System and method for detection and classification of findings in images
A system for detection and classification of findings in an image, comprising at least one hardware processor configured to: receive the image; process the image by a plurality of convolutional and pooling layers of a neural network to produce a plurality of feature maps; process one of the feature maps by some of the layers and another plurality of layers to produce a plurality of region proposals; produce a plurality of region of interest (ROI) pools by using a plurality of pooling layers to downsample the plurality of region proposals with each one of the plurality of feature maps; process the plurality of ROI pools by at least one concatenation layer to produce a combined ROI pool; process the combined ROI pool by a classification network comprising some other of the convolutional and pooling layers to produce one or more classifications; and output the one or more classifications.
1. A system for detection and classification of findings in an image, comprising:
at least one hardware processor configured to:
receive the image;
process said image by a first plurality of convolutional and pooling layers of a neural network to produce a plurality of feature maps, each one of said plurality of feature maps comprising a plurality of location coordinates related to said image;
process one of said plurality of feature maps by some of said first plurality of convolutional and pooling layers, a second plurality of convolutional and pooling layers and a first plurality of fully connected layers of said neural network to produce a plurality of region proposals;
produce a plurality of region of interest (ROI) pools by using a third plurality of pooling layers of said neural network to downsample said plurality of region proposals with each one of said plurality of feature maps;
process said plurality of ROI pools by at least one pool concatenation layer of said neural network to produce a combined ROI pool;
process said combined ROI pool by a classification network comprising some other of said first plurality of convolutional and pooling layers to produce one or more classifications; and
output said one or more classifications.
2. The system of claim 1 , wherein said first plurality of convolutional and pooling layers comprises a plurality of layer blocks, each layer block comprising a sequence of convolutional and pooling layers;
wherein said plurality of layer blocks are connected in a topology selected from a group consisting of a linear graph and a directed acyclic graph.
3. The system of claim 2 , wherein said plurality of layer blocks of said first plurality of convolutional and pooling layers are connected linearly;
wherein said first plurality of convolutional and pooling layers comprises a first sequence of layers consisting of: an input layer, followed by a first block consisting of a first sequence of convolutional and rectified linear unit layers, followed by a first pooling layer, followed by a second block consisting of a second sequence of convolutional and rectified linear unit layers, followed by a second pooling layer, followed by a third block consisting of a third sequence of convolutional and rectified linear unit layers, followed by a third pooling layer, followed by a fourth block consisting of a fourth sequence of convolutional and rectified linear unit layers, followed by a fourth pooling layer, and followed by a fifth block consisting of a fifth sequence of convolutional and rectified linear unit layers;
wherein an output of said third sequence of convolutional and rectified linear unit layers is connected to a first normalized convolutional layer;
wherein an output of said fourth sequence of convolutional and rectified linear unit layers is connected to a second normalized convolutional layer;
wherein a first of said plurality of feature maps is produced by said first normalized convolutional layer;
wherein a second of said plurality of feature maps is produced by said second normalized convolutional layer;
wherein a third of said plurality of feature maps is produced by said fifth sequence of convolutional and rectified linear unit layers; and
wherein said one of said plurality of feature maps is said third of said plurality of feature maps.
4. The system of claim 1 , wherein each one of said plurality of region proposals comprises a boundary box, an object likelihood score and a non-object likelihood score.
5. The system of claim 1 , wherein said third plurality of layers comprises a second sequence of layers consisting of: an ROI pooling layer, followed by a fifth pooling layer, followed by a fully connected layer and followed by a pooling reshaping layer.
6. The system of claim 1 , wherein said classification network further comprises some of said first plurality of fully connected layers, a fourth plurality of fully connected layers, at least one bounding box regressor layer and one or more loss layers.
7. The system of claim 1 , wherein said image is a medical image.
8. The system of claim 1 , wherein said image is a medical image; and
wherein said at least one hardware processor is further configured to:
execute an electronic record management (ERM) system for storing a plurality of medical records comprising a plurality of medical images;
store said image in said ERM system;
retrieve said medical image from said ERM system before processing said image by said first plurality of convolutional and pooling layers; and
store said one or more classifications in said ERM system.
9. The system of claim 8 , wherein each one of said one or more classifications comprises a bounding box, a class and a probability score of said class at said bounding box; and
wherein said at least one hardware processor is further configured to issue an alert upon determining said class of at least one of said one or more classifications is a member of a predetermined set of abnormal classes and said probability score of said at least one classification is greater than a threshold score value.
