The goal of a routine blood examination is to measure the number and concentration of blood cells in a patient’s blood in order to determine, among other things, if the patient has anemia or an infection. Therefore, high-precision blood cell detection technology is essential for human health diagnostics [1]. There are five different types of white blood cells, neutrophils, eosinophils, basophils, lymphocytes, and monocytes. There are now two available types of cell detection and counting technologies. Traditional image processing techniques like automatic cell recognition technology based on Hough transform [2], watershed method based on distance transform [3], and method based on threshold [4] were used in the early stages, but they will not be accurate enough for real-world applications, especially in areas with a lot of cell overlap. In the earliest stages of pre-blood cell identification, which is divided into two categories, traditional image processing technology was mostly used. The most prevalent technique was hand-staining, followed by manual inspection and classification under a microscope [5]. This method requires examiners to have a comprehensive understanding of cell morphology and to have mastered the capacity to spot and recognize patterns via consistent study and practice. Due to the vast amount of microscopic image data and the diversity of blood cell morphology, traditional blood cell identification methods are becoming more unsuitable for addressing the various complex challenges confronted by blood cell analysis inspectors. Many effective deep learning-based object detection algorithms, such as R-CNN, Mask-R-CNN [6], YOLO [7], and RetinaNet [8], have been presented in recent years, and they are now frequently employed in blood cell detection applications.

Francois Chollet proposed CNNs that are superior to Inception by substituting depth-wise separable convolutions. An experiment with a larger dataset of 350 million images and 17,000 classes for image classification [9] shows that the Xception can improve the performance of CNNs. Christian et al. scaled networks using correctly factorized convolutions and aggressive regularization [10]. He et al. [11] introduced a residual learning framework and evaluated residual networks with up to 152 layers using the ImageNet dataset. Han et al. [12] proposed a three-step strategy to eliminate superfluous connections to circumvent the issues that neural networks are difficult to deploy on embedded systems and that traditional neural networks can correct the design prior to training. Start by training the network and identifying the vital links. Second, eliminate any unnecessary connections. The network must then be retrained. Szegedy et al. [13] proposed the Inception architecture for deep convolutional neural networks. In addition, the Hebbian principle and multi-scale processing intuition were used to maximize quality in design decisions. Ma et al. [14] created the novel ShuffeNet V2 architecture to analyze the direct metric on the target platform. Wang et al. [15] proposed an end-to-end training method for a residual attention network employing an attention mechanism and a hybrid feed-forward network architecture. The proposed network achieves state-of-the-art object performance on three benchmark datasets, including CIFAR-10, CIFAR-100, and ImageNet. Zoph et al. [16] made a new search space called the “NASNet search space” and a new regularization method called Scheduled DropPath to make the NASNet models more general.

Hegde et al. [17] categorized white blood cells using both traditional image processing techniques and deep learning methods. The output of the classifier was evaluated using a neural network with hand-crafted features that had an average accuracy of 99.8 percent. In addition, white blood cells were classified with approximately 99% accuracy using CNN and transfer learning. Kutlu et al. [18] utilized Regional Convolutional Neural Networks (R-CNN) and merged AlexNet, VGG16, GoogLeNet, and ResNet50 architectures to recognize WBC cells. When the sample size is small, Baydilli et al. [19] classified WBC using capsule networks. Shahin et al. [20] proposed transfer learning using deep activation characteristics and fine-tuning of existing deep networks for the identification of WBCs. In addition, “WBCsNet,” a novel end-to-end convolutional deep architecture, was proposed and utilized to balance the categorization of the WBCs dataset. Sharma et al. [21] created a deep learning system that uses convolutional neural networks to make the whole process of classifying white blood cells (WBC) automatic. Sahlol et al. [22] proposed a more effective hybrid technique for classifying WBCs. The hybrid strategy consisted primarily of two steps, Initially, features were extracted from WBC images using VGGNet. The collected features were then filtered using the Salp Swarm Algorithm (SESSA). Zhang et al. [23] added the deformable convolution layer to the standard U-Net structure to deform U-Net (dU-Net) and improve the accuracy of RBC classification in order to further address the low image quality resulting from noise and artifacts. Iqbal et al. [24] utilized machine learning techniques such as logistic regression, support vector machine (SVM), and CNN to differentiate between healthy and diseased cells. Kassim et al. [25] introduced the RBCNet pipeline, which consists of a U-Net first stage for cell-clusteror superpixel segmentation and a faster R-CNN second refinement stage for recognizing tiny objects within connected component clusters. Jiang et al. [26] proposed a new deep learning method called attention-YOLO, which adds a channel attention mechanism and a spatial attention mechanism to a feature extraction network in order to effectively realize the automation of blood count and solve the problems of inadequate positioning of the bounding boxes in the YOLO detection method. Compared to RBC cell classification, Aliyu et al. [27] proposed a method to classify abnormalities based on images of RBCs with deformed shapes using SVM and Deep learning. Boldú et al. [28] used pre-trained CNNs (VGG16, ResNet101, DenseNet121, and SENet154) to adjust their layers to datasets, after which they selected the optimal architecture and created a system with two parallel modules (ALNet). Mohamed et al. [29] classified WBC using ten distinct pre-trained models and six distinct machine learning techniques. Reference [30] proposed geometric-feature spectrum ExtremeNet (GFS-ExtremeNet), a geometry-aware deep-learning technique for cell detection. Khouani et al. [31] developed an approach based on deep learning for automatically recognizing white blood cells in images of peripheral blood and bone marrow. In particular, the application of combined predictions and corrections increased model outputs, as did the use of a new technique that improved segmentation quality by collaborating model results with geographical information. Tran et al. [32] used cutting-edge deep learning semantic segmentation techniques to separate red and white blood cells in blood smear images.

