metastases, even with missing MRI sequences. This input-level dropout model demonstrates superior performance compared to existing methods, improving clinical utility.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Brain metastases detection and segmentation are crucial for treatment planning.
- Deep learning models show promise in medical image analysis.
- The challenge of missing MRI sequences can hinder model performance.
Purpose of the Study:
- To assess the clinical value of a deep learning (DL) model for automatic detection and segmentation of brain metastases.
- To evaluate a novel input-level dropout (ILD) layer simulating missing MRI sequences during training.
- To compare the ILD model's performance against a state-of-the-art DeepLab V3 model.
Main Methods:
- Retrospective, multicenter study involving 165 patients with brain metastases.
- Training a neural network on four distinct MRI sequences using an ILD layer.
- Validation and testing on datasets with and without specific MRI sequences.
- Comparison of segmentation results using Dice score, IoU, and ROC statistics.
Main Results:
- The ILD-model achieved comparable AUC to DeepLab V3 (0.989 vs. 0.989).
- ILD-model demonstrated significantly higher Dice score (0.795 vs. 0.774) and IoU score (0.561 vs. 0.492).
- ILD-model exhibited a significantly lower false positive rate (3.6/patient vs. 7.0/patient).
Conclusions:
- The ILD-model facilitates accurate detection and segmentation of brain metastases.
- The ILD approach enhances model robustness, even with missing MRI sequences.
- This method holds potential for multicenter application in brain metastases management.
