Blood Cell Image Dataset

In this study, the blood cell images database was collected from an opensource repository; it is the Blood Cell Count Dataset (BCCD), which has 364 images of blood cells. These images contain 4888 different cells: 4155 RBCs cells, 372 WBCs cells, and 361 platelets ^{(BCCD, 2020}). Figure 1 shows RBCs, WBCs, and platelets on an input blood smear image. The deep learning model is trained with 256 images, and the model is validated with 54 blood cell images. Once the model is generated, it is tested with 54 blood cell images having 800 different blood cells, and counting is performed.

Note: derived from research.

Figure 1 Different Blood cells. 

Proposed Method

Artificial intelligence is significantly transformed with the arrival of deep learning (^{LeCun et al., 2015}; ^{Krizhevsky et al., 2017}). This technology is a part of machine learning that is utilized to solve complex problems with state-of-the-art performance on computer vision and image processing (^{Alom et al., 2018}). Deep learning models have been widely used on medical images, giving high results in detection, segmentation, and classification tasks. Proper annotation and labeling are performed on collected blood cell images by using the “LabelImg” application, which is further divided for training, validation, and test ing set. The dataset images are trained with convolutional neural networks (CNN) to generate a model to perform detection of different blood cells from the test blood cell images and perform the counting of the recognized blood cells. An overview of the proposed work is shown in Figure 2 for the detection and counting of the blood cells.

CNN is named after a mathematical operator known as convolution. In the CNN architecture with multiple numbers of layers stacked on it, the convolution layer extracts features from the input image dataset by applying filters on it. Computational performance is enhanced with a max-pooling layer that reduces the dimensionality of layers. The final layer for performing classification tasks is known as the fully connected layer. CNN consists of these layers, and to perform classification and detection tasks, internal parameters of these layers are adjusted.

Instead of developing a deep learning model from scratch, a more coherent approach is to design the architecture utilizing already established CNN models. Therefore, Darknet-53 architecture is used for this study as a backbone network that increases the speed of object detection from the input image. Darknet-53 has 53 layers trained

Note: derived from research.

Figure 2 Methodology of the blood cell detection. 

Note: derived from research.

Figure 3 Architecture of the proposed model on processed data. 

on ImageNet datasets (^{Krizhevsky et al., 2017}). Then, these layers were stacked with 53 more layers of YOLO-v3 algorithm to perform the detection task. The proposed work has a state-of-the-art architecture design for blood cells detection. Figure 3 shows the schematic diagram of the proposed work with processed data.

The input blood cell images are divided into N X N grids. Grid cells are responsible for detecting objects if the centers of the objects lie in those grid cells: they predict bounding boxes and determine the confidence score associated with those boxes. YOLO-v3 calculates bounding boxes on three different scales like features pyramid network (^{Lin et al., 2017}), and these prediction results are more significant for detecting minor-sized blood cell targets. The algorithm was already trained on COCO datasets ^{(Lin et al., 2014}) and implemented using pre-trained weights.

The input blood cell images are divided into N X N grids. Grid cells are responsible for detecting objects if the centers of the objects lie in those grid cells: they predict bounding boxes and determine the confidence score associated with those boxes. YOLO-v3 calculates bounding boxes on three different scales like features pyramid network (^{Lin et al., 2017}), and these prediction results are more significant for detecting minor-sized blood cell targets. The algorithm was already trained on COCO datasets ^{(Lin et al., 2014}) and implemented using pre-trained weights. Predictions were made on three different scales in the proposed work, as shown in Fig. 3. Thus, an input smear image of 416 x 416 dimension was divided into grids of 13 × 13, 26 × 26, and 52 × 52 for the respective stride values of 32, 16, and 8. The confidence score describes the confidence of the model that the object lies in the box and the accuracy of the box it predicted. Each grid in the input image predicts B bounding boxes with confidence scores and C class conditional probabilities. The confidence score formula is given in equation 1:

where represents the common value between predicted and reference bounding box and pred(Obj) = 1, if the target is in the grids, otherwise it would be 0.

Experiment and Results

A convolutional neural network needs to be trained and tested to develop a framework for blood cell detection. For this purpose, an Intel Xenon processor with 64 GB RAM and an NVIDIA Quadro P600 graphics processing unit (GPU) with 24 GB graphics memory are used. Image pre-processing, training, and testing are done on Anaconda3 (Python 3.7), and other libraries, such as TensorFlow and OpenCV, required to train the model. The statistics of the number of images in the datasets are given in Table 1.

A dataset from all sources is divided for training, testing, and validation in 70%, 15%, and 15%, respectively. The deep learning model is trained with blood cell images containing 4888 cells for several iterations until the loss becomes saturated. Generated trained models are analyzed with multiple images in test datasets to obtain overall performance. Yolo-v3 based convolutional neural network is trained for 400 epochs with an initial learning rate of 10-3 and IoU of 50%.

The training is done for different batch sizes of 4, 8, and 16 images. Its steps in each epoch are 512, 256, and 128 while training in a batch size of 4, 8, and 16, respectively. Once trained, models are tested with never-seen blood cell smear images. Some of the tested results are shown in Figure 4, where various blood cells are detected with a higher percentage using the proposed framework.

Average precision (AP) is a measure that integrates recall and precision for good results. It splits into 4 kinds to perform a binary classification on the image: true positive TP, false negative FN, true negative TN, and false positive FP, according to the fusion of predicted class and labeled class of the proposed model. Average precision is the mean of the precision scores after every significant document is retrieved. The formula to calculate precision and mean average precision is given in equation 2:

where Q is the number of queries. Table 2 shows the AP value of trained objects.

Note: derived from research.

Figure 4 Result of the generated detection model. 

Different batches achieve the highest accuracy for different cells. Batch size 8 gives the highest AP for WBC and RBC. Instead, batch 16 returns the highest AP for platelets and the highest mAP value that is 85.35% on 0.5 IoU. The performance of the deep learning model depends on the detection and counting of different cells from the input blood cell image. The statistical analysis is done to count different blood cells where the counting of RBC, WBC, and platelets is compared with the ground truth value of different samples. For WBC counting with all three trained CNN models, accurate results reached 100%. The counting range for RBC is -6 to 7 for different batches. Table 3 shows the counting error (CE) for different blood cells in various error ranges; it also displays most test images that gave zero errors.

The performance of model trains in a batch size of 16 gives the best result with the least counting errors and the highest mAP value with 85.35% higher than the models used in (^{Alam & Islam, 2019}).

The computation time to process the image frames having dimensions of 640×480 pixels is 0.111 seconds, and about 9 frames can be processed per second using the NVIDIA Quadro P600 GPU. Blood cell detection and counting can be done fast and accurately.

The YOLO-v3 detects the blood cells in three different scales. The system can work in real time and give accurate outputs, especially WBCs. This system can help pathologists and medical professionals with rapid analysis to provide recommendations and consultation to patients based on the generated report.

Conflict of Interest

The authors declare no competing interests.

Author contribution statement

All the authors declare that the final version of this paper was read and approved. Total contribution by the authors towards conceptualization, preparation, drafting, and editing of this article is as follows: R.C.J. 25%, S.Y. 25%, M.K.D. 25%, and C.M.T.G. 25%.

Data Availability Statement

Data sharing is not applicable since no new data was created or analyzed in this study. The data supporting the results of this study will be made available by the corresponding author, M. K. Dutta, upon reasonable request.