White blood cell recognition for biomarker model using improved convolutional neural network (CNN)
White Blood Cell (WBC) is one of important elements in protecting human’s immunity system. WBC composed of two main elements namely cytoplasm and nucleus, with almost the same in color contrast, and hence make it hard to be detected and analyzed. Manual WBC analysis is less efficient, therefore a Co...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/8495/1/24p%20SYADIA%20NABILAH%20MOHD%20SAFUAN.pdf http://eprints.uthm.edu.my/8495/2/SYADIA%20NABILAH%20MOHD%20SAFUAN%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/8495/3/SYADIA%20NABILAH%20MOHD%20SAFUAN%20WATERMARK.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | White Blood Cell (WBC) is one of important elements in protecting human’s immunity system. WBC composed of two main elements namely cytoplasm and nucleus, with almost the same in color contrast, and hence make it hard to be detected and analyzed. Manual WBC analysis is less efficient, therefore a Computer Aided Diagnosis (CAD) based on Deep Learning (DL) model become subject of interest nowadays. However, with vast amount of WBCs data and various DL architectures available, tuning an optimal DL model is a daunting task. In this project, a diagnostic algorithm for WBCs’ DL analysis is proposed by combining transfer learning approach with fine tuning (FT) approach and tested on Kaggle dataset (9957 images). Initially, a transfer learning analysis are conducted using six well known Convolutional Neural Network (CNN) structure which are Alexnet, Googlenet, Densenet, Mobilenet, Resnet and VGG. Next, the CNN model that yield high accuracy with less prone to overfitting is selected for the FT process. Finally, the optimal FT model will undergo series of performance testing using two public WBCs dataset which are LISC and IDB-2. From series of experiments, it can be concluded that the pre-trained AlexNet model gave highest performance with 98.79% training accuracy and 99.10% testing accuracy compare to other models. When implementing the layer refinement analysis, it shows that AlexNet model with 4-layer FT and RMSProp optimizer, significantly improved the performance with 100% accuracy in training and 100% accuracy in testing. Eventually, the FT-ALexNet model tested on LISC database demonstrate 95.52% performance for the true-positive samples, and on IDB-2 100% performance for the true-negative samples. Proposed method (FT-Alexnet) which compares and applies two-stage optimization technique works well with several WBC datasets and able to improve 2.22% from the existing work. |
|---|