Presentation CBMS Cardio - tccls.computer.orgtccls.computer.org/.../08/Presentation_CBMS_Cardio.pdf · INTRODUCTION: 3 Cardiovascular Diseases (CVDs) are the leading cause of death

EJECTION FRACTION CLASSIFICATION INTRANSTHORACIC ECHOCARDIOGRAPHYUSING A DEEP LEARNING APPROACH

2018 IEEE 31st International Symposium on Computer-Based Medical Systems 18/06/2018

João Figueira Silva – joaofsilva@ua.pt

AGENDAIntroduction

Related Work

Methods

Results

Discussion

Conclusion

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH 2

INTRODUCTION:

3

Cardiovascular Diseases (CVDs) are the leading cause of death worldwide

Accounted for approximately 17.7 million deaths in 2015

Highly correlated with left ventricle (LV) function indices:

Ejection Fraction (EF)

Wall thicknessVentricular mass

…

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

INTRODUCTION:

4

Techniques to assess patient condition:

Computed Tomography (CT)

Magnetic Resonance Imaging (MRI)

Transthoracic Echocardiography (TTE)

TTE is widely available, cheaper, and portable

Considered a first line of diagnosis technique

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

5

PhysicianTTE Exam

Metrics3D-CNN

Manual Segmentation Metrics

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

INTRODUCTION:

RELATED WORK:

6

Automatic left ventricle segmentation:Active ContoursSupervised Learning Methods

Issue: require prior knowledge on the left ventricle shape

Deep learning methods using 2D echocardiographyIssue: requires prior detection of the systole and diastole frames by the user

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

RELATED WORK:

7

3D information has already been used in EF estimation with data from:Magnetic Resonance Imaging3D Echocardiography

Issue: both techniques have higher associated costs than 2D Echocardiography

OBJECTIVE: Estimate ejection fraction using TTE cineloops (2D echo), consideringtime as the 3rd dimension

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

METHODS: DATASET

8

Exams containing TTE cineloops, annotated information and free text report

Manual selection of the Apical 4 Chambers view, EF data extraction from meta-data

Extraction of 30 sequential frames from each cineloop

Automatic extraction of the region of interest to mask burned in PHI

Frame size: 128 x 128 pixels

Categorization of EF values

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

METHODS: DATASET

9EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

10EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

METHODS: DATASET

METHODS: 3D-CNN

11EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

Input Bottom Layers

3D Convolution

3D Global Avg Pool

Batch Normalization

Leaky ReLU

Top Layers

<45%

45%-55%

55%-75%

>75%

ClassificationMiddle Layers

Residual Block Y

Residual Block Z

Residual Block X

+

ResidualBlock

Convblock X/Y/Z

Batch NormalizationLeaky ReLU

Batch Normalization

Convblock X/Y/ZLeaky ReLU

+

3D Conv 3F (1x1x1)

3D Conv 6F (1x1x1)

3D Conv 3F Z(1x3x1) X(1x1x3)

Y(3x1x1)

Convblock X/Y/Z

3D Convolution

3D Global Avg Pool

Batch NormalizationLeaky ReLU

Fully -Connected Dropout

3D for temporal knowledge

Asymmetric filters

Residual blocks

Batch normalization and Dropout

Mini-batch: 64 exams

Random grid search

12

RESULTS

Accuracy: 78%

F1 score (one vs all)

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

13

DISCUSSION

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

Class imbalance is reflected on model performance

Class 2 is a transitional class: contains both healthy and unhealthy exams

Discrepancy between values in meta-data and in clinical text reports

Noise in TTE images: possible incorrect data representation

14

CVDs are the leading cause of death worldwide

First line of diagnosis often based on TTE

Metrics like EF frequently require manual annotations

Novel 3D-CNN proposed to automate the process, obtaining promising results

Possible improvements:

Automatic selection of Apical 4 chambers view in TTE exams to expand dataset

Key-point selection in TTE cineloops

CONCLUSION

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHY USING A DEEP LEARNING APPROACH

THANK YOU FORYOUR ATTENTION

João Figueira Silva - joaofsilva@ua.pt 18/06/2018

EJECTION FRACTION CLASSIFICATION IN TRANSTHORACIC ECHOCARDIOGRAPHYUSING A DEEP LEARNING APPROACH

João Figueira Silva, Jorge Miguel Silva, António Guerra, Sérgio Matos and Carlos Costa2018 IEEE 31st International Symposium on Computer-Based Medical Systems

FOLLOW US AT

@IEETA_RESEARCH