Identification of Anemia and Its Severity Level in a Peripheral Blood Smear Using 3-Tier Deep Neural Network

Abstract

The automatic detection of blood cell elements for identifying morphological deformities is still a challenging research domain. It has a pivotal role in cognition and detecting the severity level of disease. Using a simple microscope, manual disease detection, and morphological disorders in blood cells is mostly time-consuming and erroneous. Due to the overlapped structure of RBCs, pathologists face challenges in differentiating between normal and abnormal cell shape and size precisely. Currently, convolutional neural network-based algorithms are effective tools for addressing this issue. Existing techniques fail to provide effective anemia detection, and severity level prediction due to RBCs’ dense and overlapped structure, non-availability of standard datasets related to blood diseases, and severity level detection techniques. This work proposed a three tier deep convolutional fused network (3-TierDCFNet) to extract optimum morphological features and identify anemic images to predict the severity of anemia. The proposed model comprises two modules: Module-I classifies the input image into two classes, i.e., Healthy and Anemic, while Module-II detects the anemia severity level and categorizes it into Mild or Chronic. After each tier’s training, a validation function is employed to reduce the inappropriate feature selection. To authenticate the proposed model for healthy, anemic RBC classification and anemia severity level detection, a state-of-the-art anemic and healthy RBC dataset was developed in collaboration with Shaukat Khanum Hospital and Research Center (SKMCH&RC), Pakistan. To evaluate the proposed model, the training, validation, and test accuracies were computed along with recall, F1-Score, and specificity. The global results reveal that the proposed model achieved 91.37%, 88.85%, and 86.06% training, validation, and test accuracies with 98.95%, 98.12%, and 98.12% recall F1-Score and specificity, respectively.

1. Introduction

Reducing red blood cells (RBCs) mass in human blood causes anemia. It is the most common blood disease that occurs worldwide. It is responsible for reducing blood’s oxygen-carrying capacity, which leads to tissue hypoxia [1]. Changes in morphological and textural features, like the reduction in the size of circulating RBCs or diminishing red color, cause microcytic hypochromic anemia (MHA) [2]. Iron deficiency in the human body is the main stimulus of these morphological changes (shape, size, and number of RBCs per unit area) in RBCs. anemia is a global health problem with major consequences related to health, working capacity, and quality of life. According to the World Health Organization (WHO) [3], anemia affects 1.62 billion people, 24.8% of the world's total population, with a financial burden of $3.5 trillion annually worldwide.

Changes in the size of RBC elements is the alarming condition for reducing hemoglobin quantity in blood. The normal range of RBC size is 80 to 100 femtolitre/RBC (fl/RBC). RBC’s normal diameter and thickness range from 6.2 to 8.2 µm and 2 to 2.5 µm, respectively [4]. Macrocytosis (size of RBCs greater than normal reference range) and microcytosis (size of RBCs less than normal reference range) defects depend on the size of RBCs. The reference size of RBCs is expressed by an index called mean corpuscular volume (MCV). The normal size of healthy RBCs in terms of MCV is 80 to 100 fl [5]. RBCs with MCV levels below 80 fl cause MHA anemia, while MCV levels above 100 fl cause macrocytic anemia [6]. Anemic patients can be classified into three categories: severe (with hemoglobin <7.0 g/dL), moderate (with hemoglobin <7.0–8.9 g/dL), and mild (with hemoglobin <9.0–10.9 g/dL), based on WHO classification [5]. The hypochromic condition is also initiated due to the varying hemoglobin concentration in the blood. The hypochromic concentration is indicated by mean corpuscular hemoglobin concentration (MCHC). The reduced MCHC causes hypochromatism in single RBCs from the normal range, i.e., MCHC < 32 [7,8,9]. Many laboratory tests were carried out for the identification of anemia. Some of the tests are listed in Figure 1.

The visual representation of normal RBC size is shown in Figure 2a. However, in microcytic hypochromic anemia, RBC size decreases below 80 fl/RBC [2], as shown in Figure 2b. The visual or morphological analysis of peripheral blood smear in the microscopic view is a major measure for diagnosing blood-related diseases like anemia, leukemia, thalassemia, etc. [10,11]. Most of them become fatal, i.e., anemia, acute lymphoblastic leukemia (ALL), and acute myeloid leukemia (AML), if not treated in time [12]. So far, blood syndrome diagnosis greatly relies upon the hematologist’s and pathologist’s expertise and excellence. During manual blood cell analysis, a pathologist faces several issues. They require outstanding expertise and rich experience for accurate microscopic blood cell analysis. Using a simple microscope, pathologists face challenges when identifying or predicting abnormal cell structures. Pathologists also face difficulties defining the percentage change of infected cells from the normal using a compound microscope. Approximately 5–10 min are required for the examination of one slide. In the case of the overlapped structure of RBCs, pathologists cannot accurately differentiate between normal and abnormal shape and size. Manual analysis can be affected due to the fallible nature of humans in terms of fatigue, etc. Current blood analyzers perform CBC counts but do not facilitate the pathologist to examine the blood cell shape and size. The automatic analyzers also show some limitations for abnormal or neoplastic cell detection [13]. According to [14,15], an automated hematology analyzer cannot recognize abnormal red blood cell shapes and erroneously increase or decrease the results due to interfering factors. Abnormal variation in the size of red blood cells, the abnormal shape of RBC due to the sickle cell, target cell, schistocyte, etc., still needs morphological analysis for a final determination. In [16], the author points out multiple limitations of automatic analyzers in terms of the financial burden. Current blood analyzers perform CBC counts but do not facilitate the pathologist to examine the blood cell morphological structure at the pixel level.

