An Explainable Artificial Intelligence Integrated System for Automatic Detection of Dengue From Images of Blood Smears Using Transfer Learning

Dengue fever is a rapidly increasing mosquito-borne ailment spread by the virus DENV in the tropics and subtropics worldwide. It is a significant public health problem and accounts for many deaths globally. Implementing more effective methods that can more accurately detect dengue cases is challenging. The theme of this digital pathology-associated research is automatic dengue detection from peripheral blood smears (PBS) employing deep learning (DL) techniques. In recent years, DL has been significantly employed for automated computer-assisted diagnosis of various diseases from medical images. This paper explores pre-trained convolution neural networks (CNNs) for automatic dengue fever detection. Transfer learning (TL) is executed on three state-of-the-art CNNs – ResNet50, MobileNetV3Small, and MobileNetV3Large, to customize the models for differentiating the dengue-infected blood smears from the healthy ones. The dataset used to design and test the models contains 100x magnified dengue-infected and healthy control digital microscopic PBS images. The models are validated with a 5-fold cross-validation framework and tested on unseen data. An explainable artificial intelligence (XAI) approach, Gradient-weighted Class Activation Mapping (GradCAM), is eventually applied to the models to allow visualization of the precise regions on the smears most instrumental in making the predictions. While all three transferred pre-trained CNN models performed well (above 98% overall classification accuracy), MobileNetV3Small is the recommended model for this classification problem due to its significantly less computationally demanding characteristics. Transferred pre-trained CNN based on MobileNetV3Small yielded Accuracy, Recall, Specificity, Precision, F1 Score, and Area Under the ROC Curve (AUC) of 0.982 ± 0.011, 0.973 ± 0.027, 0.99 ± 0.013, 0.989 ± 0.015, 0.981 ± 0.012 and 0.982 ± 0.012 respectively, averaged over the five folds on the unseen dataset. Promising results show that the developed models have the potential to provide high-quality support to haematologists by expertly performing tedious, repetitive, and time-consuming tasks in hospitals and remote/low-resource settings.

SECTION I. Introduction

Dengue, an Aedes aegypti and Aedes albopictus mosquito-borne illness, emerged as a global health problem in the 1960s [1]. One study on the prevalence of dengue indicates that about half of the global population is in danger, with an annual estimate of 100–400 million infections [1]. Another study estimates that about $3\cdot 9$ billion people are vulnerable to the dengue virus in >100 countries where the disease is endemic [2]. As per a recent report on dengue by WHO, the Americas, Southeast Asia, and Western Pacific regions are dangerously affected, with Asia experiencing 70% of the overall burden [3]. India is among the 30 most highly endemic countries in the world [3]. Despite decades of attempts, a safe and efficacious vaccine or anti-viral drug for dengue is not in place [4]. Although many DENV infections cause mild illness, severe dengue has been the primary cause of morbidity and mortality in the dengue endemic regions. Severe dengue is characterized by severe plasma leakage, fluid accumulation with respiratory distress or shock, severe organ impairment, and severe bleeding [5]. To date, there is no precise treatment in place for dengue, and the death rates can only be reduced by early detection of severe dengue [3]. Several methods are currently available to detect DENV infection, including isolation of the virus, serology methods, or RT-PCR [5], [6]. Medical laboratories in remote/low-resource settings lack specialized resources for the diagnosis of dengue by any means other than serology methods [5]. Among serological techniques, the detection of non-structural protein 1 (NS1) antigen and Immunoglobulin M (IgM)/Immunoglobulin G (IgG) antibodies are commonly used. NS1 antigen-capture ELISA is a simple efficacious diagnostic tool that supplies qualitative positive/negative results [5], [7]. IgM and IgG antibody-capture ELISAs are beneficial in finding out whether the DENV infection is recent or previous [5], [6].

