Automatic Segmentation of Vestibular Schwannoma From MRI Using Two Cascaded Deep Learning Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of vestibular schwannoma (VS) is crucial for monitoring tumor growth and planning treatment.
• Deep learning models for VS segmentation face generalization challenges due to tumor variability.

Purpose of the Study:

• To introduce a novel deep learning model combining two Convolutional Neural Network (CNN) models for improved automatic segmentation and detection of vestibular schwannoma (VS).
• To address the generalization challenges in VS segmentation caused by tumor variability.

Main Methods:

• A sequential connection of two UNet models was developed, where the output of the first UNet is refined by a second complementary network.
• Spatial attention mechanisms were incorporated into the second network to guide the refinement process.
• Experiments were conducted on both public and private datasets using contrast-enhanced T1 and high-resolution T2-weighted MRI scans.

Main Results:

• The novel model demonstrated consistent improvements in Dice scores across 2D, 2.5D, and 3D CNN variants on public datasets, with an 8.86% enhancement for the 2D UNet on T1-weighted MRI.
• A 3.75% improvement was observed for the 2D UNet on T1-weighted MRI in the private dataset.
• T1-weighted MRI scans generally yielded better VS segmentation results compared to T2-weighted scans.

Conclusions:

• Sequential connection of UNets coupled with spatial attention mechanisms significantly enhances VS segmentation performance.
• The proposed model improves upon state-of-the-art 2D, 2.5D, and 3D deep learning methods for VS segmentation.