Method for object detection using hierarchical deep learning
A hierarchical deep-learning object detection framework provides a method for identifying objects of interest in high-resolution, high pixel count images, wherein the objects of interest comprise a relatively a small pixel count when compared to the overall image. The method uses first deep-learning model to analyze the high pixel count images, in whole or as a patchwork, at a lower resolution to identify objects, and a second deep-learning model to analyze the objects at a higher resolution to classify the objects.
1. A method comprising:
obtaining an image;
analyzing the image, using a trained machine learning model to identify one or more objects of interest in the image, the analysis occurring at a lower resolution than the native resolution of the image;
classifying the identified objects of interest at a higher resolution using a trained machine learning classifier, wherein the machine learning classifier outputs a latent vector for each classified object;
aggregating the latent vectors for a predetermined number of classified objects;
pooling the aggregated latent vectors by calculating a maximum or minimum of the data for each node of a feature map of the machine learning classifier;
performing principal component analysis on the pooled aggregated latent vectors to produce a reduced dimension latent vector;
concatenating the reduced dimension latent vector with a vector representing metadata related to the image; and
performing a binary classification for each image.
2. The method of claim 1 wherein the machine learning model places bounding boxes around the identified objects of interest.
3. The method of claim 2 wherein the machine learning model outputs, with each identified object, a probability that the identified object is an object of interest.
4. The method of claim 1 , further comprising:
separating the image into a plurality of smaller image patches;
wherein analyzing the image comprises analyzing each of the plurality of smaller image patches separately.
5. The method of claim 1 wherein the binary classification is performed using a random decision forest.
6. The method of claim 1 :
wherein the image is a high-resolution whole slide image of organic tissue; and
wherein the objects of interest are structures within the organic tissue.
7. A method comprising:
obtaining an image;
separating the image into a plurality of image patches;
analyzing each image patch at a lower resolution than the native resolution of the image using a trained machine learning model, to identify one or more objects of interest in the image patch;
classifying each identified object of interest in the image patch at a higher resolution, using a trained machine learning classifier;
aggregating latent vectors produced by the machine learning classifier for a predetermined number of classified objects of interest;
pooling the aggregated latent vectors by calculating a maximum or minimum for each node of a feature map of the machine learning classifier; and
performing principal component analysis on the pooled aggregated latent vectors to produce a reduced dimension latent vector.
8. The method of claim 7 further comprising:
concatenating the reduced dimension latent vector with a vector representing metadata related to the image; and
performing a binary classification for the image.
9. The method of claim 8 wherein the binary classification is performed using a random decision forest.
10. The method of claim 7 wherein the trained machine learning model uses a ResNet feature extractor.
11. The method of claim 7 wherein trained machine learning model places a bounding box around each identified object of interest.
12. A system comprising:
an object detection framework trained to detect objects of interest at a lower resolution within a high-resolution image;
a classification framework trained to analyze a high-resolution image of each detected object of interest to classify the object of interest; and
an aggregation framework for aggregating a predetermined number of latent space vectors produced by the classification framework with metadata relating to the high-resolution image to determine a binary classification for the high-resolution image.
13. The system of claim 12 wherein the high-resolution image is broken into a plurality of image patches and wherein the object detection framework analyzes each image patch separately to detect objects of interest within the image patch.
14. The system of claim 12 wherein the object detection framework uses a neural network trained to detect the objects of interest within the high-resolution image.
15. The system of claim 12 wherein the classification framework uses a neural network trained to classify objects of interest identified by the object detection framework.
16. The system of claim 12 :
wherein the high-resolution image is a whole slide image of organic tissue; and
wherein the objects of interest are structures within the organic tissue.
17. The system of claim 12 , further comprising:
normalizing each high-resolution image of detected objects of interest before submitting the high-resolution images to the classification framework.