This paper focuses on an alternative method of identifying dengue infection through a digital pathology approach from microscopic PBS images. The PBS analysis is a diagnostically relevant tool for evaluating various hematological disorders [8]. The manual interpretation of PBS through a microscope remains the backbone of hematological diagnostics, even though it is error-prone and time-consuming owing to its invaluable nature [9]. Automation of PBS analysis is a very active field of research that has motivated many researchers [8]. Automation can assist hematologists in yielding accurate and quick results, especially when there are tremendous amounts of samples to analyze. The digitization of PBS images using a digital microscope or a whole slide scanner, combined with the application of Artificial Intelligence (AI) – based tools, makes automated PBS image analysis feasible, limiting human intervention [8], [9].

In recent years, DL, a subset of AI, has been extensively and successfully used for automating various tasks, including healthcare-related tasks [10]. In particular, CNNs have gained popularity and become the main methodology for medical image analysis [10]. Digital pathology is one of the medical imaging areas where extensive use of CNNs is observed [10]. Whole-slide imaging systems are applied to digitize hematopathology/histopathology slides to generate images of high resolution [10]. The digitized slides are processed by incorporating CNN architectures to perform different computer vision tasks, which include classification (e.g., disease recognition), object detection (e.g., cell counting), and segmentation (e.g., nuclei identification) [10]. The ability of CNNs to extract high-level features without human supervision enables them to automatically learn the most discriminative features directly from the image [9], [11]. In particular, CNNs benefit from eliminating the tedious feature engineering process [12]. However, training from scratch demands a significantly huge labeled dataset. The inadequacy of freely accessible labeled images is one of the biggest hindrances in training CNNs for analysis of medical images (including digital pathology images) [10], [12]. An efficient solution in such a scenario is to employ pre-trained CNNs [12].

A pre-trained CNN is previously trained on millions of images from a generic dataset (e.g., ImageNet dataset) for a specific problem and could be employed to fix a fresh problem using a TL plan [10], [12]. TL signifies fine-tuning the model to fix a fresh problem [12]. The Fully Connected (FC) layers with randomly initialized weights are added to the pre-trained base model and trained on the new task-specific dataset during fine-tuning [13]. However, the base model weights are frozen to prevent them from getting updated during training to avert overfitting [13]. Commonly used CNN architectures for the analysis of medical images include AlexNet, VGGNet, ResNet, GoogLeNet, DenseNet, XceptionNet, and SqueezeNet [12]. These CNNs are trained to identify images into 1000 subcategories [14]. The advantages of transferred pre-trained CNNs include better performance, training with limited data, eliminating training from scratch, and speeding up the training process [12].

Lack of explainability is one main limitation of DL models [15], [16]. The logic behind the predictions made by these models is not clearly understood. Despite the excellent performance of the DL models in various healthcare applications, the medical fraternity still does not fully embrace them due to their black-box nature [17], [18]. The XAI technique was introduced to enable the medical fraternity to appreciate the philosophy behind the model predictions [17]. XAI explains the workings of the model and draws the users’ attention to the regions of the image that highly influence the model predictions [17]. The GradCAM XAI technique was built on the original CAM invented by Zhou et al. in 2015 [17], [19], [20]. GradCAM was developed to suit CNN architectures and is, therefore, more popular among DL models [15], [17].

A. Review of Related Literature

In recent years, enormous CNN architectures with TL have been put forward to analyze medical images, including digital pathology image analysis [10]. The following are some recently published articles that adopted pre-trained CNN architectures with a TL strategy for classifying leukocytes from digital microscopic PBS images.

