Dave Steiner, Clinical Research Scientist at Google Health, and Rory Pilgrim, Product Manager at Google Research, discussed the global shortage of medical imaging experts in fields such as radiology, dermatology, and pathology. They highlighted how machine learning (ML) technology can help by powering tools that make image interpretation more accurate and efficient. However, the development of these tools is often hindered by a lack of high-quality data, ML expertise, and computational resources.
One solution they proposed is the use of domain-specific models that utilize deep learning (DL) to create embeddings, which are compressed numerical vectors representing important features in medical images. These embeddings can be used to train models for various tasks within the specialized domain, reducing the need for extensive data, expertise, and computational power.
To address the shortage of access to medical imaging expertise, Google Health and Google Research have released two domain-specific tools for research use: Derm Foundation and Path Foundation. These tools generate specialized embeddings for dermatology and pathology images, respectively, allowing researchers to quickly develop models for their applications.
Path Foundation focuses on pathology images and has been optimized using self-supervised learning (SSL) models. The tool excels in tasks such as cancer detection and grading, outperforming traditional pre-training methods. On the other hand, Derm Foundation specializes in dermatology images and leverages a BiT ResNet-101×3 model trained using contrastive learning. It has shown significant improvements in skin condition classification tasks compared to traditional methods.
While these tools offer promising solutions for medical imaging challenges, further research is needed to evaluate their generalizability across different tasks, patient populations, and settings. Researchers interested in accessing Derm Foundation and Path Foundation can sign up for them using Google Forms to explore their potential in diagnostic tasks, image curation, and biomarker discovery.
Source link
