Deep learning-based classification of malaria through parasitized and uninfected blood smear image analysis

Abstract

Malaria remains a major global health issue, with over 229 million cases and 409,000 deaths reported annually, particularly in sub-Saharan Africa. Current diagnostic methods, such as microscopic examination of blood smears, are time-consuming and often lack accuracy due to human error and variability in slide quality. This study introduces Malaria-Net, a novel framework integrating advanced data preprocessing techniques with a Parasite Specific Attention Convolutional Neural Network (PSA-CNN) for enhanced feature extraction and Probabilistic Extremely Randomized Trees (PERT) for classification. The proposed approach begins with preprocessing steps, including image normalization, augmentation, and noise reduction to improve image quality and consistency. The PSA-CNN focuses on relevant features specific to malaria parasites, enhancing the network's ability to distinguish between different stages of infection. The PERT is then utilized for classification, leveraging its ability to handle high-dimensional data and provide probabilistic outputs. This method aims to improve diagnostic accuracy and reduce the reliance on manual interpretation, offering a more reliable and efficient solution for malaria detection. The proposed Malaria-Net achieves an accuracy of 99.937%, demonstrating its strong overall classification performance. It shows high precision (99.669%) and recall (99.337%), indicating that the model correctly identifies positive cases and minimizes false negatives. The F1-score of 99.539% reflects a balanced performance, combining precision and recall into a single metric, confirming its robustness in malaria detection.

Keywords

blood smear image analysis malaria prediction parasite-specific attention convolutional neural networks probabilistic extremely randomized trees

1 Introduction

Malaria continues to be a significant global health burden, impacting millions of people every year. The world health organization reported approximately 229 million cases of malaria around the world in 2020,¹ and 409,000 lives were lost, most of which were in sub-Saharan Africa. Less than five-year-olds are at particular risk of dying from malaria, accounting for as much as 67% of all malaria-related deaths.² Despite all the ongoing efforts to decrease its prevalence, the parasite is still resistant to drugs, and it has not been easy to make an accurate diagnosis. The ability to detect and treat early reduces the mortality rate, although, in many regions, reliable diagnostic tools are still restricted. In healthcare applications, especially in remote and resource-poor rural areas where specialist microscopists are unavailable, accurate and rapid malaria diagnosis is necessary. Although widely used, traditional microscopy³ is time-consuming and dependent on trained personnel to interpret blood smears, thus making it less feasible for large-scale screening. In such environments, automated diagnostic systems relying on sophisticated image analysis and Machine learning approaches can significantly deliver consistent, accurate results to improve patient outcomes through early intervention.⁴ In addition, it integrates automated systems in mobile health platforms, thus helping in real-time disease monitoring and management in underserved regions.

Manual microscopic examination of blood smear images using trained visual malaria parasite⁵ is the mainstay of current malaria diagnosis. Although it is considered the gold standard, this process is prone to human error, especially if the slide quality or parasite counts are low. Additional complicating factors include technical variability in the skill levels of technicians and staining technique variability that further make accurate diagnosis difficult.⁶ However, this is a labor-intensive and time-consuming method also, which can cause diagnosis and treatment delays, especially in regions with very high patient loads during malaria outbreaks. However, the rise of artificial intelligence⁷ and machine learning⁸ in healthcare is an opportunity to overcome these challenges. Blood smear images were processed and analyzed by artificial intelligence-powered systems at speed and accuracy, thereby reducing the need for human expertise and increasing diagnostic accuracy. Deep learning and artificial intelligence techniques were used tremendously in medical image analysis, extracting complex features and patterns from large datasets. These capabilities can provide the basis for the development of artificial intelligence driven systems that could revolutionize malaria diagnosis through the ability to automate the identification of parasites in blood smears to return a faster, more reliable result. However, despite the promise of artificial intelligence and machine learning,⁹ existing systems still face several limitations.

Another major problem is the requirement of large, high-quality labeled datasets to train these models properly, which are not always used for malaria detection. Furthermore, real-world performance is decreased by the variability in image quality caused by slide preparation, staining, and lighting conditions, which can pose challenges for artificial intelligence models. The other issue is overfitting, where the models have high accuracy on training data but not new and unseen data. Furthermore, many artificial intelligence systems are black boxes,¹⁰ and healthcare professionals have trouble trusting or interpreting the decisions made by these models, making their widespread adoption by clinical practice a significant barrier. It is critical to leverage AI fully for malaria diagnostics.

