Enhanced CATBraTS for Brain Tumour Semantic Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Early and precise brain tumour identification is crucial for patient survival, often requiring efficient segmentation of medical images.
• Deep learning models in computer vision have advanced automatic tumour segmentation, improving boundary delineation accuracy.

Purpose of the Study:

• To introduce Enhanced Channel Attention Transformer (E-CATBraTS), a novel automated model for brain tumour semantic segmentation.
• To demonstrate improved segmentation accuracy and statistical stability using diverse multi-modal MRI datasets.

Main Methods:

• Developed E-CATBraTS, integrating convolutional neural networks and Swin Transformer with channel shuffling and attention mechanisms.
• Utilized a vision transformer architecture (3D CATBraTS) as a foundation for the new model.
• Evaluated the model on four datasets comprising 3137 brain MRI scans.

Main Results:

• E-CATBraTS significantly improved segmentation accuracy on two datasets, outperforming state-of-the-art models by a mean DSC of 2.6%.
• The model maintained high accuracy comparable to top performers on other datasets.
• Demonstrated robust generalization abilities across varied dataset acquisition parameters.

Conclusions:

• E-CATBraTS achieves high brain tumour segmentation accuracy and superior generalization capabilities.
• The model's robustness to dataset variations makes it a reliable tool for clinical applications.
• This advancement aids in precise tumour identification, potentially improving patient prognoses.