Back to results
US 10,832,406 B2Granted
Quantitative pathology analysis and diagnosis using neural networks
AssigneePresident and Fellows of Harvard College
InventorKun-Hsing Yu
Life of this patent
Estimating the term…
Abstract
Histological images are preprocessed and classified among pathology states using a neural network. In preprocessing, the digital image is electronically partitioned into a grid of identically sized, overlapping subimages, and images are selected for analysis based on a parameter such as density.
Claims30
A pathology classification system comprising:
- a processor;
- a computer memory for storing a digital image of tissue to be analyzed, the digital image comprising an m×n array of pixels, where m and n are non-zero integers;
- an image preprocessor, executed by the processor and configured to (i) partition the digital image into a grid of identically sized, overlapping subimages thereof, each of the subimages comprising an o×p array of pixels that includes pixels from a subimage adjacent thereto along a single overlap dimension, where o and p are non-zero integers and
- o
- ×
- p
- m
- ×
- n
- ≤
- 0.01
- ,
- (ii) compute an average pixel density for each of the subimages, and (iii) select q subimages with densities higher than the densities of the other subimages, where q is a non-zero integer equal to or greater than 10;
- a neural network, executed by the processor and configured to receive the q selected subimages as feature vectors and classify the selected subimages among pathology states;
- and
- a postprocessing module, executed by the processor and configured to aggregate the subimage classifications and classify tissue among a plurality of types within the pathology states.
13 more elements
A method of classifying tissue, the method comprising the steps of:
- electronically storing a digital image of the tissue, the digital image comprising an m×n array of pixels, where m and n are non-zero integers;
- electronically partitioning the digital image into a grid of identically sized, overlapping subimages thereof, each of the subimages comprising an o×p array of pixels that includes pixels from a subimage adjacent thereto along a single overlap dimension, where o and p are non-zero integers and
- o
- ×
- p
- m
- ×
- n
- ≤
- 0.01
- ;
- computing an average pixel density for each of the subimages;
- selecting q subimages with densities higher than the densities of the other subimages, where q is a non-zero integer equal to or greater than 10;
- using a neural network to classify the selected subimages among pathology states;
- and
- aggregating the subimage classifications and classifying tissue among a plurality of types within the pathology states.
13 more elements
Dependencies are inferred from the claim text.