Slide 1

Slide 1 text

breast cancer diagnosis using digital repositories improving the César Suárez Ortega uaa@ceta-ciemat.es 18-20 March 2013 International Work-Conference on Bioinformatics and Biomedical Engineering

Slide 2

Slide 2 text

The IMED project

Slide 3

Slide 3 text

introduction » IMED is a collaboration among: » CETA-CIEMAT (Trujillo, Spain) » INEGI (Porto, Portugal) » FMUP-HSJ (Porto, Portugal) » Since 2008 » Created to ease the cancer diagnosis. » Focused on breast cancer.

Slide 4

Slide 4 text

Breast cancer is most common cancer for women. According to World Health Organization: http://www.who.int/cancer/events/breast_cancer_month/en/ 1 1 1

Slide 5

Slide 5 text

Early detection is the cornerstone of its control.

Slide 6

Slide 6 text

Double checking of cases reduces the ratio of missed cancer R. Ramos-Pollan, et al., Discovering Mammography-based Machine Learning Classifiers for Breast Cancer Diagnosis. Journal of Medical Systems (9 April 2011), pp. 1-11. 1 1

Slide 7

Slide 7 text

But this is expensive!

Slide 8

Slide 8 text

But this is expensive! Solutions?

Slide 9

Slide 9 text

But this is expensive! Solutions? Computer-aided diagnosis

Slide 10

Slide 10 text

But this is expensive! Solutions? Computer-aided diagnosis CAD Systems 1 1

Slide 11

Slide 11 text

what are CAD systems? » Assistance in the diagnosis of medical images » Multiple disciplines are combined: » Artificial intelligence » Digital image processing » Radiology

Slide 12

Slide 12 text

what are CAD systems? » Assistance in the diagnosis of medical images » Multiple disciplines are combined: » Artificial intelligence » Digital image processing » Radiology They don't replace doctors, they support them.

Slide 13

Slide 13 text

MIWAD BCDR MLCs three pillars Breast Cancer Digital Repository Machine Learning Classifiers Mammography Image Workstation for Analysis and Diagnosis

Slide 14

Slide 14 text

MIWAD BCDR MLCs three pillars BCDR – A repository of annotated breast cancer cases. MLCs Machine learning classifiers for breast cancer diagnosis MIWAD – A workstation prototype with a CAD system.

Slide 15

Slide 15 text

MIWAD BCDR MLCs

Slide 16

Slide 16 text

MLCs » Artificial intelligent systems. » They learn from data.

Slide 17

Slide 17 text

benign malign Examples MLC info + classification info + classification

Slide 18

Slide 18 text

benign malign Examples MLC – benign? malign? benign classification info + classification info + classification info

Slide 19

Slide 19 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them!

Slide 20

Slide 20 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree

Slide 21

Slide 21 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule

Slide 22

Slide 22 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule Neural Networks

Slide 23

Slide 23 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule Neural Networks Genetic Programming

Slide 24

Slide 24 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule Neural Networks Genetic Programming Support Vector Machines

Slide 25

Slide 25 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule Neural Networks Genetic Programming Support Vector Machines Clustering

Slide 26

Slide 26 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule Neural Networks Genetic Programming Support Vector Machines Clustering Bayesian Networks

Slide 27

Slide 27 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule Neural Networks Genetic Programming Support Vector Machines Clustering Bayesian Networks Inductive Logic

Slide 28

Slide 28 text

MLCs » Artificial intelligent systems. » They learn from data. » Lot of them! Decision Tree Association Rule Neural Networks Genetic Programming Support Vector Machines Clustering Bayesian Networks Inductive Logic Sparse Dictionary

Slide 29

Slide 29 text

What is the best one for breast cancer?

Slide 30

Slide 30 text

We are working on it :)

Slide 31

Slide 31 text

current results » GRID and HPC to speed up the research. » Metrics » ROC → True positives VS. False positives » AUC → Probability of a true positive » Our classifiers?

Slide 32

Slide 32 text

current results » GRID and HPC to speed up the research. » Metrics » ROC → True positives VS. False positives » AUC → Probability of a true positive » Our classifiers? AUC = 85%

Slide 33

Slide 33 text

What data we use to train our MLCs?

