MRI images have always been a significant challenge in medical image analysis due to their complexity and large data volume. To train large language models (LLMs) for MRI analysis, developers have had to slice the acquired images into 2D images. While this approach is feasible, it limits the model's ability to analyze complex anatomical structures, particularly in intricate cases such as brain tumors, skeletal diseases, or cardiovascular diseases.
Image Source Note: Image generated by AI, image authorized by service provider Midjourney
However, GE Healthcare unveiled the industry's first whole-body 3D MRI foundational model (FM) at this year's AWS re:Invent conference, marking a significant advancement as MRI models can now utilize 3D images of the entire body. This model is built on over 173,000 images from 19,000 studies, and the development team stated that the computational power required for training has been reduced by five times compared to previous methods.
Although GE Healthcare has not yet commercialized this foundational model and it is still in the research phase, early evaluator Massachusetts General Hospital (Mass General Brigham) is set to begin experiments using this model. GE Healthcare's Chief AI Officer, Parry Bhatia, expressed hopes to empower the technical teams in healthcare systems with these models to help them develop research and clinical applications more quickly and cost-effectively.
The introduction of this model will enable real-time analysis of complex 3D MRI data. The GE Healthcare team has a decade of experience in advanced technology, and their flagship product, AIR Recon DL, is a deep learning reconstruction algorithm that helps radiologists obtain clear images more quickly, potentially reducing scan times by up to 50%. Additionally, this 3D MRI model can support image and text searches, linking, and disease segmentation and classification, aiming to provide healthcare professionals with more detailed scan information than ever before.
In terms of data processing, the development team employed a "tune and adapt" strategy, allowing the model to handle various datasets, even when some image data is incomplete, enabling the model to skip missing parts. Furthermore, a semi-supervised student-teacher learning method has been utilized to enhance the model's learning capabilities under limited data conditions.
To address the computational and data challenges encountered during the construction of this complex model, GE Healthcare leveraged Amazon's SageMaker platform, combined with the distributed training capabilities of high-performance GPUs, significantly improving data processing speed and model training efficiency. All of this is done while ensuring compliance with standards such as HIPAA, aiming to provide more personalized healthcare services to patients.
Currently, while the model focuses on the MRI field, developers see significant opportunities for expansion into other medical areas. In the future, this foundational model may provide faster and more efficient solutions in fields such as radiation therapy.
Key Points:
🧠 GE Healthcare has launched the industry's first whole-body 3D MRI foundational model, significantly enhancing image analysis capabilities.
💻 The new model reduces computational resource consumption and improves training efficiency through adjusted data processing strategies.
🚀 This model is expected to expand into other medical fields in the future, aiding in more precise healthcare services.
