Classification of white blood cells using deep learning

Nowadays, medical image analysis has become an increasingly indispensable tool during the diagnosis and treatment of many illnesses. Particularly, white blood cell images observation matters. The amount trend of different classes of white blood cells can serve as a prediction of many blood diseas...

Full description

Saved in:
Bibliographic Details
Main Author: Zhang, Mengxin
Other Authors: Wen Bihan
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2022
Subjects:
Online Access:https://hdl.handle.net/10356/158876
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Nowadays, medical image analysis has become an increasingly indispensable tool during the diagnosis and treatment of many illnesses. Particularly, white blood cell images observation matters. The amount trend of different classes of white blood cells can serve as a prediction of many blood diseases. Deep learning has been widely used in the medical image analysis area. However, in the medical imaging area, because of reasons such as privacy and imbalance of patients, rich data sets with accurate annotations are always hard to make. As a result of that, transfer learning sometimes is chosen as a method to make up for this deficiency. The experiment of this dissertation is based on a newly proposed white blood cells data set Raabin-WBC with the objective of classifying five different classes of leukocytes: neutrophils, monocytes, lymphocytes, eosinophils, and basophils. Three kinds of traditional models with great performance in other tasks are used: VGG, ResNet, and ResNext. For each architecture, different network depths are used. Transfer learning is used for models with large depth. The results show that those models perform well in test set A which is similar to the training set while performing badly in test set B which differs a lot from the training set. Besides, several comparative analyses are made to show the effect of different influence factors such as network depth.