An essential component of illness diagnosis and therapy planning is blood cell analysis. Precise categorization of blood cell types, including neutrophil, lymphocyte, monocyte, and eosinophil, aids in focused therapy and individualized care by offering important insights into a variety of illnesses and disorders. Manual blood cell categorization, however, may be laborious, arbitrary, and prone to human mistakes. Therefore, automated methods that effectively categorize blood cell pictures are required in order to provide prompt and precise diagnosis [1].

Our proposed system model integrates the IoMT devices, data transmission protocols, preprocessing techniques, and deep learning models to enable efficient blood cell image transmission and classification. The model shown in Fig>(1) is designed to automate the process of capturing blood cell images, transmitting them securely to a centralized location, preprocessing the images for enhancement, and classifying them into Eosinophil, Lymphocyte, Monocyte, and Neutrophil types.

The IoMT device we utilize is a microscopic camera designed specifically for medical imaging purposes. This device integrates a high-resolution camera with microscopic capabilities, allowing for detailed visualization of microscopic objects, including blood cells. The camera is compact and portable, making it suitable for point-of-care applications and remote healthcare settings. With this microscopic camera, we capture high-quality images of blood cells, enabling detailed analysis and classification. The device offers adjustable magnification levels, focus control, and lighting options to optimize image quality and clarity. It is equipped with advanced imaging technologies, such as image stabilization and autofocus, ensuring accurate and sharp images even during dynamic movements or vibrations.

Table 1 provides an overview of the technical details specific to the IoMT device used in our system, which is a microscopic camera. It highlights the sensor technology, connectivity options, power management features, data security and privacy considerations, as well as data processing and analysis capabilities. These details demonstrate the key aspects of the microscopic camera and its role in capturing high-resolution images of blood cells for further analysis and classification in the proposed IoMT-based blood cell image transmission and classification system.

Fig. (1). IoMT-based blood cell image transmission and classification system.

3.6. Explainability and Interpretability

Our novel CNN architecture, shown in Fig>(2), focuses on providing explainability and interpretability for the classification decisions made by the model. We incorporate visualization techniques, such as class activation mapping or saliency maps, to highlight the important areas of the image that contribute to the classification decision. This enables clinicians and researchers to advance insights into the features and characteristics that the model utilizes for classification, enhancing trust and understanding.

The novelty of our approach lies in the combination of multi-scale feature extraction, attention mechanism, transfer learning, domain adaptation, and explainability techniques specifically tailored for blood cell image analysis. This unique CNN architecture not only improves the accuracy and efficiency of blood cell image classification but also provides insights into the reasoning behind the model's decisions. The proposed novelty addresses the challenges in accurately classifying blood cell images, paving the way for advanced diagnostics, personalized medicine, and improved patient care.

3.6.1. Input Representation

Let X represent an input blood cell image. This image is a matrix, and we can represent it as:

(1)

where H is the height, W is the width, and C is the number of channels, e.g., for a grayscale image C = 1 and for a color image C = 3.

3.6.2. Convolutional Layers

Let f_i denote the i^th filter or kernel in the convolutional layer. The convolution operation * between X and fi is represented as:

(2)

where j and k are the spatial indices, and m and n are the filter indices.

Fig. (2). Multi-scale attention-based cnn for blood cell image classification.

3.6.3. Activation Functions

Let A_i denote the activation map after applying the activation function e.g., ReLU, to the convolutional output. The activation function sigma can be represented as:

(3)

3.6.4. Pooling Layers

Pooling reduces the spatial dimensions. Let P_i denote the i^th pooled feature map obtained by max-pooling or average-pooling A_i . The pooling operation can be represented as:

(4)

3.6.5. Attention Mechanism

The attention weights α_i,j,k for the i^th filter at spatial location j, k can be computed using an attention mechanism:

(5)

where θ_i,j,k is a learnable parameter associated with the attention weight.

3.6.6. Fully Connected Layers

Let F be the set of features obtained from the last layer, and W_fc represents the weights of the fully connected layer. The output of the fully connected layer can be represented as:

(6)

3.6.7. Softmax Layer

The softmax function softmax(.) is applied to obtain the probability distribution over the blood cell types:

(7)

where K is the number of blood cell types.

