In order to bring you the best possible user experience, this site uses Javascript. If you are seeing this message, it is likely that the Javascript option in your browser is disabled. For optimal viewing of this site, please ensure that Javascript is enabled for your browser.

The free consultation period for this content is over.

It is now only available year-round to EACVI Silver Members, Fellows of the ESC and Young combined Members

Automatic classification of CMR image sequences with convolutional neural networks

Session Poster session 3

Speaker Nay Aung

Event : EuroCMR 2019

  • Topic : imaging
  • Sub-topic : Imaging - Other
  • Session type : Poster Session

Authors : N Aung (London,GB), AM Lee (London,GB), MM Sanghvi (London,GB), K Fung (London,GB), JM Paiva (London,GB), RJ Thomson (London,GB), MY Khanji (London,GB), PB Munro (London,GB), SE Petersen (London,GB)

Authors:
N Aung1 , AM Lee1 , MM Sanghvi1 , K Fung1 , JM Paiva1 , RJ Thomson1 , MY Khanji1 , PB Munro1 , SE Petersen1 , 1Queen Mary University of London, William Harvey Research Institute - London - United Kingdom of Great Britain & Northern Ireland ,

Citation:
European Heart Journal - Cardiovascular Imaging ( 2019 ) 20 ( Supplement 2 ), ii504

Background:  The CMR protocol information in DICOM meta-data is vendor-dependent and may be erased during the anonymisation process.

Purpose: To automatically categorise the CMR image sequences using only image voxel-intensity data.

Methods: We used the UK Biobank CMR dataset (training/test set: 29,899/15,317 unique non-contrast images; 15 sequence classes), acquired using a 1.5T MAGNETOM Aera Siemens scanner. We constructed three separate convolutional neural networks (CNNs) adapted from VGG16 (model 1), Inception ResNet V2 (model 2) and Inception network V3 (model 3), which were all trained and tested using Tensorflow. We then compiled these three models into a final ensemble network (model 4). 

Results: All three CNNs classified 15 image sequence types in the test data with high accuracy and low error rates: 1.9%, 1.2% and 3.0% for models 1, 2 and 3, respectively. The final ensemble network (model 4) outperformed all three CNNs with an overall accuracy approaching 1. The performance metrics of model 4 for each image class are outlined in Table 1 and the confusion matrix is presented in Figure 1.

Conclusion: The CNN models can classify the common CMR sequence types with a very high level of accuracy and can be incorporated into a future fully-automated post-processing pipeline.

Precision Recall F1 score Validation sample size
SAX Cine 1 1 1 999
Inline VF 1 1 1 1318
LAX 3Ch Cine 1 1 1 1000
LAX 4Ch Cine 1 1 1 1000
LAX 2Ch Cine 1 1 1 1000
LVOT Cine 1 1 1 1000
Tagging 1 1 1 1000
T1 1 1 1 1000
Aortic flow 1 1 1 1000
Aortic distensibility 1 1 1 1000
Localiser: Heart Scout 1 0.99 0.99 1000
Localiser: SAX 0.99 1 1 1000
Localiser: LAX 1 0.99 0.99 1000
Localiser: Survey 0.99 0.97 0.98 1000
Localiser: Coronal/transverse thorax 0.96 0.99 0.98 1000
Average/total 1 1 1 15317

Members get more

Join now
  • 1ESC Professional Members – access all resources from general ESC events 
  • 2ESC Association Members (Ivory, Silver, Gold) – access your Association’s resources
  • 3Under 40 or in training - with a Combined Membership, access all resources
Join now

Our sponsors

ESC 365 is supported by Bayer, Boehringer Ingelheim and Lilly Alliance, Bristol-Myers Squibb and Pfizer Alliance, Novartis Pharma AG and Vifor Pharma in the form of educational grants. The sponsors were not involved in the development of this platform and had no influence on its content.

logo esc

Our mission: To reduce the burden of cardiovascular disease

Who we are