Learning General-Purpose Biomedical Volume Representations using Randomized Synthesis

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Current 3D biomedical foundation models exhibit limited generalization due to small, non-diverse public datasets.
• Existing models struggle with variations in medical procedures, conditions, anatomy, and imaging protocols.

Purpose of the Study:

• To develop a representation learning method that improves the generalization of 3D biomedical foundation models.
• To enable a single 3D network to perform various voxel-level tasks across diverse medical contexts.

Main Methods:

• A data engine was created to synthesize highly variable training samples, anticipating domain shifts.
• A contrastive learning method was developed to pretrain a 3D network for stability against simulated imaging variations.
• The method enables dataset-agnostic initialization for fine-tuning on new datasets.

Main Results:

• The proposed method sets new benchmarks in both multimodality registration and few-shot segmentation.
• Achieved state-of-the-art performance without pre-training on any real-world medical image datasets.
• Demonstrated robust feature representations for downstream tasks.

Conclusions:

• The developed representation learning approach significantly enhances the generalization capabilities of 3D biomedical vision models.
• This method offers a powerful, dataset-agnostic initialization for various medical imaging tasks.
• The approach overcomes limitations of current models by proactively addressing domain shifts during training.