This work's contributions include developing a novel Malaria-Net integrating enhanced PSA-CNN and PERT to improve blood smear image quality. Then, the PSA-CNN was introduced for efficient and optimal feature selection from blood smear images, and the challenge of variability in blood smear images was addressed. Finally, the PERT classifier was leveraged to achieve high accuracy and robustness in malaria detection.

The rest of the paper is organized as follows: Section 2 offers a detailed review of related work and current methods in malaria classification. Then, Section 3 describes the proposed method through image preprocessing, feature selection, and classification. In section 4, the results and the performance of the proposed method are presented. Section 5 concludes the paper by discussing findings and future research directions.

2 Literature review

This section provides a detailed analysis of each related work. Table 1 summarizes the related work, highlighting its strengths and limitations. In,¹¹ authors proposed a microscopic parasite malaria classification system, which, following feature selection using Generalized Normal Distribution Optimization (GNDO), was then used to propose a relationship between the various features. While the GNDO feature selection algorithm can perform well with high dimensionality when the feature number does not increase with the number of samples, it has limitations when applied in real-world applications where a large dataset of many images is involved. In,¹² authors applied deep learning-based malaria detection over analysis of blood samples using CNNs to identify malaria parasites in blood smear images. Although the proposed method works for image classification, it is prone to overfitting because the training data does not have sufficient diversity, especially in non-standard images from different sources.

Table 1.
Summary of related work.

Authors, year, Ref. No, and Method/Model Limitations

Amin, Javeria et al. (2024)¹¹ Utilizes GNDO for feature selection.
Struggles with high dimensionality in large datasets, limiting scalability in real-world applications.

Raju, M. et al. (2024)¹²

Uses CNNs for effective malaria detection from blood smear images.
Prone to overfitting due to insufficient training data diversity, especially with non-standard samples.

Muhammad, Sheharyar et al. (2024)¹³

The CNN-based for malaria detection.
It relies on a single deep learning model, limiting generalization across varied microscopy images.

Hoyos, Kenia and William Hoyos (2024)¹⁴

deep learning with data augmentation helps mitigate limited training data.
Augmented images do not fully capture real-world image complexities, leading to potential misclassification.

Dev, A. et al. (2024)¹⁵

A hybrid deep learning framework increases detection accuracy.
High computational demands make it impractical for resource-constrained environments.

Asif, Sohaib et al. (2024)¹⁶

Efficient malaria parasite detection using deep learning.
Struggles with image quality inconsistencies and noise, leading to false negatives.

Kumar, S. Aanjan et al. (2024)¹⁷

The hybrid capsule network model improves spatial hierarchy learning for detection.
It is computationally expensive, limiting its use in fast, high-speed diagnostic environments.

Sukumarran, Dhevisha et al. (2024)¹⁸

The optimized YOLOv4 model improves detection speed.
Performance drops when detecting images with low parasite counts, limiting early-stage detection.

Choudhary, Ashutosh Kumar et al. (2024)¹⁹

Bioinspired CNN enhances detection accuracy by mimicking biological processes.
Sensitive to noise and artifacts, reducing reliability with low-quality images.

Murmu, Anita and Piyush Kumar (2024)²⁰

The hybrid DLRFNet approach leverages the strengths of both methods.
Increased complexity, longer training times, and challenging hyperparameter tuning.

Kundu, Tamal Kumar et al. (2024)²¹

Combines image processing with CNNs for malaria detection.
Limited by the inability to handle various image formats, affecting deployment in diverse settings.

Ramos, Jonathan da Silva et al. (2024)²²

Transfer learning for malaria identification in conventional blood smear images.
High variability in image quality affects consistency and reduces performance.

Özbilge, E. et al. (2024)²³

Ensembling object detection models enhances robustness and reliability.
Increases computational overhead and complexity, hindering deployment in resource-limited settings.

Chaharou, Ibrahim Mouazamou Laoualy et al. (2024)²⁴

Image cropping improves feature extraction by focusing on regions of interest.
Misses’ contextual information outside the cropped area reduces accuracy when parasites are sparse.

