CarveMix: A simple data augmentation method for brain lesion segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain lesion segmentation is crucial for clinical diagnosis and research.
• Convolutional Neural Networks (CNNs) excel at image segmentation tasks.
• Data augmentation, particularly image mixing, is vital for improving CNN performance.

Purpose of the Study:

• To develop a lesion-aware data augmentation method for brain lesion segmentation.
• To address limitations of existing image mixing methods in brain lesion segmentation.
• To improve the accuracy and robustness of CNNs in segmenting brain lesions.

Main Methods:

• Propose CarveMix, a lesion-aware data augmentation approach for CNN-based brain lesion segmentation.
• Stochastically combine annotated brain lesion images, preserving lesion information via a carved Region of Interest (ROI).
• Incorporate harmonization steps for heterogeneous data and model mass effect for tumor segmentation.

Main Results:

• CarveMix significantly improves the accuracy of brain lesion segmentation across multiple datasets.
• The proposed method demonstrates effectiveness in enhancing CNN performance for brain lesion segmentation.
• Experimental validation confirms the superiority of CarveMix over existing augmentation techniques.

Conclusions:

• CarveMix offers a simple yet effective solution for brain lesion segmentation data augmentation.
• The lesion-aware approach enhances the utility of image mixing for medical imaging tasks.
• This method provides a valuable tool for advancing CNN-based brain lesion analysis.