Synthetic MRI pretraining for medical imaging classification

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep learning (DL) models excel in medical imaging but require large, diverse datasets.
• Transfer learning, using models pretrained on natural images (e.g., ImageNet) or large unlabeled medical datasets, is a common solution.
• Scarcity of large-scale medical imaging data limits current DL model performance.

Purpose of the Study:

• To generate and utilize a synthetic MRI dataset for pretraining DL architectures.
• To evaluate the effectiveness of synthetic pretraining for medical imaging tasks with minimal computational cost.
• To compare synthetic pretraining against ImageNet-pretrained, foundation, and self-supervised models.

Main Methods:

• Generation of a synthetic MRI dataset.
• Pretraining of DL architectures using the synthetic dataset.
• Evaluation across downstream tasks: brain tumor classification and MedMNIST benchmarks (2D and 3D).

Main Results:

• Synthetic pretraining consistently improved feature representations and downstream task performance.
• The proposed method outperformed ImageNet-pretrained, foundation, and self-supervised models.
• A new state-of-the-art was established on the MedMNIST benchmark.

Conclusions:

• Synthetic MRI pretraining offers an effective and computationally efficient solution for medical DL.
• This approach significantly enhances model performance on various medical imaging tasks.
• Synthetic data pretraining represents a promising direction for advancing medical AI.