The rest of this work is arranged as follows, Section 2 contains the deep learning classification model and model classification assessment criteria, such as model architecture design and analysis, and Section 3 contains the model optimization approach. Section 3 presents a feature distribution study of the four blood cells as well as a comparison of nine deep learning models for classification. The Section 4 of this paper brings the paper to a conclusion.

On numerous occasions, deep neural networks outperform conventional machine learning techniques. Complex nonlinear relationships between input and output variables are modeled by DNNs. Given sufficient data, deep neural networks can provide a reliable solution for hemodynamic classification tasks. CNNS learns features implicitly from training data. Because neurons on the same feature map surface have identical weights, the networks can be trained concurrently. The structure of local weight sharing reduces network complexity and provides unique benefits for image classification tasks. Resnet’s residual structure enables the network to be deeper, to converge more rapidly, and to be optimized with fewer parameters and less complexity than earlier models. Based on InceptionV3, Xception replaces the Inception module with depth-separable convolutions and combines Resnet skip-connection structures. This expedites the model’s convergence and increases its precision. However, the computational process of the InceptionV3 network structure is relatively dispersed, and the performance of existing convolutional neural networks for classification tasks is inadequate. SqueezeNet is a traditional lightweight network, and on the ImageNet dataset, AlexNet can achieve comparable results with 50 times fewer parameters than SqueezeNet. GoogleNet includes an additional classifier during training, which can mitigate the effects of vanishing gradients. Once the model has been trained, the supplementary classifier is no longer required. Additionally, the fully connected layer and a global average pooling layer are utilized to reduce the number of parameters in the model, allowing it to be successfully applied to image classification tasks. SuffleNetV2’s primary function is to alter the data flow using channel shuffling. It reduces the number of parameters and computation required for small models. Its disadvantage is that it creates a boundary effect in which only a small portion of the input channel is used to generate the output channel. Attention 56 performs better when the output sequence and input sequence are not in the same order. NASNet can be applied to arbitrary classification tasks by stacking network units. NASNet is essentially a more complicated version of Inception. We think that these traditional neural networks can do a good job of analyzing how blood moves through the body.

In this paper, a multi-deep learning model-based classification analysis for hemodynamics is proposed. This paper uses a total of 12,444 images, which are divided into training sets, verification sets, and test sets, totaling 7965, 1992, and 2487 images, respectively. These images are categorized as eosinophils, neutrophils, monocytes, and lymphocytes. All datasets are subjected to uniform data preprocessing, including normalization distribution testing. The CNNs, Xception, InceptionV3, ResNet50, Squeeze, GoogleNet, SuffleNetV2, Attention56, and NasNet deep learning models are used for the four categories. The performance of the proposed approaches is measured by the TP, TN, FP, FN, accuracy, and model loss.

Fig. 1 illustrates the process flow for this paper. In the classification task’s overall architecture, the original data set is processed before the training set, validation set, and test set are proportionally divided. The processing networks CNNs, Xception, InceptionV3, ResNet50, Squeeze, GoogleNet, SuffleNetV2, Attention56, and NasNet receive the processed data. Then, visualize the training epoch’s curve using the loss and accuracy values obtained during model training. At the conclusion, each model’s test results were evaluated using a confusion matrix.

This section contains an evaluation of feature distribution and performance. Visulization examines the typical distribution of the four sample categories. To make sure that the final classifications are correct, the performance of the eight deep learning models is compared to that of the benchmarking methods.

In this study, nine models of deep learning were employed to classify four types of blood cells, eosinophils, neutrophils, monocytes, and lymphocytes. CNNs, Xception, InceptionV3, ResNet50, Squeeze, GoogleNet, SuffleNetV2, Attention56, and NasNet are among these models. Also presented are the statistical data derived from the aforementioned deep learning model, as well as the experiments based on 12,444 blood cell images and the experimental parameters setup. ResNet50, NasNet, and Xception achieve minimal model loss of 0.0080, 0.0098, and 0.0129 in training, validation, and testing datasets, respectively, and NasNet, NasNet, and Xception achieve 98%, 97% and 89% classification accuracy in training, validation, and testing datasets, despite blood cells with varying granulocyte appearance, diameter, and refractive index. Future research will investigate the analysis of clinical datasets utilizing improved techniques.