Because of this, it is necessary to build computer-aided diagnostic (CAD) techniques that assist the hematologists with accurate and efficient morphological analysis of blood elements. The key working of the CAD system is to extract the hidden features and attributes of blood elements that are beyond the human visual range. These features include morphological, textural, and color features at the pixel level. Physical and quantitative descriptors like morphology, texture, and color play a crucial role in diagnosing several blood morphology-related diseases [13]. This research aims to develop a CNN-based model for automatically identifying anemic and healthy RBC elements in the microscopic image. This model will also find the severity level of anemic RBC elements by targeting changes in the shape and size of infected RBC. We have developed a state-of-the-art anemic RBC dataset named the Anemic-RBC dataset. The proposed dataset comprises a total of 11,500 images with approximately 750,000 RBCs elements. Out of 11,500 images, 5750 are normal and 5750 are anemic images with manually generated ground truth (binary and pixel-wise). The key contributions of the proposed research work are:

• A CNN-based three-Tier deep convolutional fused neural network (3-TierDCFNet) architecture that performs two-stage classification of RBC images.
• Module-I classifies the input image into two classes, i.e., healthy and anemic images. Module-II detects anemic image severity levels and classifies them into mild or chronic.
• Module-II of 3-TierDCFNet also provides accurate detection of overlapped structures of anemic RBCs.
• We have developed a standalone RBC microscopic image dataset along with manually segmented ground truth images of both healthy and Anemic-RBCs under the supervision of a hematologist/pathologist for cross-match analysis.

2. Related Work

Azam et al. used a grey level matrix and feature bank to classify microcytic hypochromic anemic RBCs [4]. They got an accuracy of 96% with adaptive synthetic sampling and locality-sensitive discriminant analysis (LSDA). In [17], the authors proposed a technique for the automated identification of nucleated RBCs from the peripheral blood smear. They achieved an accuracy of 99.42% for detecting nucleated red blood cells using a multilevel threshold approach marker-controlled watershed algorithm. A total of 950 nucleated blood cells were examined from 50 blood smear images. In [18,19,20], the authors performed simultaneous segmentation of RBCs and WBCs and achieved with 94.8% and 97.2%, 97.45%, 93.43%, and 96.00% accuracies, respectively, using Marker-controlled watershed, modified SegNet model, and deep convolutional network. In [21,22] the author proposed a novel technique for automated diagnosing of several blood diseases like AIDS, iron deficiency, a blood disorder, platelets, malaria, leukemia, and anemia. Shirazi et al. [23] developed a statistical-based thresholding method for the segmentation of RBCs followed by the Fuzzy C-means using the ALL-IDB blood cell dataset. They performed single cell-based RBCs classification and achieved 96% accuracy. In [24], the authors applied the thresholding technique for the segmentation and detection of RBCs corners and recorded an accuracy of 87.9%. This method failed to detect the edges of the image accurately for the overlapped structure of cells. In [25], the author proposed a novel approach for splitting Rouleaux red blood cells from thin blood smears by applying distance transform and local maxima. Three evaluation parameters were used with evaluation rates of true positive rate (TPR) = 96%, error rate (ER) = 4% and accuracy (AC) = 98%. Semantic segmentation was applied by [26,27] using SegNet architecture and VGG- for segmentation and counting of WBCs and RBCs cells element within the ALL-IDB1 dataset, respectively. A capsule network-based model was adopted [28,29] to classify the blood elements in peripheral blood smear to enhance the incremental training procedure. In [30,31,32,33], normal and abnormal RBCs have been detected based on form factor, perimeter, and area features with an accuracy of 94%, but fail with noisy images using random forest (RDF) and support vector machine (SVM) as classifiers. Malarial infected cells have been segmented by [34,35,36] with the help of Otsu’s method using the intensity threshold approach and CNN but failed to segment the overlapped structure of blood cells and blurry images.