Aziz et al. adopted a DL-based method for leukocyte classification (Munich AML Morphology dataset) from blood smear images. Leukocytes were segmented using K-means clustering in the color space - $\text{L}^{\ast} \text{a}^{\ast} \text{b}$ . The authors performed classification by employing transferred pre-trained AlexNet and ResNet18. Classification accuracies of 93.30% for AlexNet, and 93.85% for ResNet18 were achieved [21]. Roy et al. presented a method for localizing and classifying leukocytes (LISC dataset) using a DL approach. The authors localized the leukocytes by semantic segmentation using DeepLabv3+ and cropped the leukocytes. Further, the cropped leukocytes were classified by employing pre-trained AlexNet with TL. An average classification accuracy of 98.87% [22] was observed. Khaled et al. explored transferred pre-trained CNNs - VGG, ResNet, and DenseNet for classifying different leukocytes (LISC dataset). The cropped leukocytes were augmented by using image transformations and generative adversarial networks (GAN). A classification accuracy of 98.8% using DenseNet-169 [23] was obtained. Li et al. combined GAN with ResNet to classify leukocytes (BCCD dataset). The authors adopted GAN for increasing the training data, transferred pre-trained ResNet for the classification, and reported an accuracy of 91.7% with a modified loss function [24]. Sharma et al. presented a strategy for the automatic leukocyte classification (Kaggle dataset) using pre-trained CNN DenseNet121. With augmentation and transfer learning, a classification accuracy of 98.84% was achieved [25]. Cengil et al. employed pre-trained CNN architectures – AlexNet, ResNet18, and GoogleNet, with TL for automatically classifying leukocytes (Kaggle dataset). ResNet18 yielded the best classification accuracy of 99.83% [26]. Liu et al. put forward a DL technique for the leukocyte classification. Various pre-trained CNNs were employed. ResNet-50 excelled with an average classification accuracy of 96.7% over C-NMC, ALL-IDB2, PBC, and LISC datasets [27]. Chen et al. proposed a DL framework by integrating two pre-trained networks - ResNet and DenseNet with an attention system for accurately classifying the cropped leukocytes. An overall classification accuracy of 97.96% and 98.71% for the LISC and Raabin datasets, respectively, was achieved with data augmentation and transfer learning. The GradCAM XAI technique was employed to understand the logic behind the predictions made by the model [28]. Meenakshi et al. adopted deep features extracted from CNNs that are pre-trained for automatically classifying leukocytes. The authors extracted 3,000 deep features, 1,000 each from pre-trained CNNs AlexNet, GoogleNet, and ResNet50. Mayfly Algorithm with Particle Swarm Optimization was used to select the most important 1,000 features. These features were then given to the RNN - LSTM classifier to perform the classification. An overall classification accuracy of 95.25% was achieved [29]. Dong et al. developed a novel ensemble CNN framework to classify the five types of leukocytes. The prediction results of the three transferred pre-trained CNN models VGG16, ResNet50, and InceptionV3 were integrated through the Bagging process. The Gompertz function was incorporated to formulate the combination strategy, which yielded an average classification accuracy of 96.5% with ten-fold cross-validation [30]. Dipto et al. presented a strategy for identifying the types of leukocytes (BCCD dataset) by utilizing a pre-trained vision transformer (VT) and pre-trained CNN VGG19. An overall classification accuracy of 84% and 85% was achieved for VT and VGG19, respectively. However, VT demonstrated significantly faster learning compared to VGG19. The GradCAM XAI technique was integrated to understand the logic behind the predictions made by the model [31]. To classify leukocytes from blood smear images, Bhatia et al. employed various pre-trained DL models, such as DenseNet121, Xception, MobileNetV2, ResNet50, and VGG16. DenseNet121 outperformed with an average classification accuracy of 98.59%. An XAI technique was integrated, thereby leveraging local interpretable model-agnostic explanations (LIME) into the models to make the model’s predictions explainable [32].