Swaminathan, Dilipkumar et al. (2024)²⁵

Advanced deep learning techniques improve malaria diagnosis accuracy.
It relies on extensive training data and overfits with high-dimensional features, limiting real-world applicability.

Authors, year, Ref. No, and Method/Model	Limitations
Amin, Javeria et al. (2024)¹¹ Utilizes GNDO for feature selection.	Struggles with high dimensionality in large datasets, limiting scalability in real-world applications.
Raju, M. et al. (2024)¹² Uses CNNs for effective malaria detection from blood smear images.	Prone to overfitting due to insufficient training data diversity, especially with non-standard samples.
Muhammad, Sheharyar et al. (2024)¹³ The CNN-based for malaria detection.	It relies on a single deep learning model, limiting generalization across varied microscopy images.
Hoyos, Kenia and William Hoyos (2024)¹⁴ deep learning with data augmentation helps mitigate limited training data.	Augmented images do not fully capture real-world image complexities, leading to potential misclassification.
Dev, A. et al. (2024)¹⁵ A hybrid deep learning framework increases detection accuracy.	High computational demands make it impractical for resource-constrained environments.
Asif, Sohaib et al. (2024)¹⁶ Efficient malaria parasite detection using deep learning.	Struggles with image quality inconsistencies and noise, leading to false negatives.
Kumar, S. Aanjan et al. (2024)¹⁷ The hybrid capsule network model improves spatial hierarchy learning for detection.	It is computationally expensive, limiting its use in fast, high-speed diagnostic environments.
Sukumarran, Dhevisha et al. (2024)¹⁸ The optimized YOLOv4 model improves detection speed.	Performance drops when detecting images with low parasite counts, limiting early-stage detection.
Choudhary, Ashutosh Kumar et al. (2024)¹⁹ Bioinspired CNN enhances detection accuracy by mimicking biological processes.	Sensitive to noise and artifacts, reducing reliability with low-quality images.
Murmu, Anita and Piyush Kumar (2024)²⁰ The hybrid DLRFNet approach leverages the strengths of both methods.	Increased complexity, longer training times, and challenging hyperparameter tuning.
Kundu, Tamal Kumar et al. (2024)²¹ Combines image processing with CNNs for malaria detection.	Limited by the inability to handle various image formats, affecting deployment in diverse settings.
Ramos, Jonathan da Silva et al. (2024)²² Transfer learning for malaria identification in conventional blood smear images.	High variability in image quality affects consistency and reduces performance.
Özbilge, E. et al. (2024)²³ Ensembling object detection models enhances robustness and reliability.	Increases computational overhead and complexity, hindering deployment in resource-limited settings.
Chaharou, Ibrahim Mouazamou Laoualy et al. (2024)²⁴ Image cropping improves feature extraction by focusing on regions of interest.	Misses’ contextual information outside the cropped area reduces accuracy when parasites are sparse.
Swaminathan, Dilipkumar et al. (2024)²⁵ Advanced deep learning techniques improve malaria diagnosis accuracy.	It relies on extensive training data and overfits with high-dimensional features, limiting real-world applicability.

In,¹³ authors used CNN algorithms, whose dependence on a single deep learning model restricts its generalization ability for variabilities of microscopy image quality and lighting conditions. In,¹⁴ deep learning and data augmentation were used to diagnose malaria, and they were supported by synthetic image-generated data augmentations to the training set. The drawback of this approach is the limited training data. However, the augmented images do not entirely describe real-world images, which results in misclassification when applied against actual patient samples. In,¹⁵ the authors implemented a hybrid deep learning model consisting of the tail made of traditional CNNs and reinforced by some preprocessing techniques for detecting malaria. This method achieves high classification performance, but the hybrid nature of the framework in conjunction with the computational demands makes it less useful for resource-constrained environments.