In [37,38,39], the authors targeted the shape of red blood cells to identify thalassemic and anemic RBCs using a multilayer perceptron. [37] used texture and color features to classify thalassemia microscopic images and obtained an accuracy of 93.77%. Watershed transform was applied by [34] to detect hypochromic and normochromic anemia with an accuracy of 96.7%. In [40], the author used CNN, KNN, and SVM to detect microcytic hypochromia using CBC and blood film features. In [41], the authors diagnoses three types of anemia, i.e., (1) iron deficiency anemia (IDA), (2) α-thalassemia trait, and (3) β-thalassemia trait using weka software. The author used two parameters, (1) highest accuracy and (2) lowest mean absolute error, to evaluate the data mining technique. In [42], the authors analyzed the classification approaches, i.e., decision tree (KNN, CNN, SVM, Logistic regression) and association rule mining for statistical analysis of anemia. Random prediction (Rp) classification algorithms were used by [43] for the selection of anemia in pregnant women. The authors in [44,45,46] adopted classical approaches, i.e., WEKA, Naïve Bayes, multilayer perception, and J48 algorithms to predict anemia types using CBC reports. The authors in [47] proposed a computer-aided system to diagnose blood disorders like anemia by classifying red blood cells. K-Medoids algorithm was used to extract the white blood cells from the image. The RBCs and WBCs images were collected from the Kasturba Medical College, Karnataka [48], Hemato-pathology laboratory [49], and Internet resources. They obtained an accuracy rate of 98% for the correct identification of anemia. In [50], the authors proposed a deep CNN to classify RBCs in sickle cell anemia. They developed a high-throughput framework that consists of three stages. A total of eight sickle cell disease patients have been selected to collect 7000 single RBCs images. Four various artificial learning methods were used by [51] to classify the general type of anemia. They used Support Vector Machines, Naïve Bayes, and Ensemble Decision Tree for classification purposes on 1663 samples. They achieved the highest accuracy of 85.6% with the Bagged Decision Tree algorithm. The summary of the literature is given in

Table 1. Summarizes the anemia prediction techniques and datasets used for RBC detection. It also shows the relevant test and performance evaluation matrices commonly used for image segmentation and classification.

S #	Paper	Project Aim	Database/ Dataset	Methods	Targeted Features	Performance Evaluation
1	[41]	Diagnosis of (1) iron deficiency anemia, (2) α-thalassemia trait and (3) β-thalassemia trait	793 individuals 184 IDA 200 healthy 203 β-thalassemia 206 α-thalassemia	Weka Software	CBC	Accuracy = 96.343% Mean absolute error = 0.0183%
				Hybrid vote algorithm
				J48, IBK and Naïve Bayes algorithms
2	[42]	Statistical analysis of anemia	NFHS-4	Decision tree	Hemoglobin level	Accuracy with only hemoglobin = 97.35% Accuracy for mother-child relation DT = 44%
				Association rule
3	[43]	Feature selection and computational time for anemia prediction	Dataset with 2120 samples and 19 features	Median vector feature selection	Median Vector Feature Selection	Algorithm accuracy 98%.
				RandomPrediction (Rp) algorithm
4	[52]	Analysis of Anemia Using Data Mining Techniques with Risk Factors Specification	539 anemic patients	Weka Software	CBC, MCV, MCH	Accuracy = 86.1%
				Naïve Bayes, Bayesian Network
				Logistic regression, Multilayer Perceptron
5	[53]	Social determinants of health in anemia classification	6935 instances with 986 variables.	KNN, RF, ANN, SVM	Correlation, Gradient boosting, recursive feature selection	Evaluate the performance of different classification algorithms
6	[54]	anemia disease prediction using CBC test results	200 test samples with seven attributes.	NB, RF, DT algorithm	CBC	Evaluate the performance of different classification algorithms
7	[55]	Hematological data classification	425 samples	RF	CBC	Evaluate the performance of different classification algorithms
				Multilayer Perceptron
8	[56]	Blood diseases detection	668 records	RF, KNN, SVM, DT	Not mentioned	Evaluate the performance of different classification algorithms
9	[47]	anemia diagnosis by RBC classification	1000 images manually collected	K-Medoids algorithm, Modified Watershed algorithm	Area, Perimeter, Diameter, Shape, geometric	Accuracy
	[50]	Classification of RBCs in sickle cell anemia	7000 single RBCs	CNN model	Geometric transformations	Accuracy, Mean evaluation accuracy
	[51]	Classifying anemia types	1663 samples	Support Vector Machines, Naïve Bayes, and Ensemble Decision Tree	HGB and MCV	Classification Error, Area Under Curve, Precision, Recall, and F1-score

.

The literature reveals that most of the current research on blood image processing pays more attention to simple segmentation and classification, especially for WBCs. The present research work generally focused on identifying RBC and WBC elements within an image. They are not targeting the identification of any specific disease. We have targeted the identifying normal and abnormal morphological features of RBC elements considering specific diseases, i.e., anemic RBC. The proposed work helps to count the number of infected RBC elements within an image along with the severity level of anemic images as mild or chronic. The proposed work also paid attention to the separation of the overlapped structure of RBC. In the case of the overlapped structure, pathologists cannot differentiate accurately between normal and abnormal shape and size. The overlapped structure of abnormal RBC elements hides the blood element density and morphological features (shape and size) in a unit volume.

3. Methodology

The process was initiated by preparing blood smear slides from 50 patients at Shaukat Khanum Hospital and Research Center (SKMCH&RC), Pakistan. Out of 50 patients, 25 were non-anemia, and 25 were anemic patients. After that, the Olympus Dp27 8.9-megapixel CMOS sensor captured RBC images with 4 K resolution with a microscopic digital camera. The resolution of the collected microscopic RBC was 1920 × 1080. After pre-processing, the collected images were given as input to the CNN-based model to identify healthy and anemic images. The proposed model was executed on GEFORCE RTX-3060 GPU with 12 GB of RAM using a Python 3.7 and PyTorch v1.10 environment.

