AI-enabled precise brain tumor segmentation by integrating Refinenet and contour-constrained features in MRI images

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Transformer-based models in healthcare

Background:

• Medical image segmentation is crucial across various medical fields.
• The Segment Anything Model (SAM) shows promise but struggles with medical image complexities like variable styles and unclear boundaries.
• Direct application of SAM to medical imaging is limited by these challenges.

Purpose of the Study:

• To improve the Segment Anything Model (SAM) for medical image segmentation.
• Introduce the SAM-RCCF framework to enhance robustness and generalizability for intracranial tumors.
• Improve segmentation precision for gliomas, metastatic tumors, and meningiomas.

Main Methods:

• Utilized 484 T1CE-weighted MRI scans from brain tumor patients (glioma, metastatic, meningioma).
• Developed the SAM-RCCF framework integrating RefineNet and conditional control for precise feature recognition.
• Employed five-fold cross-validation for performance evaluation.

Main Results:

• SAM-RCCF achieved high performance in glioma segmentation (IOU: 0.90, DSC: 0.912) and meningioma segmentation (IOU: 0.9214, DSC: 0.93).
• The model significantly outperformed classic segmentation models on these tasks.
• Demonstrated superior segmentation accuracy and robustness on complex brain tumor MRI data.

Conclusions:

• The SAM-RCCF algorithm significantly surpasses the original SAM in segmenting brain tumors (glioma, metastatic, meningioma).
• Validated the framework's effectiveness for complex and variable medical images.
• Enhanced segmentation accuracy and robustness in medical image analysis.