The unique and distinctive contribution introduced by this proposed work is an XAI-integrated, computationally efficient deep-learning approach for automatically detecting dengue fever from digital microscopic PBS images. PBS analysis is a gold standard for diagnosing various hematological disorders, including dengue fever. The literature review revealed very few published works utilizing similar methodologies on comparable datasets. Most of the published work associated with automatic dengue detection was centered on tabular datasets containing symptoms/vital signs/blood profile data [33], [34], [35], [36]. Hence, this proposed work can potentially fill this gap in the literature. Only a few articles have been published on the automatic diagnosis of dengue from PBS images [37], [38]. In one of our published works [37], we employed MobileNetV2 (a pre-trained CNN) for extracting features of the lymphocyte nucleus. Further, these deep features were administered to popular supervised classifiers to detect dengue-infected smears against the normal ones. A classification accuracy of 95.74% was obtained with the Support Vector Machine (SVM). This paper presents an explainable DL approach for dengue detection from 100x digital microscopic PBS images using pre-trained CNNs with TL. Here, an end-to-end system performs the classification process, bypassing segmentation and feature extraction/selection. To the best of our knowledge, there has been no prior publication on the use of this technique for dengue detection from PBS images.

Significant contributions of this article are listed as follows:

1. Novel end-to-end computationally efficient DL system for automatically detecting dengue from PBS images.

2. Integration of GradCAM XAI technique that provides confidence to the clinicians in the DL model’s predictions.

3. Utilizes an original hospital dataset (i.e., microscopic PBS images of dengue-infected and normal controls) collected systematically under ethical clearance.

The arrangement of the article is as follows. In Section II, the details of the image dataset and the implementation process of the models are presented. Section III provides the detailed experimental results and discussion. Finally, Section IV provides the conclusions and directions for future works.

SECTION II. Methodology

This paper proposes an explainable DL approach for detecting dengue fever from microscopic images of blood smears. State-of-the-art pre-trained CNNs are used to lower the expenses of model training from scratch. Heat maps are generated using the GradCAM XAI technique to highlight which areas on the image are concentrated in the predictions of the CNN model. The clear-cut details concerning the data, and the implementation of the TL and the XAI are described here.

A. Image Dataset

The dataset adopted is authentic hospital data accumulated from the Hematology Lab, Kasturba Hospital, Manipal. The data is acquired under ethical clearance (114/2020) granted by the Institutional Committee. Normal blood samples are collected from the Department of Immunohematology, Kasturba Hospital, Manipal, from blood bank donors visiting the department. Figure 1 shows the dataset preparation process. The dataset contains 888 PBS images (446 dengue-infected and 442 normal controls) garnered from 116 Leishman-stained blood smear thin glass slides (60 dengue-infected and 56 normal controls). A high-quality technology brightfield microscopic imaging system - Olympus DP25 digital microscope facilitated with a 5-megapixel high-precision digital camera and DP2-BSW software is used to capture the digital images of exceptional quality ($2560\times 1920$ pixels) from the slides. The microscope is coupled with a camera and linked to a computer. The images captured are in RGB color space with a color depth of eight bits per color channel. First, the body region of the slide is identified by focusing the slide with a 40x objective lens. Then, the images/oil immersion fields (ROIs) are focused and captured with a 100x objective lens with a resolution of $2560\times 1920 \times 3$ . The images are resized to $640\times 480 \times 3$ due to memory constraints. The dataset is split up randomly into training and validation, respectively, in a ratio of 4:1.

FIGURE 1.

Dataset preparation.

The emphasis is on the lymphocytes (one among the five types of leukocytes) throughout the work. Dengue alters the morphology of the lymphocytes. Studies show that this is an important indicator for dengue diagnosis [39]. PBS images of dengue-infected and normal control subjects are shown in Figure 2.

FIGURE 2.

PBS images. (a) Dengue-infected and, (b) Normal control.

B. Implementation Details

TL is executed on three state-of-the-art CNNs – ResNet50, MobileNetV3Small, and MobileNetV3Large and is utilized to discriminate dengue-infected smears from normal. The pipeline for dengue detection from PBS using transferred pre-trained CNNs is displayed in Figure 3.

FIGURE 3.

Pipeline for the detection of dengue using pre-trained CNNs.

