Loading presentation...

Present Remotely

Send the link below via email or IM

Copy

Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.

DeleteCancel

Make your likes visible on Facebook?

Connect your Facebook account to Prezi and let your likes appear on your timeline.
You can change this under Settings & Account at any time.

No, thanks

2-SiMDoM: A 2 Sieve Model for Detection of Mitosis in Multispectral Breast Cancer Imagery

No description
by

Ardhendu Tripathi

on 9 April 2015

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of 2-SiMDoM: A 2 Sieve Model for Detection of Mitosis in Multispectral Breast Cancer Imagery

2-SiMDoM: A 2 Sieve Model for Detection of Mitosis in Multispectral Breast Cancer Imagery
Ardhendu Tripathi*, Atin Mathur*, Mohit Daga*, Manohar Kuse^, Oscar C. Au^
The LNM Institute of Information Technology, Jaipur, India*,
Hong Kong University of Science and Technology, Hong Kong^
17 September, 2013
at
20th IEEE International Conference on Image Processing, Melbourne, Australia
Relevance of Mitotic Count in Breast Cancer Prognosis
"All change is a miracle to contemplate, but it is a miracle which is taking place every instant"
-Henry David Thoreau
Challenges
a) Mitosis
b) Mitosis
c) Non Mitosis
d) Non Mitosis
Dataset Description
MITOS Dataset (ICPR 2012 Mitosis Detection Dataset)
48 slides
All slides were H&E stained.
Multispectral imaging
Images were captured over 10 visible bands.
For each band digitization was performed over 17 different focus planes resulting in stacks.
Each focus plane was separated by 500 nm.
System Design
Preprocessing and Segmentation
Feature Extraction and Selection
Handling Imbalanced Dataset
2-Sieve Classification
Schematic for Preprocessing and Segmentation
Entropy based stack selection
Image entropy is defined as follows:
The best quality stack
was selected such that:
The most informative stack for each band was used for all future computations.
Cell Segmentation
The region based active contour model as proposed by Chan et. al. was employed.
This technique was effective here due to the difference in the average pixel intensity levels inside and outside the cell.
To obtain the seed points, the histogram equalized higher visible contrast band BB07 was used.

Feature Extraction
Training was based on texture features
5 GLCM features were computed in the wavelet domain (4 components - LL, LH, HL, HH).
3 level decomposition chosen based on the minimum entropy algorithm.
Total number of features = 5 X 4 X 3 X 10 bands = 600
9 additional GLEM features were extracted resulting in 90 features (9 X 10 bands).
Optimal feature set selection
Feature set selection to get rid of statistically irrelevant set of features.
Our aim - to find out the set of features which best discriminates the mitosis from non-mitosis.
Supervised Dimensionality Selection based on Neighborhood Examination (SDSNE)
Selects the set of features which minimizes the error rate i.e. the number of misclassified mitosis on the basis of the neighborhood majority rule in the feature space.
690 initial features reduced to 43
Data Imbalance
Imbalanced dataset causes inaccurate training of the classifier.
Biased towards the majority class.
Imbalanced dataset dealt by:-
1) Oversampling of Mitotic instances
2) Data Cleaning
Oversampling of Mitotic instances
To balance the dataset oversampling of the mitotic instances was done based upon:
a) Perturbations - image
b) SMOTE (Synthetic Minority Oversampling Technique) - feature space
Data Cleaning
Removal of class label noise and borderline examples.
Instances participating in Tomek Links were eliminated.
Tomek links - points that are each others closest neighbors but do not share the same label.
The 2-Sieve Model
First Level Sieve
SVM with radial basis kernel.
Trained on 33 HPF's and tested on 15 HPF's.
Sensitivity = 85.29% and PPV = 59.58%.
Second Level Sieve
Textural differences around a mitotic and a non-mitotic cell.
a, b - mitosis and c, d - non-mitosis
A set of textural features were extracted from a window of 100 X 100 around the bounding box of each segmented cell.
Second Level Sieve contd.
48 phase gradient and 48 gabor features extracted.
3 bands used - Red, Green, Blue.
For each of these 96 features, 4 statistical measures calculated:-
a) mean
b) skewness
c) kurtosis
d) standard deviation
Training set - All Mitosis (after First Level Sieve) + Equal number of randomly selected non mitotic instances from the original dataset.
Second Level Sieve Contd.
An ensemble of Random Projections and SVM (linear kernel) with a majority rule was used for final mitosis prediction.
Final Sensitivity = 82.35%
Final PPV = 73.04%
Experimental Results
ICPR 2012 Mitosis Detection Contest Participants
SDSNE (Supervised Dimensionality Selection based on Neighborhood Examination)
Quantitative Evaluation
Qualitative Evaluation
Segmentation and Data Imbalance
The mean and the standard deviation of the distance between the detected mitosis and the ground truth centroids were found out to be 0.87 pixel and 0.45 pixel respectively.
An experiment was carried out to test the confidence limit of our proposed scheme:
Testing data: HPF's having less than or equal to 2 mitosis
Sensitivity = 81.13% and PPV = 74.97%
1. Nottingham Grading Scheme:
a) Mitotic count
b) Nuclear Atypia
c) Tubule formation
2. Some studies suggest that mitosis count can be as predictive as the grading system
Mitosis - Small objects with large variation in shapes.
Similar looking lymphoid/inflammmatory and apoptic cells.
High degree of dataset imbalance.
IEEE Signal Processing Society
The LNM Institute of Information Technology, Jaipur, India.
Acknowledgments
Training - 33 slides
Testing - 15 slides
Seeds for Active contour
BB-07
BB-06
Histogram Equalization
Gray Level Thresholding
Smoothing blobs
Entropy based stack selection
Band Selection
Active contours without edges - Tony F. Chan and Luminita A. Vese
Feature Set
Optimal Feature Set
Mitosis
Non- Mitosis
SMOTE + Tomek Links Removal
Full transcript