Abstract:
Part of the work carried out within the framework of this thesis involves the automatic segmentation
of echocardiographic images.
The separation and identification of the different structures from accurate delineation, called
semantic segmentation, is the first step to measure surfaces or volumes.
However, segmentation in echocardiography is a particularly difficult task due to the lack of
clear boundaries, a low signal-to-noise ratio, the speckled texture specific to ultrasound images,
and the presence of numerous and complex image artifacts such as as reverberations and loss of
signal.
We have presented a fully automatic deep learning approach based on the U-NET architecture
by integrating EfficientNet as an encoder.
Our network has achieved 97% accuracy on training data as well as validation data, which
makes our network powerful.
The results of the test on the CAMUS challenge dataset clearly show this with a Dice score
above 0.8. As prospects, we want to make changes to our network using the transfer learning
technique in order to improve it and solve the problem of metadata of the 19 patients.