1) Pre-Trained CNNs—ResNet50, MobileNetV3Small, and MobileNetV3Large

ResNet50 is the most common type of residual network that was introduced in 2015 to address the problems of vanishing gradient and performance degradation associated with deep CNNs [40], [41]. It has 107 layers (49 convolution layers + an FC layer) and approximately 26 million parameters [42]. ResNet50 architecture begins with a convolution layer and is followed by 16 stacked building blocks (residual) and terminates with an FC layer [43], [44]. Each residual block is a bottleneck block consisting of a stack of three convolution layers ($1\times 1$ , $3\times 3$ , and $1\times 1$ ) [41]. The building blocks of ResNet50 utilize residual connections, or skip connections, to propagate information directly through the network and overcome the degradation and vanishing gradient problems [41].

MobileNetV3, introduced in 2019, is the latest version of lightweight MobileNets [45], [46]. MobileNetV3 was built upon the MobileNetV1 and V2 structures. MobileNetV1 uses lightweight depth-wise convolutions to reduce the number of parameters [47]. MobileNetV2 was built upon MobileNetV1 with a new resource-efficient feature added: Inverted residual blocks with linear bottlenecking structure [47]. To make the structure more accurate and efficient, MobileNetV3 was introduced. Major improvements in MobileNetV3 include the Squeeze and Excitation - attention module added to the residual block and the use of hard-swish non-linearity instead of ReLU [47]. Moreover, platform-aware network architecture search and NetAdapt algorithm are engaged to optimize the architecture at the block and layer levels, respectively [47], [48]. Furthermore, MobileNetV3 includes small and large versions that operate on the same principle but vary in depth and trainable parameters [45], [49]. These networks have drastically lower parameter counts (approximately 2.9 million for small and approximately 5.4 million for large) and can be effectively implemented on resource-constrained devices [50].

2) Transfer Learning

The transfer learning plan accommodates the pre-trained CNNs to the dengue dataset. The pre-trained CNNs ResNet50, MobileNetV3Small, and MobileNetV3Large are base models. TL signifies fine-tuning the CNNs to resolve a new problem [12]. During fine-tuning, the trainable layers with randomly initialized weights are added on top of the pre-trained base model and trained on the new task-specific dataset [13]. Trainable layers include a flattened layer, a dense layer with 512 neurons and ReLU activation, a batch normalization layer, a dropout layer with rate=20%, a dense layer with 256 neurons and ReLU activation, a dropout layer with rate=20%. The output is fed to the final dense layer with two classes and Softmax activation to classify. These layers are added after the frozen base model, as illustrated in Figure 4.

FIGURE 4.

The architecture of transferred pre-trained CNNs based on (a) ResNet50 (b) MobileNetV3Small (c) MobileNetV3Large.

3) Hyperparameters

The hyperparameters used to train the transferred pre-trained CNN architectures are tabulated in Table 1. These hyperparameters are fine-tuned to attain the optimum performance. Dropout is used after each dense layer to mitigate overfitting by randomly dropping out 20% of the neurons during training. The Learn Rate is initialized to 0.0001, and a scheduler is employed to dynamically decay the Learn Rate by a factor of 0.1 when the validation loss does not improve over four consecutive epochs. The size of the batch is fixed to 32 due to memory constraints, and the network is trained with maximum epochs set to 40. The early stopping mode is triggered to halt the training when the validation loss does not improve over six consecutive epochs, and the best weights are restored.

TABLE 1 Hyperparameters Used to Train the Transferred Pre-Trained CNN Architectures

The randomly initialized weights of the trainable layers are upgraded during the training of the models, depending on the optimization algorithm, to lessen the loss function [51]. Here, the loss function is cross-entropy. Cross-entropy loss is represented by Eq. (1)

$$\begin{equation*} L_{CE}=-\sum \nolimits _{i=1}^{n} {t_{i}log\left ({p_{i} }\right)} \tag{1}\end{equation*}$$

In Eq. (1), n is the number of classes, $t_{i}$ corresponds to the truth label, and $p_{i}$ corresponds to the Softmax probability of $i^{th}$ class. Gradient descent is a common optimization algorithm that updates the weights of the network iteratively. The gradient is mathematically represented as a partial derivative of the loss concerning the weight. A single update of the weight is represented in Eq. (2).

$$\begin{equation*} w: =w- \alpha \partial L/\partial w \tag{2}\end{equation*}$$

In Eq. (2), w stands for the weight, $\alpha$ stands for learn rate, and L stands for the loss function. Many improvements to the gradient descent algorithm have been proposed, such as stochastic gradient descent with momentum (sgdm), RMSprop, and Adam [51], [52]. Here, the binary cross-entropy loss during training has been curtailed using an Adam optimizer (Solver). Figure 5 illustrates the pseudo-code for implementing TL for the proposed method.