In,¹⁶ a deep learning framework was used in MozzieNet to detect malaria parasites from blood smear images efficiently. While the method performs well, it cannot handle noisy or unclear images as the source of false negatives and with the inconsistency of image quality between datasets. In,¹⁷ a hybrid capsule network model of malaria parasite detection is introduced. It also learns spatial hierarchies in the data, which helps improve detection. However, capsule networks are computationally more expensive than CNNs and impractical in environments where diagnosis must be done quickly. In,¹⁸ authors used an optimized YOLOv4 deep learning model to detect malaria in blood smear images. However, the model reduces detection speed slightly while decreasing its effectiveness for early-stage malaria detection as it is less effective for low counts. In,¹⁹ authors designed a bioinspired CNN for automated malaria detection. The detection accuracy is improved through biological processes using this model. Although bioinspired approaches are sensitive to noise and artifacts in the data and produce inaccuracy given low-quality images, they are not sufficiently reliable for clinical applications.

In,²⁰ authors proposed the Deep Learning-based Random Forest Network (DLRFNet) framework for the malaria parasite classification. Then, a mix of the two methods, such as pixel intensities, are mapped into one dimension. Here, a random forest classifier is combined with a deep learning approach that increases the model's complexity and can entail longer training times and more complex hyperparameter tuning. In,²¹ the authors employed machine learning and deep learning-based image analysis to detect malaria parasites. In their approach, a system is developed that integrates image processing with CNNs, which allows for the processing of a wide variety of images but is limited in its deployment in many clinical settings with varied imaging technologies due to its inability to support several different image formats. In,²² transfer learning-based malaria parasite identification in conventional microscopic blood smear images. Despite this, there are challenges in the high variability of image quality and preparation, which causes inconsistencies in results and poor performance. In,²³ the proposal was to use an ensemble of object detection models for robust and reliable malaria parasite detection in thin blood smear images. Model ensembling is advantageous but not easily deployed in resource-constrained settings with limited computational resources due to the increased computational overhead and complexity. In,²⁴ the authors finally proposed an image-cropping method for malaria parasite detection due to heterogeneous data. Image cropping helps in feature extraction by focusing on the region of interest. However, the approach fails to capture the contextual information at the cropped area, resulting in reduced classification accuracy when parasites are sparse or irregularly distributed in the smear. In,²⁵ authors introduced the deep learning approach for malaria parasite detection to improve malaria diagnosis from blood smear images. Nevertheless, this method relies on extensive training data and the possibility of overfitting with high dimensional features, making the method unsuitable for application to smaller or less diverse datasets encountered within real-world clinical scenarios.

3 Proposed methodology

This work proposes Malaria-Net, a new algorithm that is a novel combination of advanced techniques not presented by existing surveys. This method integrates a unique data preprocessing pipeline with a PSA-CNN and PERT, improving existing drawbacks in malaria classification. Specialized preprocessing within Malaria-Net addresses image quality, consistency variability, and noise issues common to existing methods. By concentrating on parasite-specific features, the PSA-CNN increases the accuracy of feature extraction. Then, PERT also offers robust classification over high dimensional data and probabilistic results. This innovative combination provides a comprehensive and efficient solution superior to manual examination and improves diagnostic performance.

Figure 1(a) shows the training architecture of Malaria-Net. The blood smear image dataset is initially normalized to standardize lighting conditions and color intensity. Then, advanced noise reduction algorithms are used to clear up images and make the malaria parasites clear to remove artifacts and background noise inherent in manually taken images. The histogram equalization enhances the contrast between malaria-infected cells and background, improving cell structure visibility to identify and segment each cell. Further, The PSA-CNN is designed after preprocessing to highlight features typical of malaria parasites. This attention mechanism motivates network attention to be relevant to the image, which leads to more accurate feature extraction. Deep convolutional layers were used for the network to learn the complex patterns and hierarchical features of malaria parasites. It makes it easier to distinguish and classify the different stages of infection with higher precision. Then, train test splitting is applied, which splits the PSA-CNN features into 80% for training and 20% for testing. Here, $X_{t r a i n}$ and $Y_{t r a i n}$ are PSA-CNN features and target labels applied to PERT for training. Then, the test PSA-CNN features such as $X_{t r a i n}$ are applied to PERT to classify the features extracted into categories like infected or non-infected. It is probabilistic, so the system effectively accommodates uncertainties and variations in the data. With massive extractions by the PSA-CNN, PERT's ability to process high-dimensional data allows accurate classification despite having an extensive feature set. It decreases the risk of overfitting and makes the model run better. Finally, performance estimation is performed by comparing the predicted output of PERT ( $Y_{p r e d i c t}$ ) with its original label ( $Y_{t e s t}$ ). Here, the comparisons help to measure Accuracy (A), Precision (P), Recall (R), F1-score, and Confusion Matrix (CM). Further, the PSA-CNN and PERT model files are saved to avoid the retraining process.

