Institutional Repository of Vilnius University
Neuroninių tinklų taikymas vaizdų atpažinime /
Abstract
dc:descriptionThe aim of the bachelor thesis was to develop a model based on convolutional neural networks. The purpose of the model is to recognize B-lines in lung ultrasound imagery. To achieve this goal, a literature review was first carried out in order to review and compare existing B-line solutions in the scientific literature. B-lines are artifacts appearing in lung ultrasound imagery that correlate with various lung diseases. A total of 3 existing solutions in the literature were reviewed and compared. After the literature analysis, I proceeded to program the neural network model. To achieve this, it was first necessary to adapt a publicly available lung ultrasound dataset available on the internet. The authors of the dataset originally used the dataset to train a convolutional neural network that recognized lung diseases. However, the dataset was also suitable for the task of B-line classification, as the videos and photos were labelled as to whether or not B-lines existed (in a given video or photo). The dataset was also split into two separate datasets, the first dataset was further split into training, validation and testing parts. The second dataset was used only for testing the model. Once the datasets were ready to be used, the convolutional neural network had to be programmed. The Tensorflow and Keras deep learning libraries (python programming language) were used to achieve this task. The created model performed very well during training, validation and testing. However, when the trained model was tested with the data of the second dataset, the results deteriorated significantly. This could be due to several reasons. All possible reasons are discussed in the paper. Most likely, in order to adapt the dataset to the B-line classification task, the first dataset was inappropriately split into training, validation and testing parts. The main conclusion of the paper is that a model may perform very well on one dataset, but this does not necessarily mean that it will perform well and generalize on another dataset.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Vinkleris, Kęstutis,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD146243111&prefLang=en_US