Synthetic MRI Pretraining for Medical Imaging Tasks

Deep learning (DL) models have reached remarkable achievements in medical imaging, but their performance heavily depends on the availability of large and diverse datasets. To overcome this, transfer learning has emerged as a widely adopted solution, where models pretrained on large datasets are fine-tuned for specific medical tasks. Due to the scarcity of large-scale medical imaging datasets, most existing models are pretrained on natural image datasets such as ImageNet, while recent studies have explored high complex models trained on a vast amount of diverse medical unlabelled datasets. In this work, we generate and leverage a dataset of 9,000 MRI scans to pretrain DL architectures, demonstrating effective model learning with minimal computational cost and robustness to pretraining dataset size. We evaluate our approach across multiple clinical tasks, including brain tumour classification and ten benchmark tasks from MedMNIST covering 2D and 3D modalities. Compared with ImageNet-pretrained, foundation, and self-supervised models, synthetic pretraining consistently improves feature representations and downstream performance. Our results show that models pretrained on synthetic medical data outperform both ImageNet-pretrained models and those trained from scratch. Overall, our approach outperforms competing methods, establishing a new state-of-the-art on the MedMNIST benchmark.