I have chosen my thesis theme, because I’m interested in Deep Learning and I want to be absorbed in it. When I was choosing my thesis theme, I considered that I do not have extensive knowledge of medicine and medicine software’s. I have found it interesting and I believe that it is a good opportunity to upgrade my knowledge about this field. That’s why I have chosen this topic.
The problem is complex, because in medicine there is no accepted unified solution which segment heart MR images. Of course, there are already existing solutions (parts), initially these directed me to the right way.
This thesis presents among others the basic concepts of deep learning, the technology which helps to create these types of solutions and known heart MRI segmentation solutions (or just parts of them). Beyond this knowledge it tackles the main problems what need a solution and after I will present my own solution. I came up with two solutions, with the help of these the second method gives me results which are accurate enough (ROI-B). In the results I will evaluate the model’s accuracy during training.
Based on this thesis, there is a re-creatable neural network architecture, which has got predictions which are accurate. The average error is around 6 pixels. This means that the distance of the predicted and the real results of the region of interests are 6-pixel width. This can be very useful for the segmentation problem when it is used as a preprocessing step.