This represents a simplified mathematical model of the CNN architecture used for blood cell image classification. Table 2 summarizes the components of the CNN structure proposed for blood cell image classification. It provides a brief description of each component, highlighting its role in the network's functionality. Together, these components enable the CNN to effectively extract features, classify blood cell images, and provide insights into the reasoning behind the classification decisions.

Table 2. Multi-Scale Attention-Based CNN description

Component	Description
Input Layer	Receives blood cell images as input.
Convolutional Layers	Perform multi-scale feature extraction with parallel pathways using different filter sizes.
Activation Functions	Apply ReLU activation to introduce non-linearity and enhance the network's ability to learn patterns.
Pooling Layers	Downsample the feature maps while retaining important information through max pooling or average pooling.
Attention Mechanism	Dynamically focus on relevant regions of the blood cell images by assigning attention weights.
Fully Connected Layers	Act as a classifier and learn high-level representations by combining extracted features.
Dropout and Regularization	Mitigate overfitting through dropout layers and control model complexity with regularization techniques.
Output Layer	The> layer with a softmax activation function provides probabilities for each blood cell type.
Explainability and Interpretability	Incorporate visualization techniques to highlight important regions for explainability and interpretability.

Fig. (3). Transmission of blood cell images to cloud via wireless protocol in IoMT microscopic camera.

Table 3. Dataset with additional cell type labels.

Dataset	Type	Total Images	Eosinophil	Lymphocyte	Monocyte	Neutrophil
'dataset-master'	Original Images	410	-	-	-	-
'dataset2-master'	Augmented Images	2,500	750	750	750	250
Total	All Images	2,910	750	750	750	250

This section describes the experimental setup conducted for our blood cell classification work. It encompasses dataset acquisition, preprocessing, data augmentation, and the utilization of deep learning models. The methodology is visually depicted in Fig, (3).

4.2. Data Augmentation

Data augmentation methods are employed to increase the diversity and size of the dataset, mitigating overfitting and improving model generalization. Various image transformations, such as rotation, flipping, shifting, and zooming, are applied to generate additional augmented images from the original dataset. Augmented images are created with different variations and added to the dataset to provide a richer training experience for the deep-learning models. Data augmentation involves various transformations applied to an original image X to generate augmented images X_aug. Let T denote a transformation function. The augmented image can be represented as:

(8)

4.3. Deep Learning Models

CNNs are utilized for the classification of blood cell images. The proposed CNN architecture, such as the Multi-Scale Attention-Based CNN, is implemented using deep learning frameworks in MATLAB, such as the Deep Learning Toolbox. The CNN architecture consists of multiple convolutional layers for feature extraction, pooling layers for down sampling, and fully connected layers for classification. Hyperparameters, including the number of layers, filter sizes, pooling types, activation functions, and regularization techniques, are determined based on experimental design and model selection. Let f represent the CNN model with parameters W (weights). The CNN architecture processes an input image X through its layers, yielding an output prediction Y:

(9)

4.4. Experimental Workflow

The dataset is split into training, validation, and testing sets, typically using a 70-15-15 ratio, respectively. The CNN model is trained on the training set using an optimization algorithm (e.g., Adam optimizer) and a defined loss function (e.g., categorical cross-entropy). The validation set is used for hyperparameter tuning, model selection, and early stopping to prevent overfitting. Model performance is evaluated using the testing set, measuring metrics such as accuracy, precision, recall, and F1 score. Let D represent the original dataset. The dataset is split into training (D_Train), validation D_Val, and testing D_test sets using a ratio:

(10)

4.5. Performance Evaluation and Analysis

The performance of the trained CNN model is compared with baseline models or existing methods to assess its effectiveness in blood cell classification. Evaluation metrics, including accuracy, precision, recall, and F1 score, are calculated to quantify the model's performance. Visualizations, such as confusion matrices and classification heatmaps, are generated to gain insights into the model's predictions and identify potential areas of improvement. By following this experimental setup, we aim to train and evaluate the proposed deep-learning models for blood cell classification using the BCCD dataset. The results obtained from this setup will validate the effectiveness and accuracy of our proposed approach, contributing to the advancement of automated blood cell analysis and diagnosis. Performance metrics such as accuracy, precision, recall, and F1 score can be calculated using standard formulas:

The IoMT microscopic camera captures high-resolution images of blood cells, including Eosinophil, Lymphocyte, Monocyte, and Neutrophil types, which are then seamlessly transmitted to the cloud-based infrastructure using a wireless protocol. In this depiction, Fig. (3) showcases the wireless transmission process from the IoMT device to the cloud. The microscopic camera, integrated into the IoMT system, captures detailed images of blood cells, representing each category distinctly. These images shown below in the figure are then securely and wirelessly transmitted through a designated wireless protocol, ensuring efficient and reliable data transfer. This also serves to emphasize the crucial role of wireless communication in IoMT systems, enabling healthcare professionals and researchers to access, analyze, and interpret blood cell images remotely, fostering collaborative diagnostics and facilitating timely medical interventions.

Upon reaching the centralized location, the transmitted blood cell images captured by the IoMT microscopic camera are directed toward a proposed classification model for accurate and efficient classification. At the centralized location, a sophisticated infrastructure awaits the arrival of the transmitted images. These images are meticulously stored and processed, ensuring their integrity and security throughout the classification process. The proposed classification model, specifically designed for blood cell analysis, employs advanced machine learning algorithms and deep neural networks to classify the blood cells into their respective categories. The classification model scrutinizes the intricate details and unique characteristics of each blood cell type present in the received images. By leveraging the power of artificial intelligence, the model identifies and categorizes the blood cells accurately, distinguishing between Eosinophils, Lymphocytes, Monocytes, and Neutrophils. This centralized approach to blood cell classification offers several advantages. It enables the utilization of robust computing resources, such as high-performance servers and powerful GPUs, to handle the computational demands of the classification model. Additionally, the centralized location provides a centralized database of classified blood cell images, which can be further utilized for research, analysis, and reference purposes. By employing cutting-edge technology and efficient data management strategies, the proposed classification model plays a pivotal role in enhancing the accuracy and speed of blood cell analysis. The seamless integration of the IoMT device, wireless transmission, and the centralized classification model creates a comprehensive and efficient system for blood cell analysis, ultimately facilitating better healthcare outcomes and advancing medical research Fig. (4).

The figure encapsulates the outcomes of blood cell classification through a confusion matrix, providing a detailed account of correct and incorrect classifications for Eosinophil, Lymphocyte, Monocyte, and Neutrophil. Examining the matrix reveals specific counts for each class: For Eosinophil, 700 samples were correctly classified, with 30 instances misclassified as Lymphocyte, 10 as Monocyte, and 10 as Neutrophil. In the case of Lymphocytes, 700 were correctly classified, along with 40 misclassified as Eosinophil, 5 as Monocyte, and 5 as Neutrophil. Monocytes had 725 correct classifications, 10 were misclassified as Eosinophil, 5 as Lymphocyte, and 10 as Neutrophil. Neutrophils had 235 correct classifications, with 5 misclassified as Eosinophil, 5 as Lymphocyte, and 5 as Monocyte. This numerical breakdown within the confusion matrix allows for a comprehensive analysis of classification performance, aiding in the identification of patterns and areas for model improvement.

The Confusion Matrix Heatmap shown in Fig. (5) provides a visual representation of the classification results in the blood cell classification. Each cell in the heatmap corresponds to a specific combination of true and predicted classes, and the color intensity represents the proportion or percentage of samples. The heatmap allows for an easy interpretation of the confusion matrix, highlighting patterns, misclassifications, and class imbalances, thereby aiding in the evaluation and understanding of the classifier's performance for different blood cell types.