10. The system of claim 8 , wherein said ERM system stores said image in a format compliant with Digital Imaging and Communications in Medicine (DICOM).
11. The system of claim 1 , wherein said image is received by said at least one hardware processor via a network communications protocol compliant with DICOM.
12. A method for detection and classification of findings in an image, comprising:
receiving the image;
processing said image by a first plurality of convolutional and pooling layers of a neural network to produce a plurality of feature maps, each one of said plurality of feature maps comprising a plurality of location coordinates related to said image;
processing one of said plurality of feature maps by some of said first plurality of convolutional and pooling layers, a second plurality of convolutional and pooling layers and a first plurality of fully connected layers of said neural network to produce a plurality of region proposals;
producing a plurality of region of interest (ROI) pools by using a third plurality of pooling layers of said neural network to downsample said plurality of region proposals with each one of said plurality of feature maps;
processing said plurality of ROI pools by at least one pool concatenation layer of said neural network to produce a combined ROI pool;
processing said combined ROI pool by a classification network comprising some other of said first plurality of convolutional and pooling layers to produce one or more classifications; and
outputting said one or more classifications.
13. The method of claim 12 , wherein said image is captured by a mammography of a breast; and
wherein said method further comprising:
after receiving said image:
removing background and pectoral muscle from said image to produce a breast tissue image;
receiving a plurality of other images;
producing a plurality of visual attributes by:
extracting Fibro-glandular tissue from said breast tissue image and said plurality of other images;
extracting at least one saliency map of said breast tissue image and said plurality of other images; and
processing said breast tissue image and said plurality of other images to produce a plurality of vessel identifications; and
dividing said breast tissue image into a plurality of overlapping sub-images; and
after producing said one or more classifications and before outputting said one or more classifications:
projecting one or more bounding boxes of said one or more classifications on said image.
14. The method of claim 12 , further comprising after producing said one or more classifications and before outputting said one or more classifications:
clustering said one or more classifications according to redundant and overlapping bounding boxes to produce a one or more clustered classifications; and
projecting one or more bounding boxes of said clustered classifications on said image.
15. A system for detection and classification of findings in an image, comprising:
at least one hardware processor configured to:
receive the image;
receive a plurality of values of a plurality of attributes related to said image;
process said plurality of values by a first plurality of fully connected layers of a neural network to produce a plurality of processed values;
integrate said plurality of processed values into a plurality of layers of said neural network;
process said image by a first plurality of convolutional and pooling layers of a neural network to produce at least one feature map, each one of said at least one feature map comprising a plurality of location coordinates related to said image;
process one of said at least one feature map by some of said first plurality of convolutional and pooling layers, a second plurality of convolutional and pooling layers and a second plurality of fully connected layers of said neural network to produce a plurality of region proposals;
produce a plurality of region of interest (ROI) pools by using a third plurality of pooling layers of said neural network to downsample said plurality of region proposals with each one of said at least one feature map;
process said plurality of ROI pools by at least one pool concatenation layer to produce a combined ROI pool;
process said combined ROI pool by a classification network comprising some other of said first plurality of convolutional and pooling layers to produce one or more classifications; and
output said one or more classifications.
16. The system of claim 15 , wherein said plurality of processed values is integrated into said plurality of layers by processing said plurality of processed values and said plurality of ROI pools by said least one pool concatenation layer to produce said combined ROI pool.
17. The system of claim 15 , wherein said plurality of processed values is integrated into said plurality of layers by using said plurality of processed values as modulators for a plurality of convolutional kernels of said first plurality of convolutional and pooling layers before processing said image to produce said plurality of feature maps.
18. The system of claim 15 , wherein said image is a medical image; and
wherein said plurality of attributes are selected from a group consisting of: a clinical record of a patient, a positive mask corresponding with an anatomical feature, a negative mask corresponding with an anatomical feature, a computed visual attribute, a density, a symmetry property, an additional view of said image, and a feature detected in another image.
19. The system of claim 18 , wherein said medical image is captured by a mammography of a breast;
wherein said positive mask corresponds to Axilla tissue or Fibroglandular tissue; and
wherein said negative mask corresponds to a blood vessel.
20. The system of claim 18 , wherein said medical image is captured by a method selected from a group consisting of: a computerized tomography of a brain, a magnetic resonance imaging of a brain, an ultrasound of a brain, a positron emission tomography of a brain and an X-ray radiography of a brain; and
wherein said positive mask corresponds to white matter.