After the ethical committee's approval of SKMCH&RC, we collected real-time microscopic images of a peripheral blood smear. These images were used to train and test the proposed CNN-based model to diagnose anemia and its severity level from a digital image like an expert medical consultant. Before the training and testing process, collected images were annotated and manually classified into relevant classes under the supervision of a medical consultant, as shown in Figure 3 and Figure 4. The methodology section comprises the following steps.

• Image collection
• Pre-processing
• Dataset arrangement
• Proposed CNN model architecture
• Loss function

3.1. Image Collection

The proposed system used blood image data collected, technically coordinated, and annotated by consultant pathologists of SKMCH&RC, Pakistan. The blood smear slides were examined under an Olympus Dp27 at a 40× magnification rate. The 8.9-megapixel CMOS sensor was used for image acquisition with 1920 × 1080 pixel resolution. We have collected 500 images from 50 blood smear slides, comprising 250 normal and 250 anemic RBC images. In Figure 3a,b images are collected from patients without anemia, while the images in Figure 3c–f were collected from patients suffering from anemia.

Dataset Preparation

Each image consists of approximately 1500 RBC elements. Thus, a total of 375,000 normal and 375,000 anemic RBC elements were used for training and testing purposes. Clinical pathologist experts critically examined each image to classify it as a normal or anemic image. This classification was used to authenticate the proposed model results for the cross-match analysis, making this model more generic and authenticated.

To identify two classes of RBCs, i.e., Healthy and Anemic-RBCs, 1000 high resolution fine-tuned binary (500), and pixel-level (500) masks of all approximate 750,000 RBCs elements from 500 microscopic images were generated. The generated masks were used as ground truth during the training and testing process of the proposed model. The term binary means that the image contains only two values, 1 or 0. In a binary image, 0 represents the RBC element, while 1 represents the background. The binary images are useful for segmentation. Mostly binary versions of images are used for segmentation validation. We have therefore converted all images into the binary format. Pixel-wise ground truth preserves actual values of the RBC elements and only removes the background. This evidence is helpful for the detection of the severity level of anemic RBC elements. So for the ease of the research community, the binary and pixel-wise masks of original images were also generated. In Figure 4, a,b are binary images of normal RBCs, and in c,d are anemic RBC images. Images in e–h represent pixel-level RBC elements of healthy and anemic-RBCs.

3.2. Pre-Processing

In CNN-based deep learning models, pre-processing is the key step to re-arranging the data for better execution. Unwanted features and the noise in the blood images were removed during this step to prepare them for further processing. Pre-processing involves the following steps, as shown in Figure 5.

• Rescaling of image pixels to sharpen the edges for the separation of the region of interest (ROI) from the background
• Removal of noisy, blurry patterns and detection of RBCs edges
• Enhancement of image quality
• Resizing of the input image according to the underlying model

The collected 500 images have high resolution, i.e., 1920 × 1080, so all the images were cropped into 300 × 300. Consequently, the total image for execution becomes 11,500. The contrast normalization [57] technique was applied to sharpen the RBC edges, which helped to identify overlapped areas. Due to the variation in the pixel range of RGB images, the loss function also varied. To address the variation in the pixel range, the image’s pixels were projected into a standardized (0,1) range to make the total loss of all pixels constant. The rescaling of image pixels was also helpful for the standardization of the learning rate. A second-order derivative Laplacian filter [58] was applied using the local noise estimator function to reduce noisy, blurry patterns. This filter gives the best performance on a 3 × 3 window size with λ = 0.5. The primary objective of the Laplacian filter is the simultaneous detection of horizontal and vertical directions along with edge detection. The quality of the image was enhanced using morphological operations like erosion and dilation. In abnormal images, erosion was applied, followed by dilation to enlarge the ROI for better segmentation. This phenomenon helps to increase the distance between the ROI and the background features, leading to the prominent visibility of the RBC elements.

3.3. Dataset Arrangement

The collected 11,500 original images comprise normal (5750) and anemic (5750) images. The anemic images are further divided into two classes, i.e., mild and chronic, based on disease severity, as shown in

Table 2. Detailed description of original and segmented images used during the training, validation, and testing of the proposed model.

Image Type		Healthy Images	Anemic Images		Total Healthy + Anemic	Original + Segmented
			Mild	Chronic
Original Images		5750	2875	2875	11,500
Manual Segmented Images		5750	2875	2875	11,500	23,000
Training Images	Original	4025	2012	2013	8050
	Segmented	4025	2012	2013	8050	11,500
Validation Images	Original	575	287	288	1150
	Segmented	575	287	288	1150	2300
Test Images	Original	1150	575	575	2300
	Segmented	1150	575	575	2300	4600
Total RBC Elements	Original	375,000	187,500	187,500	~750,000
	Segmented	375,000	187,500	187,500	~750,000	1,500,000

. Manual evidence was also generated of all images under the supervision of an expert pathologist. Thus, the proposed dataset is equipped with 11,500 pairs of images. Every pair contains an original image and respective ground truth. The generated ground truth is used for cross-match analysis and model performance evaluation. The collected images were divided into three data groups: (1) the training group, (2) the testing group, and (3) the validation group, with a 70:20:10 ratio, respectively. Out of 11,500 images, 8050 were used to train the proposed CNN model. A validation test is performed after each tier of training to check the effectiveness of the CNN model learning procedure. For this purpose, a validation group of the dataset comprised of 1150 pairs of images was used, while for testing purposes, 2300 pairs were used. Table 2 gives a detailed description of images used during training, validation, and testing.