Figure 1.

Proposed system model. (a) model training. (b) model testing.

Figure 1(b) shows the prediction architecture of Malaria-Net. A single image predicts normal and malaria classes with PSA-CNN feature extraction and PERT classification. The pre-trained models of Malaria-Net are used for faster prediction. Finally, the user can predict the classes from various images instantly.

3.1 Parasite specific attention convolutional neural network

The PSA-CNN is proposed to enhance malaria parasite classification with convolutional feature extraction and an attentive mechanism to emphasize relevant features in blood smear images. The proposed PSA-CNN structure is depicted in Figure 2. The operation of the PSA-CNN feature extraction algorithm is presented in Table 2, including several important steps, such as feature extraction through convolutional layers, the application of an attentive mechanism, and then the integration of these features for classification.

Figure 2.

Proposed parasite specific attention convolutional neural network architecture.

Table 2.

Proposed parasite specific attention convolutional neural network feature extraction algorithm.

Input: Input Images

Output: Predicted Class-Specific Features

Step 1: Image Processing: Load the blood smear image, which was grayscale or RGB. Normalize the image pixel values to ensure stable learning.

Step 2: CNN for Feature Extraction: Apply multiple convolutional layers to detect spatial and temporal patterns as presented in equation (1). Apply ReLU and MaxPooling layers to reduce image size while preserving key information as presented in equation (3).

Step 3: Spatial Feature Analysis: CNN outputs feature maps representing different malaria-related structures. These feature maps contain both parasite-related and background information.

Step 4: Attention Mechanism: The model assigns different importance levels to different regions in the feature maps. Areas likely to contain malaria parasites receive higher attention weights, while irrelevant areas receive lower importance, as presented in equation (4) and equation (5).

Step 5: Loss Calculation: The model compares the predicted classification with the actual label to compute the error. As presented in equation (6) to equation (8), a loss function measures the model's performance.

Step 6: Optimization using Adam Optimizer: The optimizer adjusts the model parameters to reduce the loss, as presented in equation (9). It dynamically updates the learning rate for faster and more stable training.

Step 7: Weights Adjustment & Model Training: The CNN layers and attention mechanism continuously refine their weights. This process is repeated for multiple training cycles (epochs) to improve performance.

Step 8: Predicted Class-Specific Features: The final feature maps are flattened and passed through fully connected layers, as presented in equation (10). A classification function (such as SoftMax) assigns probabilities to different malaria infection classes.

3.1.1 Input image

In the first stage of PSA-CNN, the blood smear image is processed to enhance relevant features and normalize the input. Images are operated on to resize, normalize, and augment all training and testing samples consistent with one another. The input image is represented as a matrix $I (i, j)$ with pixel intensity values, which is the foundation for subsequent feature extraction.

3.1.2 CNN for spatial-temporal feature extraction

Once the input image is prepared, it is processed through a series of convolutional layers designed to extract spatial-temporal features. Each convolutional layer applies a set of filters $K (m, n)$ to detect significant patterns such as edges, textures, and structures indicative of malaria parasites. Equation (1) represents the convolution operation. It is mathematically expressed as follows:

F_{o u t} (i, j) = \sum_{m = 0}^{M - 1} (\sum_{n = 0}^{N - 1} I (i + m, j + n) \cdot K (m, n))

(1)

Here, $F_{o u t} (i, j)$ is the value at position $(i, j)$ in the output feature map, I is the input image, K is the convolutional filter, and M and N are the filter's dimensions. Equation (2) presents the applying non-linear activation functions, such as Rectified Linear Unit (ReLU) in convolution layer, which enhances feature extraction by introducing non-linearity.

F_{r e l u} (i, j) = max (0, F_{o u t} (i, j))

(2)

After convolution and activation, pooling layers are applied to reduce the spatial dimensions while retaining crucial information as presented in equation (3). A max pooling ( $maxpool$ ) operation selects the highest value in each region.

