Sammendrag
The thesis explores the use of transfer learning for automated segmentation of 2D echocardiograms in the situation where few labeled data are available for end-to-end training. The main focus is on transfer learning involving data sets of the left heart chambers and the right heart chambers, with a goal of improving segmentation performance on the latter kind of data. A custom version of the U-Net neural network was implemented for automated segmentation and trained on two left heart data sets and one right heart data set. The worst performing models in direct training on right heart data significantly improved through pre-training on left heart data. Their multi-class Dice Score rose by 6% on average, while the score for RV epicardium improved by 16%. Predictions made by the models were also explored, revealing that the left heart chambers and the right heart chambers share certain features in the images that are useful for learning segmentation. It is concluded that transfer learning appears to be a feasible approach for echocardiogram segmentation when there is a lot of data available in the source task data set and far less data in the target task data set. Using transfer learning with the right setup can therefore reduce the amount of right heart ultrasound data that needs to be collected for training AI segmentation models. However, further investigation is required to quantify some observations made throughout the work.