Improving Representation of High-frequency Components for Medical Visual Foundation Models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Foundation models show broad applicability but lack precision in high-frequency details crucial for medical imaging.
• Intricate anatomical structures and subtle features in medical scans necessitate improved representation capabilities.

Purpose of the Study:

• To introduce a novel pretraining strategy, Frequency-advanced Representation Autoencoder (Frepa), for medical image foundation models.
• To enhance the representation of high-frequency components and fine-grained details in medical images.

Main Methods:

• Frepa employs high-frequency masking, low-frequency perturbation, and embedding consistency learning for pretraining 2D and 3D medical images.
• A histogram-equalized image masking strategy extends Masked Autoencoder principles to various architectures beyond Vision Transformers (ViT).

Main Results:

• Frepa significantly outperforms existing self-supervised pretraining methods across nine modalities and 32 downstream tasks.
• Improvements include a +15% dice score in retina vessel segmentation and +8% IoU in lung tumor detection without fine-tuning.
• Quantitative analysis confirms Frepa's superior high-frequency representation and preservation in image embeddings.

Conclusions:

• Frepa addresses limitations of current foundation models in capturing critical high-frequency details in medical imaging.
• The proposed strategy demonstrates potential for developing more generalized and universal medical image foundation models.