Semantic segmentation of peripheral white blood cells using neural networks
Semantic segmentation is the differentiation of the meaningful parts on an image. It has been used in many distinct fields, such as traffic or medical areas. One of these uses in the medical field is the blood smear examination. White blood cells (WBC) are part of the immune system and their countin...
| Autor: | |
|---|---|
| Tipo de documento: | dissertação |
| Data de publicação: | 2019 |
| País: | España |
| Recursos: | Universitat Oberta de Catalunya (UOC) |
| Repositório: | O2, repositorio institucional de la UOC |
| OAI Identifier: | oai:openaccess.uoc.edu:10609/90265 |
| Acesso em linha: | http://hdl.handle.net/10609/90265 |
| Access Level: | Acceso aberto |
| Palavra-chave: | DeconvNet U-Net SegNet Machine learning -- TFM Aprenentatge automàtic -- TFM Aprenendizaje automático -- TFM |
| Resumo: | Semantic segmentation is the differentiation of the meaningful parts on an image. It has been used in many distinct fields, such as traffic or medical areas. One of these uses in the medical field is the blood smear examination. White blood cells (WBC) are part of the immune system and their counting and determination are often performed by medical specialists for diagnosis. The shape and size of the nucleus of leukocytes can determine the type of WBC, by visual examination of an expert. WBC segmentation had been proposed before, but the convolutional neural network architectures were not tried for this task. Therefore, in this project, the semantic segmentation was performed on free access dataset, which is composed of microscopic images and segmented ground truth images, of WBC, made by experts. The dataset was filtered, transformed and augmented in order to be used in an artificial neural network. Some segmentation models, such as U-Net, SegNet and DeconvNet, were chosen, adapted and trained to/with this data. After training, the models were evaluated, using different metrics (accuracy, Jaccard similarity index and Sørensen¿Dice similarity coefficient), with the same dataset. Jupyter notebook from the free Google platform called Colaboratory was used for the training and evaluation of the models. Although all three models achieved very high scores in distinct metrics. U-Net architecture resulted in being the best model for segmenting, as well as the fastest one for the training process. |
|---|