Automatic analysis in echocardiography using machine learning
MetadataShow full item record
Echocardiography is the cornerstone of modern cardiac imaging due to its availability, low cost and real-time functionality. The modality has enabled sophisticated non-invasive evaluation of the hearts morphophysiology, with a wide range of clinical parameters of high diagnostic and prognostic value. However, despite the clinical impact, quantitative measurements are often omitted in clinical practice by being labor intensive, time consuming and difficult to reproduce. Automation can reduce some of these limitations and redefine parts of the clinical workflow, but the design of generic algorithms is complex due to the inherent variability of echocardiography data and the expertise required for interpretation. The overall goal of this work was to investigate the use of deep learning (DL) methods for fully automating several image analysis steps of an echocardiography exam. Emphasis was given to method adaptation for ultrasound (US) image processing, as well as addressing fundamental domain limitations such as noise and acquisition variability. Real-time support and workflow enhancements was also important features in the development. The thesis consists of three technical contributions and one clinical feasibility study. In the first part, a method for cardiac view classification with convolutional neural networks (CNNs) is presented. Further, we describe a recurrent CNN method for cardiac event detection. The third part presents a DL based motion estimator, and the integration of several DL components into a pipeline for automated longitudinal strain (LS) measurements. The last part is dedicated to a feasibility study comparing the latter with a commercially available solution. Results indicate that the different components can benefit or even be improved with DL. The flexibility of learning-based approaches helps to surpass conventional methods on inherent limitations of US. Integrating DL components in a pipeline for fully automated measurements was feasible, and yielded encouraging results by being comparable to intervendor variability. Despite several limitations described in the thesis, we can be optimistic about the future employment of DL in echocardiography.
Has partsPaper 1: Østvik, Andreas; Smistad, Erik; Aase, Svein Arne; Haugen, Bjørn Olav; Løvstakken, Lasse. Real-time Standard View Classification in Transthoracic Echocardiography using Convolutional Neural Networks. Ultrasound in Medicine and Biology 2018 https://doi.org/10.1016/j.ultrasmedbio.2018.07.024
Paper 2: Fiorito, Adrian Meidell; Østvik, Andreas; Smistad, Erik; Leclerc, Sarah; Bernard, Olivier; Løvstakken, Lasse. Detection of Cardiac Events in Echocardiography using 3D Convolutional Recurrent Neural Networks. Proceedings - IEEE Ultrasonics Symposium 2018 https://doi.org/10.1109/ULTSYM.2018.8580137
Paper 3: Østvik, Andreas; Salte, Ivar Mjåland; Smistad, Erik; Nguyen, Thuy Mi; Melichova, Daniela; Brunvand, Harald; Haugaa, Kristina; Edvardsen, Thor; Grenne, Bjørnar; Løvstakken, Lasse. Myocardial Function Imaging in Echocardiography Using Deep Learning. IEEE Transactions on Medical Imaging 2021 ;Volum 40.(5) s. 1340-1351 https://doi.org/10.1109/TMI.2021.3054566 This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Paper 4: Salte, Ivar M.; Østvik, Andreas; Smistad, Erik; Melichova, Daniela; Nguyen, Thuy Mi; Karlsen, Sigve; Brunvand, Harald; Haugaa, Kristina H.; Edvardsen, Thor; Løvstakken, Lasse; Grenne, Bjørnar. Artificial Intelligence for Automatic Measurement of Left Ventricular Strain in Echocardiography. JACC Cardiovascular Imaging 2021 ;Volum 14.(10) s. 1918-1928 https://doi.org/10.1016/j.jcmg.2021.04.018 This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)