Slide 34

Slide 34 text

MIWAD BCDR MLCs

Slide 35

Slide 35 text

BCDR is a wide-ranging annotated repository of anonymous breast cancer studies

Slide 36

Slide 36 text

BCDR » Our reference dataset for MLCs. » It contains a custom subset of DICOM and: » Reviewed by specialists and biopsy proven » Cases classified using BI-RADS » Useful for researchers. 23 17 per lesion region intesity, texture and shape features clinical features per case

Slide 37

Slide 37 text

data model patients images lesions segmentations classifications studies * * * * * *

Slide 38

Slide 38 text

data model patients images lesions segmentations classifications studies * * * * * * ?

Slide 39

Slide 39 text

segmentation of a lesion

Slide 40

Slide 40 text

two flavours BCDR-FMR 1010 cases 795 segmented lesions 3073 mammographies 720x1168 TIFF images (8bpp)

Slide 41

Slide 41 text

two flavours BCDR-DMR 600 cases 828 segmented lesions 2837 mammographies 3328x4084 TIFF images (14bpp)

Slide 42

Slide 42 text

two flavours BCDR-DMR 600 cases 828 segmented lesions 2837 mammographies 3328x4084 TIFF images (14bpp) + 2073 ultrasound images 800x600 (8bpp)

Slide 43

Slide 43 text

No content

Slide 44

Slide 44 text

No content

Slide 45

Slide 45 text

http://bcdr.inegi.up.pt

Slide 46

Slide 46 text

http://bcdr.inegi.up.pt

Slide 47

Slide 47 text

http://bcdr.inegi.up.pt

Slide 48

Slide 48 text

http://bcdr.inegi.up.pt

Slide 49

Slide 49 text

MIWAD BCDR MLCs

Slide 50

Slide 50 text

MIWAD » A workstation for analysing and diagnosing. » It's multiplatform (Windows, Linux & Mac) » Composed by two connected apps: MIWAD-DB MIWAD-CAD

Slide 51

Slide 51 text

MIWAD-DB » It manages... » ...the clinical data » ...the mammographies » BCDR as data source. » It operates over BCDR data. » Executes MLCs with a 1-click process.

Slide 52

Slide 52 text

No content

Slide 53

Slide 53 text

patients studies lesions segmentations classifications

Slide 54

Slide 54 text

mammographies

Slide 55

Slide 55 text

classifiers

Slide 56

Slide 56 text

results

Slide 57

Slide 57 text

MIWAD-CAD » Screening tool (based on TUDOR Viewer) » It includes image processing tools. » Tools to help in the segmentation process. » Three functions: 1 It stores the segmentations in BCDR. 2 It calculates image features. 3 It helps doctors to find lessions.

Slide 58

Slide 58 text

CAPTURAS CAD

Slide 59

Slide 59 text

No content

Slide 60

Slide 60 text

No content

Slide 61

Slide 61 text

this is real

Slide 62

Slide 62 text

No content

Slide 63

Slide 63 text

CAD DB Bruno São João Hospital

Slide 64

Slide 64 text

summarizing...

Slide 65

Slide 65 text

MIWAD BCDR MLCs BCDR – A repository of annotated breast cancer cases. MLC Machine learning classifiers for breast cancer diagnosis MIWAD – A workstation prototype with a CAD system.

Slide 66

Slide 66 text

MIWAD BCDR MLCs BCDR is used to train MLCs

Slide 67

Slide 67 text

MIWAD BCDR MLCs BCDR is the data source of MIWAD...

Slide 68

Slide 68 text

MIWAD BCDR MLCs ...and MIWAD feeds the BCDR

Slide 69

Slide 69 text

MIWAD BCDR MLCs So, MIWAD helps to improve the MLCs

Slide 70

Slide 70 text

MIWAD BCDR MLCs The MLCs are integrated in MIWAD

Slide 71

Slide 71 text

future work » IMED goes educational! » MIWAD as a training app for radiologists: » Web architecture & responsive design. » Auto-evaluation using MLCs. » BCDR as a formative resource. » Continuation of the improvement of MLCs. (+90% AUC) » Certification for MLCs » Continuation of the BCDR growth (searching new sources)

Slide 72

Slide 72 text

conclusions » CAD systems can achieve great results in: » Improving the breast cancer diagnosis » Training specialists » Specialists demand better educational apps (difficult to find) » All is about data! » It's difficult to get reliable data. » BCDR is the most complete open breast cancer repository. » MLCs data

Slide 73

Slide 73 text

CETA-CIEMAT acknowledges the support received from the European Regional Development Fund through its Operational Programme Knowledge-based Economy.

Slide 74

Slide 74 text

any questions? thanks for your attention! :)