Let's start a new project together
Healthcare

Automated Diabetic Retinopathy Detection Platform based on AI Machine Learning

Challenge
Diabetic Retinopathy is a medical condition in which damage occurs to the retina due to diabetes and might lead to blindness. Proper and timely screening can determine potentially harmful signs of disease being developed and prescribe treatment to prevent this process.

ENBISYS needed a solution for multiple Clients that can be adopted by ophthalmology diagnostics equipment producers as a new feature to their equipment' software. Feature that is easily integrated into devices' circuit and provides unprecedented information in seconds is a great competitive advantage.
Approach
We considered various different approaches that let us build effective and precise AI platform to detect the signs of Diabetic Retinopathy. Our internal model was developed based on open data (EYEPACS, Messidor-1) which contained large number of retinal pictures with light-exposed or darkened spots, thus making it difficult to detect Diabetic Retinopathy. This was the reason we first solved the task of enhancing the quality of initial retinal picture using texture properties, pixel size of vessels relative to the whole size of the picture, and other parameters that allowed to create high quality set of data to train the model.
Based on prepared data set we trained the convolutional neural network and it became capable of grading the picture into several classes (healthy or ill) with certain probability. Also, our neural network defines the probability of disease progression based on several retina pictures of one eye. Our algorithms are capable to distinguish between left and right eye, since they were designed to consider optic disc and macula positions. Also, we trained our neural network on commercial retina images data sets and achieved accuracy of 96,5%.
Since such feature is supposed to be utilized in decision support systems we implemented the option to highlight particular spots on retina images which might be of concern for ophthalmologist. Those are the suspicious spots that neural network detects during screening. Additionally, ophthalmologists can analyze vessels mask which is derived after initial image processing when neural network screens the vessels. It is considered that such image might contain the signs of diabetic retinopathy.
Solution
We've built platform which outperforms that of other ophthalmologic companies which implement Machine Learning technology to disease diagnostics. The specificity we achieved is better and comprises now 91% and accuracy is 96,5%. Response time is 4 seconds.

Machine Learning in this Solution is used to perform the following tasks:
  • Define if retinal pictures are gradable/non-gradable
  • Identify left/right eye
  • Detect the stage of diabetic retinopathy, 1 to 4
During last year the feature was tested in several eye care centers and laboratories in Europe and was greatly appraised by ophthalmologists who performed the valuations.

Since the amount of data collected and released for scientific medical purposes increases constantly, we plan to support our partners who will implement the feature with newest updates on algorithms work and providing all required maintenance.

Automated Diabetic Retinopathy Detection Platform based on AI Machine Learning

Challenge
Diabetic Retinopathy is a medical condition in which damage occurs to the retina due to diabetes and might lead to blindness. Proper and timely screening can determine potentially harmful signs of disease being developed and prescribe treatment to prevent this process.

ENBISYS needed a solution for multiple Clients that can be adopted by ophthalmology diagnostics equipment producers as a new feature to their equipment' software. Feature that is easily integrated into devices' circuit and provides unprecedented information in seconds is a great competitive advantage.
Approach
We considered various different approaches that let us build effective and precise AI platform to detect the signs of Diabetic Retinopathy. Our internal model was developed based on open data (EYEPACS, Messidor-1) which contained large number of retinal pictures with light-exposed or darkened spots, thus making it difficult to detect Diabetic Retinopathy. This was the reason we first solved the task of enhancing the quality of initial retinal picture using texture properties, pixel size of vessels relative to the whole size of the picture, and other parameters that allowed to create high quality set of data to train the model.
Based on prepared data set we trained the convolutional neural network and it became capable of grading the picture into several classes (healthy or ill) with certain probability. Also, our neural network defines the probability of disease progression based on several retina pictures of one eye. Our algorithms are capable to distinguish between left and right eye, since they were designed to consider optic disc and macula positions. Also, we trained our neural network on commercial retina images data sets and achieved accuracy of 96,5%.
Since such feature is supposed to be utilized in decision support systems we implemented the option to highlight particular spots on retina images which might be of concern for ophthalmologist. Those are the suspicious spots that neural network detects during screening. Additionally, ophthalmologists can analyze vessels mask which is derived after initial image processing when neural network screens the vessels. It is considered that such image might contain the signs of diabetic retinopathy.
Solution
We've built platform which outperforms that of other ophthalmologic companies which implement Machine Learning technology to disease diagnostics. The specificity we achieved is better and comprises now 91% and accuracy is 96,5%. Response time is 4 seconds.

Machine Learning in this Solution is used to perform the following tasks:
  • Define if retinal pictures are gradable/non-gradable
  • Identify left/right eye
  • Detect the stage of diabetic retinopathy, 1 to 4
During last year the feature was tested in several eye care centers and laboratories in Europe and was greatly appraised by ophthalmologists who performed the valuations.

Since the amount of data collected and released for scientific medical purposes increases constantly, we plan to support our partners who will implement the feature with newest updates on algorithms work and providing all required maintenance.

Automated Diabetic Retinopathy Detection Platform based on AI Machine Learning

Challenge
Diabetic Retinopathy is a medical condition in which damage occurs to the retina due to diabetes and might lead to blindness. Proper and timely screening can determine potentially harmful signs of disease being developed and prescribe treatment to prevent this process.

ENBISYS needed a solution for multiple Clients that can be adopted by ophthalmology diagnostics equipment producers as a new feature to their equipment' software. Feature that is easily integrated into devices' circuit and provides unprecedented information in seconds is a great competitive advantage.
Approach
We considered various different approaches that let us build effective and precise AI platform to detect the signs of Diabetic Retinopathy. Our internal model was developed based on open data (EYEPACS, Messidor-1) which contained large number of retinal pictures with light-exposed or darkened spots, thus making it difficult to detect Diabetic Retinopathy. This was the reason we first solved the task of enhancing the quality of initial retinal picture using texture properties, pixel size of vessels relative to the whole size of the picture, and other parameters that allowed to create high quality set of data to train the model.
Based on prepared data set we trained the convolutional neural network and it became capable of grading the picture into several classes (healthy or ill) with certain probability. Also, our neural network defines the probability of disease progression based on several retina pictures of one eye. Our algorithms are capable to distinguish between left and right eye, since they were designed to consider optic disc and macula positions. Also, we trained our neural network on commercial retina images data sets and achieved accuracy of 96,5%.
Since such feature is supposed to be utilized in decision support systems we implemented the option to highlight particular spots on retina images which might be of concern for ophthalmologist. Those are the suspicious spots that neural network detects during screening. Additionally, ophthalmologists can analyze vessels mask which is derived after initial image processing when neural network screens the vessels. It is considered that such image might contain the signs of diabetic retinopathy.
Solution
We've built platform which outperforms that of other ophthalmologic companies which implement Machine Learning technology to disease diagnostics. The specificity we achieved is better and comprises now 91% and accuracy is 96,5%. Response time is 4 seconds.

Machine Learning in this Solution is used to perform the following tasks:
  • Define if retinal pictures are gradable/non-gradable
  • Identify left/right eye
  • Detect the stage of diabetic retinopathy, 1 to 4
During last year the feature was tested in several eye care centers and laboratories in Europe and was greatly appraised by ophthalmologists who performed the valuations.

Since the amount of data collected and released for scientific medical purposes increases constantly, we plan to support our partners who will implement the feature with newest updates on algorithms work and providing all required maintenance.

Let's discuss your Case!
Let's discuss your Case!
Let's discuss your Case!