White Blood Cell Classification Using Texture and RGB Features of Oversampled Microscopic Images

Abstract

White blood cell (WBC) type classification is a task of significant importance for diagnosis using microscopic images of WBC, which develop immunity to fight against infections and foreign substances. WBCs consist of different types, and abnormalities in a type of WBC may potentially represent a disease such as leukemia. Existing studies are limited by low accuracy and overrated performance, often caused by model overfit due to an imbalanced dataset. Additionally, many studies consider a lower number of WBC types, and the accuracy is exaggerated. This study presents a hybrid feature set of selective features and synthetic minority oversampling technique-based resampling to mitigate the influence of the above-mentioned problems. Furthermore, machine learning models are adopted for being less computationally complex, requiring less data for training, and providing robust results. Experiments are performed using both machine- and deep learning models for performance comparison using the original dataset, augmented dataset, and oversampled dataset to analyze the performances of the models. The results suggest that a hybrid feature set of both texture and RGB features from microscopic images, selected using Chi2, produces a high accuracy of 0.97 with random forest. Performance appraisal using k-fold cross-validation and comparison with existing state-of-the-art studies shows that the proposed approach outperforms existing studies regarding the obtained accuracy and computational complexity.

Keywords: Chi-squared; SMOTE; leukemia; texture features; white blood cells classification.

Figure 1

Types of WBC. (a) Neutrophils, (b) Basophils, (c) Eosinophils, (d) Monocytes, and (e) Lymphocytes.

Figure 2

Hierarchy and classification of WBCs.

Figure 3

Flow of the proposed methodology.

Figure 4

Process followed to make the hybrid feature set.

Figure 5

Approach followed for data balance using SMOTE.

Figure 6

Comparison between all machine learning approaches used for classification of WBC types.

Figure 7

RF confusion matrix using HB features.

Figure 8

Comparison of original features space and proposed features space. We visualize features in 3D space and show that the features set makes a good correlation with target classes or not. If there will be less overlapping in different target samples, this means that the feature set is good for achieving significant results. We used PCA to convert our features into 3 dimensions, and then illustrated them using a scatter plot. (a) Original dataset RGB features, (b) Original dataset texture features, (c) Hybrid features space at 0°, and (d) Hybrid features space at 90°.

Figure 9

Architecture of the state-of-the-art deep learning models. (a) CNN, (b) ResNet50 [41], and (c) VVG16.

Figure 10

Training and testing per epochs accuracy and loss graphs. (a) Accuracy on original dataset, (b) Loss on original dataset, (c) Accuracy on augmented dataset, (d) Loss on augmented dataset, (e) Accuracy on over-sampled dataset, and (f) Loss on over-sampled dataset.