Automatically identifying regions of interest of an object from horizontal images using a machine learning guided imaging system
A multimodal imaging system and method is capable of taking fundus images, automatically identifying regions of interest (ROIs) of the eye from the fundus images, and performing OCT imaging in the identified ROIs, where the OCT images can provide clinically relevant information for screening purposes. By automatically identifying the ROIs, expert intervention is not required to perform specialized OCT imaging and thus, such imaging and analysis can be provided at more facilities and for more subjects for a lower cost.
1 . An imaging method, comprising:
generating a horizontal image of an object;
automatically identifying a region of interest (ROI) of the object with a non-fully-supervised machine learning system, the non-fully-supervised machine learning system being trained to identify an abnormality of the object from the horizontal image;
generating a second image of the object within the identified ROI, wherein the second image comprises depth information of the object,
wherein the method further comprises obtaining a class activation map from the non-fully-supervised machine learning system,
wherein the non-fully-supervised machine learning system comprises a convolutional neural network (CNN),
wherein each element of the class activation map corresponds to a weighted value of an activation map for a corresponding location of the horizontal image, the activation map being one of a plurality of activation maps produced by a final convolutional layer of the CNN, and
wherein the ROI is identified based on the obtained class activation map.
2 . The method of claim 1 ,
wherein a global activation pooling (GAP) layer is attached to the final convolutional layer of the CNN,
wherein the GAP layer comprises a plurality of values, each of the plurality of values corresponding to a different one of the plurality of activation maps, and
wherein the weight applied to each activation map is determined during a training of the non-fully-supervised machine learning system, and relates a likelihood that the object has the abnormality to a value of the GAP layer at the corresponding location of the horizontal image.
3 . The method of claim 1 , further comprising:
normalizing or scaling values of the class activation map.
4 . The method of claim 1 , further comprising:
displaying the class activation map as a heat map superimposed on the horizontal image.
5 . The method of claim 1 , wherein the horizontal image is a color fundus image; an infrared fundus image; a scanning laser ophthalmoscope (SLO) image; or is derived from 3D optical coherence tomography (OCT) scan data.
6 . The method of claim 1 , wherein the second image is an OCT image.
7 . The method of claim 1 , wherein the horizontal image is derived from 3D optical coherence tomography (OCT) scan data, and the second image is an OCT image generated by extracting a portion of the 3D OCT scan data corresponding to the identified ROI.
8 . The method of claim 1 , wherein the horizontal image is derived from a 3D survey image and the second image has a greater density than the horizontal image.
9 . The method of claim 1 , wherein the abnormality is a retinopathy disorder.
10 . A method of image analysis with a trained non-fully-supervised machine learning system, comprising:
receiving a horizontal image of an object from a subject;
identifying an abnormality of the object as an output of the trained non-fully-supervised machine learning system based on the received horizontal image;
extracting information of the trained non-fully-supervised machine learning used to identify the abnormality;
identifying a region of interest (ROI) within the horizontal image as a region of the horizontal image that contributed to the identification of the abnormality,
wherein the non-fully-supervised machine learning system is trained with a plurality of horizontal images of the object from different subjects to identify the abnormality of the object,
wherein the information of the trained non-fully-supervised machine learning system is extracted by obtaining class activation maps,
wherein the non-fully-supervised machine learning system comprises a convolutional neural network (CNN),
wherein each element of the class activation map corresponds to a weighted value of an activation map for a corresponding location of the horizontal image, the activation map being one of a plurality of activation maps produced by a final convolutional layer of the CNN, and
wherein the ROI is identified based on the obtained class activation map.
11 . The method of claim 10 ,
wherein a global activation pooling (GAP) layer is attached to the final convolutional layer of the CNN,
wherein the GAP layer comprises a plurality of values, each of the plurality of values corresponding to a different one of the plurality of activation maps, and
wherein the weight applied to each activation map is determined during a training of the non-fully-supervised machine learning system, and relates a likelihood that the object has the abnormality to a value of the GAP layer at the corresponding location of the horizontal image.
12 . The method of claim 10 , further comprising: normalizing or scaling values of the class activation map.
13 . The method of claim 10 , further comprising: displaying the class activation map as a heat map superimposed on the horizontal image.
14 . The method of claim 10 , wherein the horizontal image is a color fundus image; an infrared fundus image; a scanning laser ophthalmoscope (SLO) image; or is derived from 3D optical coherence tomography (OCT) scan data.
15 . The method of claim 10 , wherein the abnormality is a retinopathy disorder.