3.4. Proposed CNN Model Architecture

This research aims to develop a CNN-based network that examines the microscopic blood image at the pixel level to perform a two-stage classification of anemic and healthy RBC elements. Conventionally, CNN-based models are multi-layered deep architectures that can inspect the microscopic images at the pixel level to extract useful information [59]. In this research, several state-of-the-art CNN models, such as VGG-Net [60], ResNet [61], CliqueNet [62], ResNeXt [63], DenseNet [64], Xception [65], GoogLeNet [66], Inception-V4 [67], NasNet [68], ShuffleNet [69] and Efficient-Net [70] have been investigated for the development of the proposed model. On the basis of preliminary experiments, DenseNet, ShuffleNet, and EfficeintNet were selected for the proposed model. DenseNet examines the images at the pixel level. EfficientNet preserves pixel-wise semantic information for analyzing morphological deformities in blood images. ShuffleNet helps to maintain accuracy with the least computational cost. The schematic diagram of the proposed model is given in Figure 6.

We have developed a condition-based three-tier densely connected fused network (3-TierDCFNet) architecture. The proposed model performs classification on two modules:

• The classification module (Anemic or Healthy)
• The anemia severity detection module (Mild/Chronic).

During module-I, the inputted image is classified as anemic or healthy. If the image is diagnosed as anemic, the anemia severity detection module finds the severity of the anemia disease. The synoptic view of the proposed three-TierDCFNet is described in Figure 7.

3.4.1. Classification Module (Anemic or Healthy)

In the classification module, we fused [64,69,70] in 3-tiers to attain optimized training feature selection with unbiased validation after the execution of each tier. There are n(n + 1)/2 direct connections for N networks. The feature maps of all preceding networks are used as inputs and their own feature map of the current network for each subsequent network, as shown in Figure 8. At tier-I, DenseNet [64] is implemented, which helps to examine the RBC elements at the pixel level. At tier-II, Efficient-Net [70] is applied, which received the feature map of the preceding network as an input along with original input. Tier-II preserves the pixel-wise semantic information for analyzing morphological deformities in blood images to detect healthy and anemic-RBC elements. Efficient-net is 18 times faster than existing models that require 75 times less floating-point operation per second and 79 times fewer parameters, while during tier-III, ShuffleNet [69] maintains comparable accuracy with the least computational power. The primary aim of using fused three-tier architecture is to simultaneously examine the input image with densely connected networks to preserve pixel-level features and semantic information. The extracted morphological features are then used to identify the severity level of the anemia.

The 3-TierDCFNet received a pre-processed original image (Oo) with 300 × 300 × 3 size input to the tier-I with a manually generated ground truth image (Og). The Og image is used for validation purposes. A threshold (Ť) value is set as condition (Ĉ) for the validation (Vo) purpose of each tier’s output. The constant weight of Ť is set as 55%, 70%, and 90% for the first, second, and third tiers, respectively. At the end of each tier training, a validation process is initiated that validates the output.

Suppose that Vo got a value greater than the predefined Ť value of the n^th tier. The output feature map (Ôi) is given as an input to the n + 1 Tier. The Ôi of n^th tier includes morphological features and pixel-level and sematic information extracted by the n^th tier model. On the other hand, if the n^th tier got a Vo value less than predefined Ť, then the Ôi of the n^th tier model is given as an input again to the same tier for more optimal features selection. Mathematically, the whole training and validation process is defined as follows:

$$\begin{gathered} V_o={\mathsf{Ô}}_i \cong \mathsf{Ť} \cong O_g \\ if V_o\ge \mathsf{Ť}, So {\mathsf{Ô}}_i \to n+1 Tier \\ if V_o!\ge \mathsf{Ť}, {\mathsf{Ô}}_i \to n^{\mathrm{th}} Tier \end{gathered}$$

(1)

Here, $\cong$ act as a comparison operator for validation ${\left(V\right)}_o$ process, Ô_i is the output feature map of n^th Tier, Ť is threshold value and O_g is a manually generated ground truth image for validation purposes. A visual look at this process is described in Figure 7.

The stride size was set as 2 with a 3 × 3 filter size on each tier of the 3-TierDCFNet. Classification loss of the model was calculated using categorical cross-entropy loss.

3.4.2. Anemia Severity Detection Module

When an input image is classified as anemic, the anemia severity detection module is activated to detect the severity level of disease in an image. The Module-II classification in Figure 7 shows the anemia detection architecture for anemia-positive images. The severity of anemic RBC was determined using morphological features like shape and size. This study categorizes anemic images into two levels, i.e., mild and chronic. To point out healthy, mild, and chronic morphological features of RBC, the proposed dataset is divided into three classes along with manually annotated labels. Figure 3 shows the annotated healthy, mild, and chronic RBC elements. A scoring mechanism was used for the severity assessment of anemic RBC elements. Three parameters were considered for setting the score range of the severity stages: (1) RBC size, (2) RBC shape, and (3) central white pallor size.

Table 3. Score range for the determination of anemia severity stages, i.e., Mild and Chronic. Column heading “Original diameter” represents the RBC size before magnification, while the column heading “After magnification” represents the size of RBC after magnification.

