Application of deep learning--detection of diabetic retinopathy

Recently, Google published in the Journal of the American Medical Council titled "Development and Validation of a deep learning algorithm for Detection of diabetic retinopathy in Reti NAL Fundus Photographs "is a deep learning algorithm that Google researchers have put forward to explain the signs of diabetic retinopathy in the retina angiography, helping doctors overcome resource shortages and make more professional diagnoses for more patients.
　　paper: Development and validation of deep learning algorithms for detecting diabetic retinopathy in retinal fundus photos
Development and Validation of a deep learning algorithm for Detection of diabetic retinopathy in retinal Fundus Photograp Hs

Summary:

　　 Importance:Deep learning is a series of methods that allow an algorithm to self-program by learning a large number of samples that show the expected behavior, which allows us to no longer need certain definite rules. The application of these methods in medical imaging needs further evaluation and verification.
　　 Target:To apply deep learning to create an algorithm that can automatically detect diabetic retinopathy and diabetic macular edema through retinal fundus photographs.
　　 Design and configuration:We used a neural network model called deep convolutional neural Networks optimized for image classification, which was trained with data sets of 128175 retinal images, each of which was evaluated 3 to 7 times for diabetic retinopathy, diabetic macular edema, and image grade levels. The resulting algorithm was validated using two separate datasets from January 2016 and February, each of which is a reference to the standard of a 7 or 8-person American certified eye doctor.
　　 main results and measures:This is used to detect the onset of diabetic retinopathy (rdr/referable diabetic retinopathy, moderate and worse diabetic retinopathy), The sensitivity of the algorithm (sensitivity) and specificity (specificity), which can be onset of diabetic macular edema or both, is based on a reference standard for most decisions in the eye panel. The algorithm is evaluated on the 2 operating points selected for two development sets, one for high specificity and the other for higher sensitivity.
* * Results: The **eyepacs-1 DataSet contains 9963 images from 4,997 patients (average age 54.4 years), 62.2% of which are women, and generally rdr,683/8878 fully-scalable images (7.8%). The Messidor-2 DataSet has 1748, 42.6% females, and is generally rdr,254/1745 fully scalable (14.6%), from 874 patients (average age 57.6 years). In order to detect RDR, the algorithm has an area of 0.991 (95% ci, 0.988-0.993) under the EyePACS-1 (ROC curve) on the test-bed, and the area under the ROC curve on Messidor-2 is 0.990 (95% CI, 0.98 6-0.995). Using the first specific operational pointcut (operating cut point), for EyePACS-1, the sensitivity is 90.3% (95% ci, 87.5%-92.7%), the specificity is 98.1% (95% ci, 97.8%-98.5%). For Messidor-2, the sensitivity is 87.0% (95% ci, 81.1%-91.0%), specificity is 98.5% (95% CI, 97.7%-99.1%). Using the second high-sensitivity operating point of the development set, for EyePACS-1, the sensitivity is 97.5% and the specificity is 93.4%; for Messidor-2, the sensitivity is 96.1% and the specificity is 93.9%.
　　 conclusions and Related:In the evaluation of this adult diabetic retinal fundus photo, the algorithm based on deep machine learning has high sensitivity and specificity in detecting suspicious diabetic retinopathy. This will confirm the applicability of this algorithm in clinical practice and determine whether the algorithm can be used to improve treatment and diagnostic outcomes compared to current ophthalmic assessments.
　　
One of the most common ways to detect diabetic ophthalmopathy is to have a specialist examine the image at the back of the eye (Figure 1) before assessing the presence and severity of the disease. The severity of the disease is determined by the types of lesions (such as micro-aneurysms, hemorrhage, hard exudate, etc.) that indicate bleeding and fluid exudation in the eyes. However, the interpretation of these photos requires specialized training, and in many parts of the world, there are not enough qualified evaluators to screen everyone who has the risk of a disease.

Figure 1: A sample of retinal fundus photos taken to screen DR. The image on the left is a healthy retina (A), while the image on the right is the retina (B) that can cause diabetic retinopathy, and can be seen in the presence of bleeding (red dots).

Paper Download: http://jamanetwork.com/journals/jama/fullarticle/2588763

Application of deep learning--detection of diabetic retinopathy