Wednesday , June 29 2022

Model AI performs better than Radiologists in the Lung Cancer Hunt



[ad_1]

FavoriteLoadingAdd to favorites

Google aims to serve the model through Cloud Healthcare API

Researchers from Google and Stanford University have developed a machine learning model that shows huge potential to catch lung cancer cases early, performing better than assessments of six radiologists – catching more cases and reducing things. mock positive – when analyzing a dataset of 45,856 chest screens.

The model – identified this week in a paper in Nature Medicine, found five per cent more cancer incidences when reducing pseudo-positive examinations by more than 11 per cent compared with unaided radiologists. . Google aims to commercialize the model through its Cloud Healthcare API, which was released in April.

The research team trained Tensorflow on three datasets available to the public and one proprietary set of Northwestern Medicine for the paper. The result: a model that can not only produce general lung cancer malignancy prediction, but also identify subtle malignant tissue in the lungs (lung nodules).

The model can also include information from previous scans, which are useful in predicting lung cancer risk, as the growth rate of suspicious lung nodules can be a sign of malignancy, Google noted in a related blog.

Google's technical leader, Shravya Shetty, said: “Radiologists usually look through hundreds of 2D images within one CT scan and cancer can be very small and difficult to identify. We created a model that can not only produce lung cancer malignancy prediction (seen in 3D volume) but also identify subtle malignant tissue in the lungs (lung nodules). The model can also include information from previous scans, which is useful in predicting lung cancer risk because the growth rate of suspected lung nodules can be a sign of malignancy. ”

He added: “These initial results are encouraging, but further studies will assess impact and usefulness in clinical practice. We are working with the Google Cloud Healthcare and Life Sciences team to serve this model through Cloud Healthcare API and are in early conversations with partners worldwide to continue research and use additional clinical validation. ”

Google does not release the code used to train the models, saying “it has a large number of dependencies on internal equipment, infrastructure and hardware, and so its release is not practical.”

The researchers added, however: “All experiments and action details are described in sufficient detail in the Methods section to allow independent duplication with non-proprietary libraries.” There are several major components of the work on get in open source stores, they add, pointing to the public datasets, along with Tensorflow machine learning tool and API Object Detection.

Read this: The Google Health Care API Goal is to Go to Dealing With Cooperation Issues with Data Standards that Are Coming To The Exposure

[ad_2]
Source link