FIGURE 5.

The pseudo-code for the implementation of transfer learning for the proposed work.

All the CNN models are executed using Keras (with TensorFlow backend) on Kaggle Notebooks with Nvidia GPU P100 Accelerator (15.9GB RAM) conducted on a 64-bit Laptop with an Intel Core i5-7200U CPU (8GB RAM).

4) GradCAM Explanations

GradCAM is the most extensively employed XAI technique for medical image analysis [17], [53]. It is very often coupled with DL models such as CNNs, which are popular for image recognition [15], [17]. The last convolution layer of a CNN model contains the most discriminative features with detailed spatial information [15], [17]. GradCAM uses the gradients of the class score with respect to the feature maps at the final convolution layer to generate heat maps [15], [17]. These heatmaps are then superimposed on the images, which help the users to see the areas of the image that are most valuable for the model predictions. GradCAM follows three steps [15], [18], [54] to generate the heat maps, as shown below:

• Step 1:

  Calculate the class score gradient c, $y^{c}$ (before the Softmax), with respect to k feature maps $A^{k}$ of the last convolution layer, i.e $\frac {\partial y^{c}}{\partial A_{i,j}^{k}}$ .

• Step 2:

  Global average pool the gradients to obtain the weights $\alpha _{k}^{c}$ as given by Eq. (3). This weight picks up the value of feature map k for a target class c.

  $$\begin{equation*} \alpha _{k}^{c}=\frac {1}{N}\sum \nolimits _{i} \sum \nolimits _{j} \frac {\partial y^{c}}{\partial A_{i,j}^{k}} \tag{3}\end{equation*}$$

  In Eq. (3), N represents the pixels in the feature map.

• Step 3:

  The GradCAM map is then a weighted combination of the feature maps with an applied ReLU as given by Eq. (4).

  $$\begin{equation*} M=\mathrm {RELU}\left ({\sum \nolimits _{k} \alpha _{k}^{c} A^{k} }\right) \tag{4}\end{equation*}$$

The ReLU activation preserves only the features that have a positive contribution to the class of interest.

SECTION III. Results and Discussion

All three models are assessed using the five-fold cross-validation scheme, wherein the data is split randomly into training (80%) and validation (20%). Further, the models are tested with unseen data. The test dataset contains 78 PBS images, of which 37 are dengue-infected, and 41 are normal controls. Six popular indices – Accuracy, Recall, Specificity, Precision, F1-score, and Area Under the ROC Curve (AUC) are used to gauge the model’s performance. These statistical metrics are derived from true positives (‘Dengue’ correctly classified; TP), true negatives (‘Normal’ correctly classified; TN), false negatives (‘Dengue’ incorrectly classified as ‘Normal’; FN), and false positives (‘Normal’ incorrectly classified as ‘Dengue’; FP) and are defined in Eq. (5)–(9).

$$\begin{align*} \mathrm {Accuracy}&=\frac {\mathrm {TP+TN}}{\mathrm {TP+TN+FP+FN}} \times 100 \tag{5}\\ \mathrm {Recall}&=\frac {\mathrm {TP}}{\mathrm {TP+FN}} \times 100 \tag{6}\\ \mathrm {Specificity}&=\frac {\mathrm {TN}}{\mathrm {FP+TN}} \times 100 \tag{7}\\ \mathrm {Precision}&=\frac {\mathrm {TP}}{\mathrm {FP+TP}} \times 100 \tag{8}\\ \mathrm {F1 Score}&=\frac {\mathrm {2\ast Pre\ast Sen}}{\mathrm {Pre+Sen}} \times 100 \tag{9}\end{align*}$$

