System and method for automated funduscopic image analysis

Abstract

A system and method of classifying images of pathology. An image is received, normalized, and segmented normalizing the image; into a plurality of regions; A disease vector is automatically determining for the plurality of regions with at least one classifier comprising a neural network. Each of the respective plurality of regions is automatically annotated, based on the determined disease vectors. The received image is automatically graded based on at least the annotations. The neural network is trained based on at least an expert annotation of respective regions of images, according to at least one objective classification criterion. The images may be eye images, vascular images, or funduscopic images. The disease may be a vascular disease, vasculopathy, or diabetic retinopathy, for example.

What is claimed is: 
     
         1 . A method of classifying a funduscopic image, comprising:
 normalizing the funduscopic image;   segmenting the funduscopic image into a plurality of regular regions;   automatically determining a quality for each respective regular region, representing at least one quantitative quality characteristic of the regular region;   automatically determining a probability of retinopathy for each respective regular region;   automatically classifying the funduscopic image with respect to a presence of retinopathy; and   outputting an indication of the presence of retinopathy in the funduscopic image.   
     
     
         2 . The method according to  claim 1 , wherein the automatic determination of the probability of retinopathy for each respective regular region is determined by a neural network. 
     
     
         3 . The method according to  claim 2 , wherein the neural network is a deep neural network trained based on expert annotation of regions of funduscopic images for retinopathy. 
     
     
         4 . The method according to  claim 1 , wherein the automatically determined quality for each respective regular region is a vector determined by a neural network. 
     
     
         5 . The method according to  claim 1 , wherein the automatically determined quality for each respective regular region comprises an assessment of at least one of a brightness, a focus, lens flare, a contrast, and a color. 
     
     
         6 . The method according to  claim 1 , wherein the automatically determined quality for each respective regular region comprises an assessment of at least two of a brightness, a focus, a lens flare, a contrast, and a color. 
     
     
         7 . The method according to  claim 1 , wherein the regions of poor quality do not contribute to a classification probability of the funduscopic image. 
     
     
         8 . The method according to  claim 1 , wherein a classification probability of a respective regular region of the funduscopic image is dependent on a multivalued vector of the quality of the regular region and the probability of retinopathy for each respective regular region. 
     
     
         9 . The method according to  claim 1 , wherein the probability of retinopathy for each respective regular region is determined by at least one of a multi-class support vector machine classifier, and a Gradient Boosting Classifier. 
     
     
         10 . The method according to  claim 1 , wherein said outputting the indication of the presence of retinopathy in the funduscopic image comprises outputting a 2D probability map. 
     
     
         11 . The method according to  claim 1 , said outputting the indication of the presence of retinopathy in the funduscopic image comprises outputting a degree of diabetic retinopathy with respect to at least three different grades. 
     
     
         12 . A method of classifying a funduscopic image, comprising:
 normalizing and segmenting a funduscopic image into a plurality of regular regions;   automatically determining a probability of retinopathy for each respective regular region;   generating a 2D probability map for a probability of presence of indicia of retinopathy in respective regular regions, each having dynamic range of probability; and   classifying the 2D probability map with respect to a presence of retinopathy in the funduscopic image.   
     
     
         13 . The method according to  claim 12 , wherein the automatic determination of the probability of retinopathy for each respective regular region is determined by a deep neural network, trained based on expert annotation of regions of funduscopic images for retinopathy. 
     
     
         14 . The method according to  claim 12 , further comprising automatically determining a multivalued, multiparameter quality vector for each respective regular region. 
     
     
         15 . The method according to  claim 14 , wherein the automatically determined quality vector for each respective regular region comprises an assessment of at least one of a brightness, a focus, lens flare, a contrast, and a color. 
     
     
         16 . The method according to  claim 12 , wherein a classification probability of a respective regular region of the funduscopic image is dependent on a multivalued vector of the quality of the regular region and the probability of retinopathy for each respective regular region. 
     
     
         17 . The method according to  claim 12 , wherein the probability of retinopathy for each respective regular region is determined by at least one of a multi-class support vector machine classifier, and a Gradient Boosting Classifier. 
     
     
         18 . A method for classifying diabetic retinopathy in a funduscopic image, comprising:
 converting a plurality of diabetic retinopathy lesion classification vectors for respective regions of a funduscopic image into a plurality of 2D probability maps, wherein a number of 2D probability maps corresponds to a number of diabetic retinopathy classification labels; and   automatically classifying the plurality of 2D probability maps with a convolutional neural network, to produce a retinal funduscopic image-level diabetic retinopathy classification vector, wherein the convolutional neural network is trained with a plurality of funduscopic images expert annotated for diabetic retinopathy.   
     
     
         19 . The method according to  claim 18 , further comprising automatically determining a quality for each respective region, representing at least one quantitative quality characteristic of the regular region, wherein an application of a respective region in the retinal funduscopic image-level diabetic retinopathy classification vector is dependent on a corresponding quality for the respective region. 
     
     
         20 . The method according to  claim 18 , further comprising:
 normalizing the funduscopic image;   segmenting the funduscopic image into a plurality of regular regions;   automatically determining a quality for each respective regular region, representing at least one quantitative quality characteristic of the regular region; and   outputting the retinal funduscopic image-level diabetic retinopathy classification vector.