Alternative data representations for a Deep Learning-based segmentation pipeline applied to fetal Doppler echocardiography

Abstract

Doppler echocardiography is a crucial image acquisition technique in fetal medicine that generates spectrums of blood velocities. The current pipeline for its segmentation is very reliant on manual quantification steps, resulting labour-intensive and time expensive. Given the rise of Deep Learning in the medical image segmentation field, some initial Deep Learning based models have been trained and tested for its automatic segmentation. A project in the scope of a grant awarded by the Bill and Melinda Gates Foundation's Global Health program, has obtained some initial good results. Their baseline solution proposed uses a W-net with 6 levels and a binary mask as data representation with values of 1 from the reference line to the curve position. However, these results could be improved. The aim of this project is to design Deep Learning based models using alternative data representations in order to find an alternative solution that overperforms the baseline solution. The dataset used contains 7063 fetal Doppler echocardiographic images which are split into training, validation and test sets. The model architectures used are U-net and W-net architectures with different levels, from 5 to 7. The data representations proposed are a binary mask around the curve position using different width values, and a linear regression. 24 models are trained combining all the architectures with the several data representations, using Dice loss for binary mask data representation models and mean square error (MSE) loss for models using linear regression. For the performance evaluation, different metrics are used when models predict unseen data from the test set. The results show that the baseline solution overperforms the alternative solutions tested in this project. It is observed that more complex and deep architectures with a data representation based on binary masks that generate big shapes work better for these images. Further alternative solutions can be studied in order to develop a much powerful segmentation tool.