Figure 6 shows the Accuracy and Loss plots of the Training and Validation for five folds of the transferred CNN based on MobileNetV3Small. Table 2 presents the classification accuracies of the transferred CNN models on the validation dataset for each of the five folds, along with the classification accuracy averaged over the five folds. Table 3 highlights the performance of the models during training and validation, averaging over five folds. Table 4 summarizes the detailed classification performance of the models on the validation dataset, averaging over five folds. Figure 7 reports the confusion matrices, and Table 5 reports the overall model’s performance on the test (unseen) data, averaging over the five folds.

TABLE 2 Classification Accuracies on the Validation Dataset for Each of the 5 Folds and the Overall (MEAN ± SD) Classification Accuracy

TABLE 3 Training/Validation Performance (MEAN ± SD) for Five-Fold Cross-Validation

TABLE 4 Performance Metrics (MEAN ± SD) for Five-Fold Cross-Validation on the Validation Dataset

TABLE 5 Performance Metrics (MEAN ± SD) for Five-Fold Cross-Validation on the Test Dataset

FIGURE 6.

Accuracy and Loss plots - Training (Blue) / Validation (Orange) for five folds of the transferred pre-trained CNN based on MobileNetV3Small.

FIGURE 7.

Confusion matrices (Mean ± SD) for five-fold cross-validation on the test dataset of the transferred pre-trained CNN based on (a) ResNet50, (b) MobileNetV3Small, and (c) MobileNetV3Large.

While all models rendered good performance (above 98% overall classification accuracy), MobileNetV3Small is the recommended model for this classification problem due to its significantly less computationally demanding characteristics. MobileNetV3Small is characterized by limited parameters, i.e., approx. 2.9 million, and therefore is the least time-consuming and has the least resource requirement.

Figures 8 and 9 demonstrate the GradCAM localization of the areas of the image that are most valuable for the predictions. As a result, the proposed transferred pre-trained CNN models (for example, based on ResNet50) recognized the changes in the morphology of the lymphocytes in Figure 8 and highlighted it with an activation map, demonstrating that it is the noteworthy region of interest for the identification of Dengue class in the image. Instead, as indicated in Figure 9, the model for the Normal class did not highlight the lymphocyte region, indicating no morphological changes in the lymphocyte. The red/yellow color in the activation maps indicated more concentrated regions, and the less concentrated regions are with lighter colors approaching green/blue.

FIGURE 8.

Activation map for Dengue class obtained from the GradCAM technique for the transferred pre-trained CNN based on ResNet50.

FIGURE 9.

Activation map for Normal class obtained from the GradCAM technique for the transferred pre-trained CNN based on ResNet50.

The outcomes of this work are inspiring and demonstrate that the pre-trained CNNs have the capacity to yield commendable assistance in the PBS analysis for dengue diagnosis. This work is crucial and can contribute significantly to healthcare as this offers some level of explainability of the inner workings of the CNNs that clinicians may relate to. Moreover, this work is better in the context that it bypasses laborious steps like segmentation of lymphocyte nucleus, feature extraction, and feature ranking, in contrast to the classical machine learning methods. In one of our previous works [38], ten, morphological and GLCM features were extracted from the segmented lymphocyte nucleus. These features, when coupled with SVM, achieved the best classification with Accuracy, Recall, Specificity, Precision, F1 Score, and AUC of 93.62%, 92.59%, 95%, 96.15%, 94.34%, and 0.96, respectively. Furthermore, this work differs from works in that the authors have taken advantage of pre-trained CNNs as feature extractors and linked them to a separate machine-learning classifier. In another previous work [37], 1,000 deep and 177 Local Binary Pattern generated features were derived from the segmented lymphocyte nucleus. Deep features were derived by utilizing pre-trained CNN MobileNetV2 as a feature extractor. Feature selection algorithm ReliefF was used to select 100 important features. These features given to SVM yielded the best classification with accuracy, recall, specificity, precision, F1 Score, and AUC of 95.74%, 98.15%, 92.50%, 94.64%, 96.36%, and 0.98, respectively. The model’s performance can be further boosted by exploring different fine-tuning strategies.

