AI companies are increasingly using medical datasets from digital photography. This technology is ubiquitous and easy to use, with the potential to break through geographical obstacles to care.
We’ve worked with clients to structure image data through our annotation services - both classifying these images and segmenting regions of interest. Examples include classification of skin lesions and segmenting retinal slit lamp videos.
Ocular Technologies partnered with Centaur Labs to bring their telehealth eye exam tool to life. On a dataset of anterior segment videos, the Centaur network annotated pupil and retinal slits in each clip. The Centaur team quickly delivered over 25,000 high quality opinions on 6,000 images of eyes, powering the first version of Ocular's tool.
ISIC and Centaur Labs designed a medical image classification task for ISIC’s archive of over 11,000 dermoscopic skin lesion images. Across three labeling tasks including malignancy classification and lesion recognition, Centaur collected over half a million opinions from its network of experts. Recent research published in the Lancet assessed how accurate board-certified experts with 10+ years experience were at this multi-class skin-lesion labeling task. The performance of the Centaur’s network without any experience exclusion criteria was more accurate (78.1%, p<0.001, vs 74.7%) than those physicians.
Auggi's gut health app transforms the lives of patients living with chronic gastrointestinal conditions. Centaur Labs classified 15,000 user-submitted stool images captured by their smartphones into Bristol stool scores. Centaur's data labels improved Auggi's model to an overall accuracy of 94%.