In Fig. (6), a bar plot vividly illustrates the performance metrics derived from the blood cell classification model's evaluation. The heights of the bars directly convey the model's effectiveness across various metrics for each blood cell type. Notably, the overall accuracy stands at approximately 97.21%, reflecting the proportion of correctly classified samples across Eosinophil, Lymphocyte, Monocyte, and Neutrophil. Examining precision, Eosinophil demonstrates a precision of about 92.23%, Lymphocyte at 93.62%, Monocyte excels with 97.29%, and Neutrophil maintains a precision of 91.36%.

Furthermore, recall values for Eosinophil, Lymphocyte, Monocyte, and Neutrophil are approximately 91.36%, 93.62%, 96.77%, and 94.64% respectively. The F1 score, a harmonic mean of precision and recall, provides a comprehensive measure of classification performance. Eosinophil achieves an F1 score of around 91.79%, Lymphocyte at 93.62%, Monocyte with an impressive 97.03%, and Neutrophil maintains a robust 92.97%. This visual representation facilitates a quick and insightful comparison of the model's performance across different blood cell types, offering valuable insights for targeted model refinement and improvement.

Fig. (7) presents a boxplot encapsulating the distribution of performance metrics across Eosinophil, Lymphocyte, Monocyte, and Neutrophil classes in the blood cell classification model. The boxplots provide a visual summary of key metrics, including precision, recall, and F1 score. For Eosinophil, the boxplot reveals a median precision of approximately 92.23%, a median recall of 91.36%, and a median F1 score of around 91.79%. In the case of Lymphocytes, the median precision stands at 93.62%, recall at 93.62%, and F1 score at 93.62%.

Fig. (4). Confusion matrix for blood cell classification.

Fig. (5). Confusion matrix heatmap for blood cell classification.

Fig. (6). Performance parameters by class in blood cell classification.

Fig. (7). Performance metrics boxplot by class in blood cell classification.

Fig. (8). Graphical user interface (GUI) process of analyzing blood cell images.

Regarding Monocyte, the boxplot displays a median precision of 97.29%, recall at 96.77%, and F1 score at 97.03%. Neutrophil exhibits a median precision of 91.36%, recall at 94.64%, and F1 score at 92.97%. These boxplots offer a comprehensive view of the variability and central tendencies of performance metrics, aiding in the identification of class-specific patterns and potential areas for model enhancement in blood cell classification.

We have developed a Graphical User Interface (GUI) shown in Fig. (8) with three buttons to streamline the process of analyzing blood cell images. This GUI was created using MATLAB and incorporates various functionalities to enhance and classify the image samples. The transmit button serves as the transmission button, allowing users to upload and send the blood cell image sample to the application. By clicking this button, the user can select the desired image and initiate the processing pipeline. The receive button plays a crucial role in receiving and processing the uploaded image sample. Once the image is transmitted, this button triggers the application to perform various image enhancement techniques. These techniques could include noise reduction, contrast adjustment, and image normalization, among others. The aim is to improve the quality and prepare the image for subsequent analysis. The process button executes the proposed classification CNN model on the processed blood cell image. By clicking this button, the trained CNN model is deployed, and the image is fed into the network for classification. The model uses its learned features to identify and classify the blood cell type accurately.

Once the classification process is complete, the GUI displays the result, providing information about the predicted class for the given blood cell sample. Our GUI, developed in MATLAB, provides an intuitive and user-friendly interface for transmitting, processing, and classifying blood cell image samples. By incorporating a trained CNN model, we ensure accurate and efficient classification, enabling quick and reliable analysis of blood cell samples for various research or diagnostic purposes.

Our MATLAB GUI, integrated with IoMT and deep learning, facilitates remote blood cell image transmission and classification, vital for disease diagnosis and treatment planning. Portable IoMT devices enable high-quality imaging at the point of care, supporting remote diagnosis, especially in underserved areas. Secure image transmission to a centralized location allows timely interventions. Automated deep learning classification ensures reliable, efficient results, empowering healthcare professionals. Diagnostic reports offer actionable insights for personalized treatment plans. This system holds the potential to transform healthcare, improving diagnostic accessibility, early disease detection, and patient outcomes through IoMT and deep learning advancements. Continuous evaluation and validation are crucial for enhancing system reliability and effectiveness.