There was a dearth of existing publications employing PBS images for dengue diagnosis, and the literature review revealed no published works from other research groups utilizing similar methodologies on comparable datasets. Most of the research work reported on automated diagnosis of dengue is by utilizing symptoms, vital signs, blood profile data, or a combination of these. Gambhir et al. proposed a PSO-optimized ANN for the diagnosis of dengue. With 16 attributes containing symptoms, vital signs, and blood profile data, authors classified the data and documented an accuracy, recall, and specificity of 87.27%, 68%, and 92.94%, respectively [55]. Mello-Roman et al. developed a symptom-based diagnostic model for dengue fever. With 38 attributes, including symptoms, the authors classified the data using MLP and documented an accuracy, recall, and specificity of 96%, 96%, and 97%, respectively [35]. Katta et al. used symptoms to develop an efficient model for dengue detection. The RF classifier yielded an accuracy and sensitivity of 94.39% and 95.60%, respectively [33]. Hoyos et al. developed a decision-support system for dengue diagnosis using the fuzzy cognitive map. With 22 features, including symptoms, vital signs, and blood profile data, the authors achieved a classification accuracy of 89.40% [36]. Table 6 summarizes the comparison of the performance of the proposed work with other works on automated detection of dengue published in the literature.

TABLE 6 Comparison of the Proposed Work With the State-of-the-Art Works From the Literature

This work incorporates an end-to-end computationally efficient DL system for automatically detecting dengue from PBS. The system is integrated with GradCAM explainability and utilizes data sourced authentically from Kasturba Hospital, Manipal, under ethical clearance. Certainly, the entries in Table 6 depict that the proposed method for dengue diagnosis outperformed in comparison with the state-of-the-art studies published in the literature.

SECTION IV. Conclusion and Future Work

The examination of the PBS is a powerful adjunct to other clinical procedures. In connection with dengue diagnosis, it can be a crucial add-on to the Complete Blood Count test and NS1 antigen capture and can substantially aid dengue diagnosis in low-resource settings. This work utilizes pre-trained CNNs for dengue fever detection from digital microscopic PBS images. The transfer learning strategy was successful in differentiating dengue-infected and normal smears. All three models rendered good performance with classification accuracy above 98%. Despite being less computationally expensive, the performance of the transferred pre-trained CNN based on MobileNetV3Small is at par with the other two models. Hence, transferred pre-trained CNN based on MobileNetV3Small is the preferred model for the proposed method of dengue diagnosis. Explainability is recognized as a key component for the acceptance of AI systems for clinical use. An explainability technique - GradCAM was integrated to the models to visualize the specific regions of the smears that were most dominant in making the predictive decisions. Promising results show that the developed models have the potential to provide high-quality support to haematologists by expertly executing tedious, repetitive, and time-consuming duties in hospitals and remote/low-resource settings. Future works could be focused on examining different fine-tuning strategies to facilitate performance improvement. The dataset size (multiple hospital data) could be further improved to increase the algorithm’s robustness. Moreover, a multiclass problem could be considered for analyzing dengue severity (normal/mild dengue/severe dengue). Furthermore, in the future, our method could be incorporated into mobile devices with a microscope attachment and utilized as a standalone product to screen for dengue in hospitals.

ACKNOWLEDGMENT

The authors thank the Department of Pathology, Kasturba Hospital, MAHE, Manipal, for providing the peripheral blood smear slides and a digital microscope to capture the peripheral blood smear images required for this work. They also would like to thank the Department of Immunohaematology, Kasturba Hospital, MAHE, Manipal, for providing blood samples from healthy blood bank donors.