*Important notice: This news reports on an unedited version of an accepted paper and is awaiting final editing. Therefore, the paper should not be regarded as conclusive or treated as established information.
To address the limited availability of medical imaging data for deep learning, researchers recently evaluated a synthetic pretraining framework that pretrains neural networks using 9000 synthetic magnetic resonance imaging (MRI) projections in less than 10 minutes.
Study: Synthetic MRI pretraining for medical imaging classification. Image Credit: PeopleImages/Shutterstock.com
Large natural datasets such as ImageNet often perform poorly on medical imaging tasks due to domain differences. To overcome this, the model established a new benchmark across two-dimensional (2D) and three-dimensional (3D) MedMNIST classification tasks, improving performance by up to 17% compared with ImageNet-based pretraining.
Saving this for later? Download a free PDF here.
The design lays the foundations for a fast and practical alternative that could reduce reliance on the ethical, financial, and regulatory challenges of collecting clinical datasets. The findings were published in Scientific Reports.
Challenges in Obtaining Medical Imaging Data
Deep learning has significantly improved medical image analysis, but its success depends on large, well-annotated datasets that are often difficult to obtain. Clinical image collections are limited by patient privacy regulations, data-sharing restrictions, and the high cost of expert annotation. As a result, many researchers rely on transfer learning, initializing neural networks with models pretrained on natural image datasets such as ImageNet.
However, the differences between everyday photographs and medical images create a domain gap that limits performance. Large medical models have attempted to overcome this by training on millions of unlabeled clinical images, but this requires extensive computing resources, long training times, and large graphical processing unit (GPU) clusters, in turn increasing costs and energy consumption and limiting accessibility for many scientists.
Novel Framework for Image Pretraining
To overcome the limitations of conventional pretraining, researchers developed a framework that combines synthetic image generation with self-supervised learning. First, three class-conditioned denoising diffusion probabilistic models (DDPMs), based on a U-Net architecture, generated 9000 synthetic MRI slices.
The images were evenly distributed across the axial, coronal, and sagittal planes and represented three diagnostic groups, including healthy controls, mild cognitive impairment, and Alzheimer's disease.
Instead of pretraining neural networks to recognize disease categories, the models learned to identify the projection plane of each synthetic image. This self-supervised task encouraged the networks to learn general anatomical structures, spatial relationships, and geometric features that could be transferred to different medical imaging tasks.
The framework was then evaluated using three neural network architectures: an eight-block convolutional neural network with 74,000 parameters, ResNet18 (residual network 18 layers) with 11 million parameters, and ResNet50 with 23 million parameters.
To extend the pretrained models from two-dimensional images to three-dimensional scans, the researchers applied axial-coronal-sagittal (ACS) convolutions, which decompose standard two-dimensional kernels into orientation-specific kernels for volumetric medical image analysis.
Superior Performance in Medical Classification
The synthetic pretraining framework generally outperformed traditional ImageNet-pretrained models, self-supervised methods, and large medical foundation models across multiple medical imaging tasks.
In linear probing evaluations for brain tumor classification, the synthetic ResNet50 achieved an accuracy of about 71.32%, compared with 66.33% for LVM-Med (large vision model for medical imaging), 60.66% for ImageNet-pretrained models, and 57.87% for CLIP (contrastive language-image pretraining).
The approach also improved performance on both two-dimensional and three-dimensional benchmark datasets. The synthetic ResNet18 outperformed its ImageNet-pretrained counterpart by 16.7 percentage points on ultrasound breast cancer classification.
When extended to volumetric imaging using ACS convolutions, the pretrained models achieved an improvement of up to 11.1 percentage points on adrenal gland classification compared with networks trained from scratch.
All pretraining was completed in less than 10 minutes on a single NVIDIA RTX A4000 GPU. Embedding analyses using dimensionality reduction and distance-based metrics further demonstrated that orientation-based pretraining learned general anatomical features rather than memorizing the training data, enabling robust performance across different medical imaging modalities.
Versatility Across Biomedical Imaging Technologies
The ability to learn modality-independent anatomical features from small datasets creates opportunities across a wide range of biomedical imaging technologies. By emphasizing spatial relationships, structural boundaries, and geometric patterns, the orientation-aware pretraining framework could later be adapted to optical coherence tomography, fundus photography, electron microscopy, and optical phenotyping with minimal additional training.
The approach is particularly valuable in imaging applications where collecting large annotated datasets is expensive, destructive, or limited by sample availability. The lightweight CNN contained approximately 74,000 parameters and used 75.6 MB of peak GPU memory during pretraining, indicating that the approach may be suitable for further investigation in resource-constrained imaging environments.
Accessibility in Diagnostics
This study shows that carefully designed synthetic datasets may reduce the need for large real-world image collections for medical vision models. By learning from structural projection tasks, neural networks develop transferable anatomical representations while reducing computational requirements, energy consumption, and the need for large clinical datasets.
Future work should extend this approach to additional imaging modalities, including computed tomography (CT), ultrasound, and histopathology, to create unified synthetic datasets.
Integrating conditional text-guided diffusion models could further improve control over pathological image generation, enabling more effective pretraining for boundary segmentation and multimodal image registration.
Journal Reference
Turrisi, R., Patanè, G. (2026). Synthetic MRI pretraining for medical imaging classification. Scientific Reports. DOI: 10.1038/s41598-026-63224-5, https://www.nature.com/articles/s41598-026-63224-5.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.