A deep network detects diabetic retinopathy
On 13 December 2016 JAMA published a Google convolutional network trained on 128,175 fundus photographs: on two independent sets it detected retinopathy needing referral to an ophthalmologist with an area under the ROC curve of 0.991 and 0.990.
Why it matters
A deep network for a medical image was tested in a clinical journal on two external sets graded by panels of ophthalmologists, and it reached sensitivity and specificity above 90% at once.
The development images were graded 3 to 7 times by 54 ophthalmologists and senior residents; the validation sets were EyePACS-1 (9,963 images, 4,997 patients) and Messidor-2 (1,748 images, 874 patients), each graded by at least seven board-certified ophthalmologists. At the high-specificity point: 90.3% sensitivity and 98.1% specificity on EyePACS-1, 87.0% and 98.5% on Messidor-2; at the high-sensitivity point, 97.5% and 93.4%, 96.1% and 93.9%. The authors say whether the algorithm improves care remains to be studied. The abstract was read; the record claims neither the architecture nor any pretraining.