SAEDNEWS: Brain specialists in Microsoft's Health Futures medical and biotechnology research group have created an AI model that identifies tumors, skin cancer, chest infections, and other health issues.
According to SaedNews, health specialists can use this AI model to detect signs of disease in conventional imaging methods that might otherwise go unnoticed.
As reported by ExtremeTech, typically, ultrasound, CT scans, MRI, and other imaging methods are sent to human specialists for careful examination. They look for unusual cell clusters, slight changes in skin cell shape, fluid indicating lung infections, etc. However, the smallest signs of danger can be missed, and the chance to catch a serious health problem before it progresses can be lost. Some detection stages, such as identifying tumors and mapping their precise shape, are performed by separate groups, leading to some potential discoveries being lost.
The BiomedParse model aims to fill these gaps by helping humans perform the three main stages of medical image processing: finding dangerous items, diagnosing them, and segmenting them. Health Futures says their model outperforms human eyes in nine imaging methods and successfully serves as a "comprehensive" analysis solution for a wide range of specialties.
To build this AI model, researchers used OpenAI's GPT-4 to create a dataset focused on the first two stages of medical image processing, including finding a dangerous item (identifying an anomaly) and diagnosis (imaging what exactly causes the anomaly in the environment). This dataset included over six million images of organs, histology, miscellaneous anomalies, and accompanying textual descriptions from across the body.
These images enable BiomedParse to detect fine details from CT scans, MRIs, X-rays, ultrasounds, pathology slides, fundoscopy (eye exams), dermoscopy, endoscopy, and OCT (optical coherence tomography).
Medical specialists provide a set of images, for example, pathology slides from a patient's liver and a simple text. The AI model then compares the slides with training images related to the disease mentioned in the text. If the model detects signs of the disease, it reports them; if not, it rejects the user's request as invalid.
Health Futures claims that the AI BiomedParse has successfully identified skin cancer, cysts, chest infections caused by COVID-19, and tumors throughout the body. It is unclear whether the group plans to commercialize this technology. Meanwhile, other AI models assist medical specialists in identifying and predicting pancreatic cancer, lung cancer, breast cancer, and other life-threatening conditions.
