Skip to Main content Skip to Navigation
Master Thesis

Application de méthodes de « Deep Learning » en intelligence artificielle pour l’identification en temps réel de vues échocardiographiques

Abstract : Introduction: echocardiography is an indispensable tool for the management of cardiovascular diseases, its access is unfortunately limited by the number of experts. The use of artificial intelligence techniques opens the possibility of assisting an operator, in the acquisition and analysis of focused echocardiography, using new ultra-portable devices, thus allowing access to the greatest number. Identifying in real-time the reference views required for this exam is the first step. Objective: develop an algorithm that can identify echocardiographic reference views in real-time with sufficient quality to ensure reliable interpretation. Method: use of echocardiographic exams from the database of Bordeaux University Hospital. Annotation of video acquisitions with 8 labels, 7 of which correspond to the mains reference views that meet defined quality criteria. The last label called "not determined" was attributed to all other views including reference views that did not meet the quality criteria. A convolutional neural network with a simplified VGG architecture was trained to distinguish these 8 labels. Performances were evaluated by image and by acquisition with cross-validation. Results: 369 studies from 358 patients, representing 5785 acquisitions, were annotated. The "not-determined" class accounted for 49.7% of acquisitions. The algorithm distinguishes the 8 classes with an average accuracy of 81.75 ± 1.27% per frame and 82.97 ± 1.42% per acquisition. The F1 score is 79.15 ± 2.50% per frame and 81.48 ± 1.91% per acquisition. Most of the classification errors are between the reference views and the "not determined" class, since only 35 (0.61%) of the reference views have been classified in the wrong reference view. Conclusion: real-time identification of seven references views of sufficient quality for performing focused echocardiography is feasible with acceptable accuracy.
Complete list of metadata

Cited literature [61 references]  Display  Hide  Download
Contributor : Médecine Ub Connect in order to contact the contributor
Submitted on : Wednesday, March 4, 2020 - 11:57:50 AM
Last modification on : Wednesday, March 18, 2020 - 1:18:19 AM
Long-term archiving on: : Friday, June 5, 2020 - 1:40:58 PM


 Restricted access
To satisfy the distribution rights of the publisher, the document is embargoed until : 2022-01-01

Please log in to resquest access to the document


  • HAL Id : dumas-02498256, version 1



Alexandre Dezellus. Application de méthodes de « Deep Learning » en intelligence artificielle pour l’identification en temps réel de vues échocardiographiques. Sciences du Vivant [q-bio]. 2019. ⟨dumas-02498256⟩



Record views