Cell Type	Healthy		Mild Stage		Chronic Stage
Parameter	Original Diameter	After magnification	Original diameter	After magnification	Original diameter	After magnification
RBC size	7.5 μm	1.2 cm	<6–4 μm Or >9–11 μm	<0.96–0.66 cm Or >1.44–1.76 cm	<4μm Or >11 μm	<0.66 cm Or > 1.76 cm
RBC Shape	Rounded		25–50% change in roundness		>50% change in roundness
Central white pallor size	1.87 μm	0.3 cm	<3–2 μm Or >4.5–5.5 μm	<0.48–0.33 cm Or >0.72–0.88 cm	<2.92 μm Or >8.25 μm	<0.33 cm Or > 1.76 cm

shows the score range of each parameter. The RBC image dataset was prepared using an Olympus Dp27 at a 40× magnification rate for experimental purposes. The magnified size of RBC elements in centimeters (cm) is described in Table 3. Due to the greater intensity of RBC elements in blood, sometimes it shows an overlapped structure that leads to the inappropriate counting of RBC elements using a simple microscope. Euclidean Distance Transform (EDT) [71] was applied to overcome this issue. Before EDT transformation, the images’ background was removed manually. This transformation works on the pixel-level marking of the binary RBC element. The threshold for marking is set as 70. We find the pixel intensity of RBC elements in overlapped areas using this technique. When pixel intensity goes beyond the threshold value, it marks the pixels and separates overlapped areas.

The severity detection Algorithm 1 was applied in Module-II classification to determine the severity of the anemia in the RBC element. This algorithm classifies the anemic images into mild or chronic using the scoring parameters described in Table 3.

Algorithm 1. Pseudocode representation of anemia Severity Detection algorithm

Start # RBC images will be loaded that are classified as anemic during Module-I classification Load Image:  Var Size of cell (S)  Var Roundness of cell (R)  Var Central pallor size of cell (CP)  Var Size Ratio (SR) Var Roundness Ratio (RR) Var Central pallor Ratio (CPR) Calculate:  If Cell Size < 0.96 and Cell Size > 0.66    S = 1 − abs(0.8 − Cell Size) × SR  Else    S = 1 − abs(1.6 − Cell Size) × SR    R = 1 − abs(38.5% − Cell Roundness) × RR    CP = 1 − abs(0.405 − Cell Paller) × CPR # Morphological parameters checking for predicting disease severity level  Mild = S + R +CP  If (Mild > 50%)    Image = “Mild”  Else    Image = “Chronic”  End  End Start  The above algorithm is a detailed description of the severity detection technique. Initially, the images are loaded that are classified as anemic by 3-TierDCFNet.

3.5. Loss Function

The classification loss (ĹCCE) of the 3-TierDCFNet model was calculated using categorical cross-entropy loss (CCE) [72]. Mathematically, CCE loss is calculated using Equation (2):

$${\mathsf{Ĺ}}_{\mathrm{CCE}}=-{\displaystyle \sum }_{i=1}^N{x}_i·\log y_i$$

(2)

Here, $N$ represents the total number of model output,$x_i$ indicates the ground truth of the healthy or anemic image, and $y_i$ represents the i^th output of the model. The minus sign guarantees that the loss becomes lesser when the distributions become closer to each other. The ĹCCE of 3-TierDCFNet is measured at each tier, and feedback is given to the network for optimal feature selection. The ĹCCE is measured consistently unless the network has a value greater than the predefined threshold value. As the network attained the expected Ĺ_CCE value, the training of the i^th tier was stopped, and output was given to the next tier for further processing with an updated threshold value.

4. Results and Discussion

In this section, we present the performance of the 3-TierDCFNet model for the classification and accurate severity detection of RBC elements. The proposed model performs RBC classification and anemia severity detection on two modules: Module-I classifies the input image as healthy or anemic. If an image classifies as anemic, it inputs Module-II to detect the anemia severity level. We emphasize optimum feature selection and preservation of semantic information during module-I to differentiate between healthy and anemic RBC elements. However, Module-II pays attention to finding the rate of change in shape, size, and central white pallor size of anemic RBC elements with respect to healthy RBCs for the detection of anemia severity level. This section is divided into four sub-sections: Training and testing, performance evaluation matrices, RBC classification, and anemia severity detection.

4.1. Training and Testing

The proposed 3-TierDCFNet model was trained and tested on 23,000 (original and manually segmented) images collected in SKMCH&RC, Pakistan. These 23,000 images comprise 1,500,000 RBC elements, including normal and anemic RBCs (Mild and Chronic), as shown in Table 2. Out of 23,000, the system randomly chooses 16,100 images for training and 4600 for testing purposes. The rest of the 2300 images were used to validate the results of each tier of the model. The model was trained on a total of 500 epochs with 200 iterations per epoch. The initial 300 epochs were executed on tier-I, the subsequent 150 epochs were conducted on tier-II, while the last 50 epochs were carried out on tier-II of the model. The initial learning rate was set as 1 × 10⁻⁴ with a minibatch size of 2. A GEFORCE RTX-3060 GPU with 12 GB of RAM was selected for experimental purposes.

