AG-MS3D-CNN multiscale attention guided 3D convolutional neural network for robust brain tumor segmentation across...

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate brain tumor segmentation from multimodal MRI is crucial for neuro-oncology but faces challenges with manual methods and current deep learning models.
• Existing deep learning approaches struggle with generalization, precise boundary delineation, and reliable uncertainty estimation.

Purpose of the Study:

• To develop an attention-guided multiscale 3D convolutional neural network (AG-MS3D-CNN) for automated and robust brain tumor segmentation.
• To improve tumor boundary delineation, provide uncertainty estimation, and ensure generalizability across diverse datasets and MRI protocols.

Main Methods:

• Proposed AG-MS3D-CNN model integrating multiscale feature extraction and spatial attention mechanisms.
• Incorporated Monte Carlo dropout for uncertainty estimation and a domain adaptation module for enhanced generalizability.
• Utilized a multitask learning framework for simultaneous segmentation, classification, and volume estimation.

Main Results:

• AG-MS3D-CNN demonstrated superior performance compared to state-of-the-art methods on the BraTS 2021 and external datasets (OASIS, ADNI, IXI).
• Achieved high Dice scores, indicating excellent segmentation accuracy and robustness across varied MRI data.
• The model provides valuable confidence scores for segmentation, aiding clinical decision support.

Conclusions:

• AG-MS3D-CNN offers a robust and accurate solution for automated brain tumor segmentation from multimodal MRI.
• The integration of attention, multiscale features, uncertainty estimation, and domain adaptation enhances clinical utility in neuro-oncology.
• This model represents a significant advancement for computer-aided diagnosis and treatment planning in brain tumor management.