Agenda

Introduction

Adventages

Pipeline

Features

Results

Problems

Conculsions

Tumor CLASSIFICATION

MRI

WSI

Tumor CLASSIFICATION

WSI Tumor CLASSIFICATION

Adventages

Cellular level

A lot of information

Expert feedback

Great potential

Universal mechanisms

Adventages

Cellular level

A lot of information

Expert feedback

Great potential

Universal mechanisms

Cellular level resolution provides necessary information about the case. It is possible to precisely classify cells from overpopulated regions and predict type and stage of tumor

Adventages

Cellular level

A lot of information

Expert feedback

Great potential

Universal mechanisms

Every cell with it's surroundings contains a lot of relevant information. It can be used to simply classify tumor or healthy tissue.  It can be also used to find unusual, atypical behavior  to better understand the nature of the case

Adventages

Cellular level

A lot of information

Expert feedback

Great potential

Universal mechanisms

It is crucial to get professional feedback. Histopathologists are very solid source of fruitful feedback

Adventages

Cellular level

A lot of information

Expert feedback

Great potential

Universal mechanisms

The development of automatic WSI analysis can significantly improve efficiency and pace at which pathologists work. Moreover it seems possible to provide diagnosis without involvement of doctors

Adventages

Cellular level

A lot of information

Expert feedback

Great potential

Universal mechanisms

Vast majority of functions used for segmentation and feature extraction can be easily generalized to different cases and different tissue

Feature details

Area
Circumference
Connectivity
Hu moments
Ellipse atributes
Density of nuclei
Fuzzy coefficient
Core of nucleous
Interior structure
- RGB
- Gray scale
- Hematoxylin and eosin

Halo effect

Shape

Color

Divide in Frames
▼
Frame segmentation
▼
Extract frames of interest
▼
Classify nuclei in a frame
▼
Classify each frame
▼
Classify the sample

General procedure

Divide in Frames
▼
Frame segmentation
▼
Extract frames of interest
▼
Classify nuclei in a frame
▼
Classify each frame
▼
Classify the sample

General procedure

Divide in Frames
▼
Frame segmentation
▼
Extract frames of interest
▼
Classify nuclei in a frame
▼
Classify each frame
▼
Classify the sample

General procedure

Divide in Frames
▼
Frame segmentation
▼
Extract frames of interest
▼
Classify nuclei in a frame
▼
Classify each frame
▼
Classify the sample

General procedure

n = 774

n = 332

n = 815

n = 68

n = 774

n = 815

Divide in Frames
▼
Frame segmentation
▼
Extract frames of interest
▼
Classify nuclei in a frame
▼
Classify each frame
▼
Classify the sample

Degree of membership

Astrocyte

Oligodendrocyte

General procedure

Divide in Frames
▼
Frame segmentation
▼
Extract frames of interest
▼
Classify nuclei in a frame
▼
Classify each frame
▼
Classify the sample
Oligo - 63%
Astro - 37%
Oligo - 24%
Astro - 76%

General procedure

Divide in Frames
▼
Frame segmentation
▼
Extract frames of interest
▼
Classify nuclei in a frame
▼
Classify each frame
▼
Classify the sample
Number of Oligodendroglioma frames  > T
▼
Oligodendroglioma

General procedure

Classify nuclei in a frame

Supported vector classification

Logistic regression

Random forests

Gaussian Naive Bayes

Supported vector machine

Explored classification algorithms

Explored classification algorithms

Supported vector classification

Logistic regression

Random forests

Gaussian Naive Bayes

Supported vector machine

Methodology

Model metaparameters were fitted with the use of genetic algorithm. 

Cost function was calculated via     score calculated on cases. 

Classify nuclei in a frame

F_1
F1F_1

Classify each frame and classify the sample

Degree of membership
▼
Frame summary x 7
▼
Classification of the frame x 7
▼
Frame set summary x 7
▼
Thresholding
\frac{\sum\limits_{i\in I}^n{\arcsin\left( 2x_i - 1 \right)}}{\pi \times n}
∑i∈Inarcsin(2xi−1)π×n \frac{\sum\limits_{i\in I}^n{\arcsin\left( 2x_i - 1 \right)}}{\pi \times n}
\max_{i\in I}\{x_i\}
maxi∈I{xi}\max_{i\in I}\{x_i\}
\frac{\sum\limits_{i\in I}x_i}{n}
∑i∈Ixin\frac{\sum\limits_{i\in I}x_i}{n}

Results and voting method

Crossvalidation
▼
Selecting best models
▼
Voting

Results and voting method

32 cases total:

  • 12 to teach nuclei classifier                  
  • 10 to teach frame classifier                   
  • 10 to test
Crossvalidation
▼
Selecting best models
▼
Voting

Results and voting method

Achieved about 73% average accuracy

Crossvalidation
▼
Selecting best models
▼
Voting

Results and voting method

Crossvalidation
▼
Selecting best models
▼
Voting

Final decision is based on the simple majority voting over the classifier output population

Problems

  • Time
  • Multitasking
  • Segmentation accuracy
  • Different data types
  • Data mining problems

Conclusions 

Can we help people?

Conclusions 

Can we help people?

I hope you agree, that we can 

Conclusions 

We plan to:

  • speed up computing (GPU) , 
  • use more efficient segmentation methods,
  • improve feature extraction methods, 
  • implement other classification algorithms, 

but also:

  • add MRI analysis as a significant part of the entire process,
  • add medical history as a significant part of the process,
  • expand our algorithm to deal with other brain tumor types

Our Team

Michał Błach

R&D Stermedia

Wrocław University of Technology

 

Witold Dyrka

R&D Stermedia

Wrocław University of Technology

Jakub Czakon

R&D Stermedia

Piotr Giedziun

R&D Stermedia

Piotr Krajewski

CIO Stermedia

Grzegorz Żurek

R&D Stermedia

Wrocław University of Technology

 

Łukasz Fuławka

Pathomorphology resident Lower Silesian Oncology Center

We are looking forward to collaborating with you

Contact us 

info@stermedia.pl

jakub.czakon@stermedia.pl

grzegorz.zurek@stermedia.pl

Thank you for attention

Made with Slides.com