The objective is the development and validation of methodology that automatically extracts diagnostic and prognostic biomarkers from echocardiographic sequences. The strategy will be to explicitly formulate the problem as a combination of sources of useful and confounding information. The intra- and inter- observer variability in echocardiography is one of the main limitations of this imaging modality, and the project will investigate strategies to reduce it through the automatic identification of the optimal vs. the sub-optimal field of view. The goals are both the provision of quality metrics and the extraction of anatomical and deformation based clinical markers. The solution will be evaluated in a unique database of 120.000 scans with associated clinical information, and the first specific envisioned application is to automatically score the response to a stress echocardiogram.
Echocardiography is the most common imaging modality in the management of cardiovascular conditions. The analysis and interpretation of these images is a key element in the training and workload of cardiologists, and many clinical guidelines recommend the acquisition of these sequences to identify the aetiology, decide treatment options and monitor progression of the disease process. Image analysis solutions, widely available from commercial products, provide semi-automatic extraction of key metrics such as mass, volume, thickness or strain. However, the utility of these metrics is limited by the problem of inter- and intra- observer variability.
On the other hand, the field of machine & deep learning is currently going through a huge explosion of interest, driven by the excellent results that this approach renders in tasks of classification or automation of otherwise quite challenging problems. The core idea is that the machine is able to identify the hidden patterns in the data that are useful for the solution of the problem, and that this technology has the potential to improve the reproducibility of measurements.
The third main component of this project is the use of statistical models of anatomy and function. This technology allows us to capture and characterise the range of variability observed in 2D, 3D or 4D datasets, and to investigate the causes of that variability by discriminant or correlation analyses.
This project explores the hypothesis that echocardiographic sequences contain further diagnostic and prognostic information beyond current metrics of mass, volume and strain, and that statistical modelling & deep learning technologies are able to capture the hidden signatures of cardiac disease in a robust manner. The two specific objectives are:
- To develop a solution to identify the optimal vs. the sub-optimal orientation of the 2D echocardiographic probe. The application will be both the reduction of variability at the acquisition stage, and the provision of quality metrics for further analysis.
- To develop a solution to extract anatomical and deformation metrics from 2D scans after removal of the confounding factor of the orientation of the probe.