Imaging segmentation using multi-scale machine learning approach

Patent metadata

Field	Value
Publication number	US-9684967-B2
Application number	US-201514921579-A
Country	US
Kind code	B2
Filing date	Oct 23, 2015
Priority date	Oct 23, 2015
Publication date	Jun 20, 2017
Grant date	Jun 20, 2017

Abstract

A robust segmentation technique based on multi-layer classification technique to identify the lesion boundary is described. The inventors have discovered a technique based on training several classifiers such that to classify each pixel as lesion versus normal Each classifier is trained on a specific range of image resolutions. Then, for a new test image, the trained classifiers are applied on the image. Then by fusing the prediction results in pixel level a probability map is generated. In the next step, a thresholding method is applied to convert the probability map to a binary mask, which determines a mole border.

First claim

What is claimed is:

1. A computer-based method for segmenting an image to identify a border using supervised machine learning, the method comprising:
a training phase including
a) accessing a training image set with at least one training image, wherein each training image includes border coordinates;
b) generating a discrete number of training sub-sections of the training image within a given set of numbers;
c) labeling each of the training sub-sections into one of a first category and a second category;
d) extracting low level features for the first category into a first classifier;
e) extracting low level features for the second category into a second classifier;
f) training a binary classifier on whether it is one of a boundary type or non-boundary type;

repeating a through f until each discrete number of sub-sections in the given set of numbers has been generated for each image in the training image set;
a segmentation phase including
g) accessing an image set with at least one image;
h) selecting a discrete number of image sub-sections of the image within the given set of numbers;
i) applying a trained classifier to each image sub-section for the discrete number of image sub-sections, the classifier representing a probability that the image sub-section is the boundary type;
j) outputting of classifiers and generating a probability map then overlaying the probability map on other probability maps generated from other classifiers containing the image sub-section;
k) in response to the overlaid trained classifier being above a settable threshold, identifying the image sub-section as a border segment; and

repeating g through k until each discrete number of sub-sections in the given set of numbers has been generated for each image in the image set.

2. The computer-based method of claim 1 , wherein the boundary type is a lesion and the non-boundary type is skin.

3. The computer-based method of claim 1 , wherein the extracting low level features for the first category into a first classifier includes using one of a color correlogram function, a wavelet texture function, color histogram function, a color wavelet function, an edge histogram function, GIST descriptor function, a multi scale local binary pattern function, or a combination thereof.

4. The computer-based method of claim 3 , wherein the extracting low level features for the second category into a second classifier includes using one of a color correlogram function, a wavelet texture function, color histogram function, a color wavelet function, an edge histogram function, GIST descriptor function, a multi scale local binary pattern function, or a combination thereof.

5. The computer-based method of claim 1 , wherein the generating a discrete number of training sub-sections of an image in the training image set includes generating each of the sub-sections as a superpixel.

6. The computer-based method of claim 1 , wherein the applying a trained classifier to each image sub-section for the discrete number of image sub-sections, includes applying an average weight to each classifier.

7. The computer-based method of claim 1 , wherein the applying a trained classifier to each image sub-section for the discrete number of image sub-sections, includes applying a weight to each classifier using an inference engine and a size of the sub-sections, confidence of classifier, and information regarding neighboring sub-sections.

8. The computer-based method of claim 1 , wherein the applying a trained classifier to each image sub-section for the discrete number of image sub-sections, includes applying a weight to each classifier using an inference engine and dermoscopic features of central dots, scattered dots, peripheral globules, eccentric globules, diffuse blue/grey dots, central globules, uniform globules, cobblestone pattern, or a combination thereof.

9. The computer-based method of claim 8 , wherein the dermoscopic features include of central dots, scattered dots, peripheral globules, eccentric globules, diffuse blue/grey dots, central globules, uniform globules, cobblestone pattern, or a combination thereof.

10. A system for segmenting an image to identify a border using supervised machine learning, the system comprising:
a memory;
a processor communicatively coupled to the memory, where the processor is configured to perform
a training phase including
a) accessing a training image set with at least one training image, wherein each training image includes border coordinates;
b) generating a discrete number of training sub-sections of the training image within a given set of numbers;
c) labeling each of the training sub-sections into one of a first category and a second category;
d) extracting low level features for the first category into a first classifier;
e) extracting low level features for the second category into a second classifier;
f) training a binary classifier on whether it is one of a boundary type or non-boundary type;

repeating a through f until each discrete number of sub-sections in the given set of numbers has been generated for each image in the training image set;
a segmentation phase including
g) accessing an image set with at least one image;
h) selecting a discrete number of image sub-sections of the image within the given set of numbers;
i) applying a trained classifier to each image sub-section for the discrete number of image sub-sections, the classifier representing a probability that the image sub-section is the boundary type;
j) outputting of classifiers and generating a probability map then overlaying the probability map on other probability maps generated from other classifiers containing the image sub-section;
k) in response to the overlaid trained classifier being above a settable threshold, identifying the image sub-section as a border segment; and

repeating g through k until each discrete number of sub-sections in the given set of numbers has been generated for each image in the image set.

11. The system of claim 10 , wherein the boundary type is a lesion and the non-boundary type is skin.

12. The system of claim 10 , wherein the extracting low level features for the first category into a first classifier includes using one of a color correlogram function, a wavelet texture function, color histogram function, a color wavelet function, an edge histogram function, GIST descriptor function, a multi scale local binary pattern function, or a combination thereof.

13. The system of claim 12 , wherein the extracting low level features for the second category into a second classifier includes using one of a color correlogram function, a wavelet texture function, color histogram function, a color wavelet function, an edge histogram function, GIST descriptor function, a multi scale local binary pattern function, or a combination thereof.

14. The system of claim 10 , wherein the generating a discrete number of training sub-sections of an image in the training image set includes generating each of the sub-sections as a superpixel.

15. The system of claim 10 , wherein the applying a trained classifier to each image sub-section for the discrete number of image sub-sections, includes applying an average weight to each classifier.

16. The system of claim 10 , wherein the applying a trained classifier to each image sub-section for the discrete number of image sub-sections, includes applying a weight to each classifier using an inference engine and a size of the sub-sections, confidence of classifier, and information regarding neighboring sub-sections.

17. The system of claim

Assignees

• IBM

Inventors

• Abedini Mani
• Chakravorty Rajib
• Garnavi Rahil
• HAYAT Munawar

Classifications

• G06V30/19147

  Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/44 Primary

  Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components · CPC title

• G06T7/13

  Edge detection · CPC title

• G06F18/217

  Validation; Performance evaluation; Active pattern learning techniques · CPC title

• G06F18/24

  Classification techniques · CPC title