Supervised machine learning based multi-task artificial intelligence classification of retinopathies
View Patent ↗An artificial intelligence (AI) system is disclosed that uses machine learning to classify retinal features contained in OCTA data acquired by a data acquisition system and to predict one or more retinopathies based on the classification of retinal features. The AI system comprise a processor configured to run a classifier model comprising a machine learning algorithm and a memory device that stores the classifier model, OCTA training data and acquired OCT A data. The machine learning algorithm performs a process that trains the classifier model to classify retinal features contained in OCT A training data. The trained machine learning algorithm uses the classifier model to process acquired OCT A data to classify retinal features contained in the acquired OCTA data and to predict, based on the classified retinal features, whether the acquired OCTA data is indicative of at least one of a plurality of retinopathies.
1. An artificial intelligence (AI) system that classifies retinal features and predicts, based on the classified retinal features, one or more retinopathies, the AI system comprising:
a processor configured to perform a machine learning algorithm, the machine learning algorithm performing a process comprising:
training a classifier model to identify, classify and predict retinal features contained in optical coherence tomography angiography (OCTA) training data, where the classifier model utilizes a hierarchical backward elimination technique to identify a plurality of combinations of retinal features used in the training process to train the classifier model to identify and classify retinal features contained in the acquired OCTA data; and
after the classifier model has been trained, using the classifier model to process acquired OCTA data acquired by an image acquisition system to classify retinal features contained in the acquired OCTA data based at least in part upon the plurality of combinations of retinal features and to predict, based on the classified retinal features contained in the acquired OCTA data, whether the acquired OCTA data is indicative of at least one of a plurality of retinopathies; and
a memory device in communication with the processor, the memory device storing the classifier model, OCTA training data and acquired OCTA data.
2. The AI system of claim 1 , wherein when the processor performing the machine learning algorithm uses the classifier model to predict, based on the classified retinal features contained in the acquired OCTA data, whether the acquired OCTA data is indicative of at least one of a plurality of retinopathies, the processor also predicts a respective stage of development of said one or more retinopathies ranging from non-proliferative stages to proliferative stages of development.
3. The AI system of claim 1 , wherein the process of training the classifier model comprises configuring the classifier model to identify the plurality of combinations of retinal features contained in the OCTA training data and to associate each of the plurality of combinations of retinal features with a respective retinopathy.
4. The AI system of claim 3 , wherein the process of training the classifier model comprises configuring the classifier model to identify the plurality of combinations of retinal features contained in the OCTA training data and to associate one or more of the plurality of combinations of retinal features with a respective stage of development of the respective retinopathy.
5. The AI system of claim 3 , wherein the process of using the classifier model comprises identifying at least one of the combinations of retinal features contained in the acquired OCTA data and predicting, based on the identification of said at least one of the combinations of retinal features contained in the acquired OCTA data, whether the acquired OCTA data is indicative of at least one of a plurality of retinopathies and a respective stage of development of the respective retinopathy.
6. The AI system of claim 5 , wherein each combination of retinal features includes two or more of: blood vessel tortuosity (BVT), blood vascular caliber (BVC), vessel perimeter index (VPI), blood vessel density (BVD), foveal avascular zone (FAZ) area (FAZ-A), or FAZ contour irregularity (FAZ-CI).
7. The AI system of claim 6 , wherein the process performed by the machine learning algorithm further comprises:
retraining the classifier model to identify at least one different combination of retinal features based at least in part on the prediction of whether the acquired OCTA data is indicative of at least one of a plurality of retinopathies.
8. The AI system of claim 2 , wherein when the processor performing the machine learning algorithm uses the classifier model to predict, based on the classified retinal features contained in the acquired OCTA data, whether the acquired OCTA data is indicative of at least one of a plurality of retinopathies and a respective stage of development of said one or more retinopathies, the processor performs multi-layer hierarchical classification comprising:
normal versus disease classification;
inter-disease classification; and
stage classification.
9. The AI system of claim 8 , wherein the inter-disease classification includes at least diabetic retinopathy (DR) versus sickle cell retinopathy (SCR) classification.
10. The AI system of claim 9 , wherein the normal versus disease classification includes at least normal versus DR classification and normal versus SCR classification.
11. The AI system of claim 10 , wherein the stage classification includes at least mild, moderate and severe non-proliferative DR (NPDR) stage classification and mild and severe SCR stage classification.
12. The AI system of claim 8 , wherein the machine learning algorithm comprises the hierarchical backward elimination algorithm constructed to identify the combinations of retinal features that achieve a best prediction accuracy, to select the identified combinations of retinal features and to configure the classifier model to perform normal versus disease classification, inter-disease classification; and the stage classification based on combinations of retinal features contained in the acquired OCTA data.
13. The AI system of claim 1 , wherein the OCTA data comprises an OCTA artery-vein map obtained by:
generating an optical coherence tomography (OCT) artery-vein map from a spectrogram dataset;
generating an OCTA vessel map from the spectrogram dataset;
overlaying the OCT artery-vein map and the OCTA vessel map to generate an overlaid map; and
processing the overlaid map by using the OCT artery-vein map to guide artery-vein differentiation in the OCTA vessel map to generate the OCTA artery-vein map.