An approach to building foundation models for brain image analysis.

Area of Science:

• Artificial Intelligence
• Neuroimaging
• Machine Learning

Background:

• Supervised machine learning for brain image analysis requires extensive labeled datasets, which are often costly and time-consuming to acquire.
• Existing models are task-specific, limiting their utility and requiring retraining for new applications.
• The need for efficient and versatile methods for analyzing complex neuroimaging data is critical.

Purpose of the Study:

• To develop a novel foundation model for brain image analysis that overcomes limitations of supervised learning.
• To leverage self-supervised learning for reduced reliance on labeled neuroimaging datasets.
• To create a versatile model applicable to diverse brain image analysis tasks.

Main Methods:

• An attention-based neural network foundation model was developed.
• A novel self-supervised approach trained the model to generate brain images patch-wise, learning intrinsic brain structures.
• High-frequency information was encoded using convolutional kernels with random weights to enhance learning of image details.

Main Results:

• The foundation model was trained on 10 public datasets and validated on five independent datasets.
• The model demonstrated competitive or superior performance in segmentation, lesion detection, denoising, and brain age estimation.
• A significant reduction in the need for labeled training data was observed across all evaluated tasks.

Conclusions:

• The proposed foundation model effectively utilizes large unlabeled neuroimaging datasets for diverse analysis tasks.
• This self-supervised approach significantly reduces the time and cost associated with acquiring labeled data in brain image analysis.
• The method offers a scalable and efficient solution for advancing machine learning applications in neuroscience.