4.2. Performance Eveluation Matrices

To evaluate the performance of the 3-TierDCFNet regarding the classification of RBC classes, the following statistical matrics [73] were used:

4.2.1. Accuracy

Generally, accuracy gives the recognition rate of the classifier. It is defined as the ratio between the sum of true positive (TP) and true negative (TP) values over the sum of TP, TN, false positive (FP), and false-negative (FN) values. The mathematical expression for accuracy is shown in Equation (3):

$$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$$

(3)

The tier-I of the proposed model achieves 85.63%, 84.53%, and 81.32% accuracies for training, validation, and test set, respectively. In contrast, tier-II achieved 91.63%, 87.57%, and 86.75% accuracy, respectively. However, Tier-III outperforms Tier-I and Tier-II and achieved 96.85%, 95.28%, and 89.29% accuracy for training, validation, and test set, respectively. Globally, the proposed model achieved 91.37%, 88.85%, and 86.06% accuracy. The confusion matrix of training, validation, and testing accuracies of Module-I classification is given in Figure 9.

Recall (Sensitivity): Recall is the ratio between true positive values and total actual positive values. Mathematically, it is written as shown in Equation (4):

$$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(4)

Recall of the proposed model was 93.49%, 95.98%, 95.87%, and 98.12% for Tier-I, Tier-II, Tier-III, and global, respectively, as shown in

Table 4. Global and individual training, validation, and test accuracies with relevant losses, Recall, F1-Score, and Specificity of 3-TierDCFNet.

Accuracy		Tier-I	Tier-II	Tier-III	Global
Training	Accuracy	0.8563	0.9163	0.9685	0.9137
	Loss	0.1437	0.0837	0.0315	0.0863
Validation	Accuracy	0.8453	0.8756	0.9528	0.8885
	Loss	0.1547	0.1244	0.0472	0.1115
Test	Accuracy	0.8132	0.8675	0.8929	0.8606
	Loss	0.1868	0.1325	0.1071	0.1394
Recall		0.9349	0.9598	0.9596	0.9895
F1-Score		0.9257	0.9562	0.9587	0.9812
Specificity		0.9165	0.9436	0.9634	0.9726

.

4.2.2. F1 Score

To identify the classification worth of the 3-TierDCFNet, we use the F1 score. F1 score considers both precision and recall. The higher the value of the F1 score indicates that the proposed model has a better prediction efficiency regarding classification. Equation (5) describes the mathematical formula for the F1 score:

$$\mathrm{F}1 \mathrm{Score}=2\times \frac{\mathrm{precision}\times \mathrm{recall}}{\mathrm{precision}+\mathrm{recall}}$$

(5)

The 3-TierDCFNet got 92.57%, 95.62%, 95.87%, and 98.12% for Tier-I, Tier-II, Tier-III, and global, respectively.

4.2.3. Specificity

Specificity is the opposite of recall. It calculates the ratio between TN values over the sum of TN and FP values. The mathematical formula for specificity is described in Equation (6):

$$\mathrm{Specificity} =\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$$

(6)

Specificity of tier-I, tier-II, and tier-III was 91.65%, 94.36%, 96.34%, respectively. However, the global specificity was 97.26%.

4.3. RBC Classification

For RBC classification into healthy and anemic categories, we trained the 3-TierDCFNet by fusing three state-of-the-art CNN models, i.e., DenseNet, EfficientNet, and ShuffleNet. This classification acted as a base for further severity detection of anemia. Thus, the role of CNN models has become crucial. The key advantage of this fusion is that DenseNet performs deep analysis on inputted images at the pixel level. Due to the RBC elements’ high density and overlapped structure, we need to preserve the pixel-level semantic information of each image. EfficientNet preserves the semantic information of each image at tier-II, while ShuffleNet maintains accuracy with the least computational power. Individual and global training, validation, test accuracies, and losses of each tier were calculated to evaluate the proposed model.

Accuracy and Loss Convergence

After completing each tier training, three different accuracies were computed, i.e., training, validation, and test accuracy with relevant losses. So a total of twelve accuracies were calculated, three of each tier and three of the whole 3-TierDCFNet model, along with relevant losses shown in Table 4. Results show that all of the tiers outperformed regarding the recall value and got 93.49%, 95.98%, 95.96%, and 98.95% for true positive results, respectively. The recall, F1-score, and specificity of each tier were also given in Table 4.

The confusion matrix of healthy and anemic RBC classification is given in Figure 9, while a visual comparison among training, validation, and test accuracies of each tier is given in Figure 10. The error bars show the difference between accuracy and relevant loss values.

4.4. Anemia Severity Detection

Hematologists face challenges when identifying and predicting abnormal RBC cell structures at the pixel level using a simple microscope due to their dense and overlapped patterns. The dense and overlapped pattern of RBC causes the unsuccessful prediction of severity and changes in anemic cells with respect to the normal cell using a compound microscope. This study’s main objective was to classify anemic RBC elements and detect anemia severity levels to further classify the anemic images into mild or chronic. In 3-TierDCFNet, module-II performs severity detection in anemic RBC elements and classifies input image as mild or chronic. The results obtained during the severity detection module are shown in

Table 5. Performance evaluation of anemia severity detection module in 3-TierDCFNet during testing.

