Developing Echocardiographic Encoders Using Self-Supervised Learning
University of West London – School of Computing and Engineering
Background:
Echocardiography is a non-invasive imaging technique that uses ultrasound waves to produce images of
the heart. It is a widely used diagnostic tool for a variety of cardiovascular conditions, including
heart failure, valvular heart disease, and congenital heart disease.
Echocardiographic images are complex and contain a wealth of information about the structure and
function of the heart. However, manually extracting this information from echocardiographic images is
a time-consuming and challenging task.
Machine learning has the potential to automate the analysis of echocardiographic images and make
them more accessible to clinicians. In particular, deep learning-based encoders can be used to extract
informative features from echocardiographic images that can be used for a variety of tasks, such as
disease classification, segmentation, and quantification.
Traditional deep learning-based encoders are trained using supervised learning, which requires a
large dataset of labelled echocardiographic images. However, such datasets are expensive and
time-consuming to collect.
Self-supervised learning is a promising alternative to supervised learning for training deep
learning-based encoders. Self-supervised learning does not require labelled data, and instead learns
features from unlabeled data by solving pretext tasks.
Research Questions:
This PhD project will explore the following research questions:
- Can self-supervised learning be used to develop echocardiographic encoders that are more efficient
and effective than traditional encoders trained on labelled data?
- What are the best self-supervised learning architectures and training strategies for
echocardiographic image analysis?
- How can self-supervised learning encoders be used to develop clinically useful applications, such
as automated image segmentation and classification?
Aim and Objectives:
The aim of this PhD project is to develop self-supervised learning encoders for echocardiographic
image analysis. The specific objectives are to:
- Develop self-supervised learning synthetic tasks tailored for training echocardiographic encoders
without explicit labels.
- Optimise encoder architectures to extract clinically relevant features from unlabeled
echocardiographic data.
- Train the self-supervised learning encoders on a large dataset of unlabeled echocardiographic
images.
- Evaluate the transferability of learned representations for downstream clinical tasks, including
classification and segmentation.
Methodology:
The aim of this PhD project is to develop self-supervised learning encoders for echocardiographic
image analysis. The specific objectives are to:
- Develop self-supervised learning synthetic tasks tailored for training echocardiographic encoders
without explicit labels.
- Optimise encoder architectures to extract clinically relevant features from unlabeled
echocardiographic data.
- Train the self-supervised learning encoders on a large dataset of unlabeled echocardiographic
images.
- Evaluate the transferability of learned representations for downstream clinical tasks, including
classification and segmentation.
Clinical Partners:
Collaboration with clinical partners at the School of Medicine, Imperial College London, involves
regular meetings and discussions to:
- Ensure alignment with clinical needs and priorities.
- Validate the clinical relevance of the developed encoders.
- Facilitate the translation of research findings into practical applications.
- Availability of patient data and expert annotationtations for model developments.
For more information about the project, please contact Professor Massoud Zolgharni