Classes	Accuracy (%)	Recall (%)	F1-Score (%)	Specificity (%)
Mild	96.82	95.96	96.89	95.96
Chronic	98.94	96.83	97.49	96.61

.

Results indicate that our model outperforms the detection of chronic anemia stages because severity detection parameters like shape, size and central white pallor size have more prominent changes in chronic stages compared to healthy RBC. The confusion matrix of severity detection measures is shown in Figure 11.

4.5. Advantages of the Proposed Model

This approach has the following advantages over [61,64,74,75], and other state-of-the-art models.

4.5.1. Impact of 3-Tier Densely Connected Architecture and Validation Function

3-TierDCFNet can preserve morphological features, pixel-level information, and the semantic pattern of images at each tier with an unbiased validation function due to the fusion of three state-of-the-art models. In a three-tier architecture, we implemented densely connected phenomena with n(n + 1)/2 direct connections for N networks. This strategy enhances the optimal pixel-level feature selection and the preservation of semantic information of morphological parameters. The output feature map of the preceding tier is given as input to the next tier, which helps to ignore the irrelevant feature selection that ultimately reduces the computational time of subsequent tiers. The validation function at each tier’s output does not allow the model to proceed with the subsequent tier execution until the training proficiency exceeds the predefined threshold value. At tier-I, DenseNet [64] is implemented, which helps to examine the RBC elements at the pixel level. At tier-II, Efficient-Net [70] preserves the pixel-wise semantic information for analyzing morphological deformities in blood images to detect healthy and anemic-RBC elements. During tier-III, ShuffleNet [69] maintains comparable accuracy with the least computational power. The main objective of using three-tier architecture is to simultaneously examine the input images with densely connected networks to preserve pixel-level features and semantic information. These morphological features are then used to identify the severity level of anemia.

4.5.2. Impact of anemia Severity Detection Module

To overcome the pathologist’s challenge regarding the detection of the change in morphological features of RBC elements, we employed a severity detection module in the proposed CNN model. The disease severity detection module helps the medical practitioners to detect anemia severity (i.e., mild or chronic) conditions based on pixel-level analysis of morphological deformities. The proposed severity detection algorithm analyzed the anemic image at pixel level and considered the RBC overlapped structure. The overlapped structure’s consideration helps to accurately detect densely placed RBCs’ ultimate boundaries to predict morphological disorders.

Various datasets have been released and are publicly available for delineation of RBCs and WBC images. Previously developed blood cell datasets like ALL-IDB-I, ALL-IDB-II [76], extended ALL-IDB [19], BCCD, IUMS-IDB [77], SMC-IDB [78], BS_DB3 [79], Ash bank, BBBC [80] dataset equipped with small no of images with less number of RBCs elements. ALL-ID-I contains only 108 images with just 39000 blood elements. All 108 images have no masks of authentication of the proposed results. Other datasets also lack detailed information regarding manual ground truth for authentication purposes and blood cell elements. Most of the prior approaches perform segmentation and classification only on a single cell at the region level and not the pixel-level. These dataset images show RBCs and WBCs with isolated, clear color variations and separated boundaries among cell constitutes. As a result, analysis is rather easy.

In contrast, the proposed model experiments on the newly proposed RBC dataset, which is more complicated in the following aspects:

(1)

Large number of images, i.e., 11,500.

(2)

Large number of RBC elements in each image, i.e., ~1500, and overall, ~750,000

(3)

In creating real-life scenarios, images are captured with heterogeneous lighting intensity.

(4)

Most of the RBC elements show an overlapped structure.

(5)

Due to the color differences, segmentation and classification become challenging.

Due to the reasons mentioned above, the test accuracy of the proposed model is lower than previously developed models.

5. Conclusions

In this study, we proposed a 3-TierDCFNet that can extract optimum morphological features of the input image. The 3-TierDCFNet comprises two modules: Module-I classifies the input image into two classes, i.e., Healthy and Anemic. However, Module-II is responsible for severity detection in anemic images and classifies it into Mild or Chronic. The proposed model introduces dense connections among three tiers of the model and a validation function at the end of each tier’s training. The validation function validates the accuracy of the preceding tier based on a threshold value. If the accuracy is greater than the predefined threshold value, the output feature map inputs the subsequent tier.

Furthermore, tier-I ensures the pixel-level analysis of the inputted image. Tier-II preserves the semantic information of the RBC element to classify relevant classes and detect severity levels, while tier-III maintains comparable accuracy with less computational resources. To evaluate the proposed model, training, validation, and test accuracies were calculated along with recall, F1-Score, and specificity. The global results reveal that the proposed model accomplished 91.37%, 88.85%, and 86.06% training, validation, and test accuracies with 98.95%, 98.12%, and 98.12% recall F1-Score and specificity, respectively.

6. Future Work

Our model works for RBC image classification and severity detection in the frequency domain. In the future, we will update this model to detect the rate of pixel change in the frequency domain during the life cycle of the healthy stage to chronic stage conversion. This strategy will assist the medical practitioners in predicting how much time will be taken by the RBC element to convert from healthy to mild and then ultimately up to the chronic stage. This prediction will coordinate the medical consultant to guide anemic patients in taking proactive measures regarding disease treatment.