F_{p o o l} (i, j) = maxpool (F_{r e l u} (i + p, j + q))

(3)

Here, $F_{p o o l} (i, j)$ is the MaxPooling value at position $(i, j)$ in the output feature map with $(p, q)$ as the current positions. This CNN based feature extraction produces initial output features that contain significant spatial representations of the malaria parasites but lack specificity in focusing on parasite-relevant regions.

3.1.3 Attention loss mechanism

To focus specifically on features pertinent to malaria parasites, the PSA-CNN integrates an attention mechanism. This mechanism assigns different weights to different regions of the feature maps, emphasizing areas likely to contain relevant features. Equation (4) presents the attention mechanism. It is expressed as follows:

A (i, j) = \frac{e x p (ϕ (F_{o u t} (i, j)))}{\sum_{P, q}^{M} (e x p (ϕ (F_{o u t} (p, q))))}

(4)

Here, $A (i, j)$ represents the attention weight at position $(i, j)$ , and $ϕ$ is a function that computes the attention score, often a learned parameter or function applied to the feature map. The attended feature map is then generated by weighting the feature maps using the attention scores as presented in equation (5).

F_{a t t} (i, j) = A (i, j) \cdot F_{o u t} (i, j)

(5)

Here, $F_{a t t} (i, j)$ is the attended feature map, emphasizing areas with higher attention weights. It allows the network to concentrate on parasite-specific features, improving the accuracy of subsequent classification. This mechanism allows PSA-CNN to prioritize regions with higher attention values, effectively reducing noise and irrelevant background information in blood smear images. The attention loss mechanism ensures the model assigns higher confidence to regions containing parasites. Equation (6) presents the loss function calculation and incorporates an attention-loss term $L_{a t t}$ , which penalizes incorrect attention allocation.

L_{a t t} = \sum_{i, j}^{M} (1 - A (i, j)) . (F_{a t t} (i, j) - F_{g t} (i, j))^{2}

(6)

Here, $F_{g t} (i, j)$ is the ground truth feature map, and M is the maximum number of features.

3.1.4 Loss value calculation

The loss value in PSA-CNN is presented in equation (7), which consists of a combination of classification loss and attention loss.

L_{t o t a l} = L_{c l s} + λ . L_{a t t}

(7)

Here, $L_{c l s}$ is the categorical cross-entropy classification loss, which is estimated using equation (8). Furthermore, $L_{a t t}$ is the attention loss, $a n d λ$ is a hyperparameter controlling the balance between classification and attention loss.

L_{c l s} = \sum_{k}^{M} y_{k} . \log (P (y = k | x))

(8)

Here, $y_{k}$ is the ground truth label, and $P (y = k | x)$ is the predicted probability.

3.1.5 Adam optimizer for weight adjustment

The Adam optimizer is used to adjust the network weights and to optimize the PSA-CNN model. Adam combines the advantages of momentum-based and adaptive learning rate optimization. The weight update rule presented in equation (9).

W_{t + 1} = W_{t} - η \frac{m_{t}}{\sqrt{v_{t}} + ϵ}

(9)

Here, $W_{t}$ is the weight at iteration t, $η$ is the learning rate, $m_{t}$ and $v_{t}$ are the first and second-moment estimates, and $ϵ$ is a small constant to prevent division by zero. The optimizer refines the network parameters to minimize the overall loss, improving classification accuracy.

3.1.6 Final output - predicted class-specific features

After applying the attention mechanism, the fully connected layers of PSA-CNN aggregate the attended features through pooling and fully connected layers to produce a final feature classification. Equation (10) presents the feature classification using a SoftMax classifier for multi-class classification.

P (y = k | x) = \frac{e x p (W_{k}^{T} x + b_{k})}{\sum_{P, q}^{M} (e x p (ϕ (F_{o u t} (p, q))))}

(10)

Here, $(y = k | x)$ is the feature probability of class k given input x, $W_{k}^{T} x$ and $b_{k}$ are the weight vector and bias for class k, and the denominator sums over all classes l. This final step outputs the probability distribution over the possible malaria infection classes, allowing for accurate diagnosis based on the input blood smear image. By combining these stages, the PSA-CNN leverages feature extraction and attention mechanisms to enhance the detection and classification of malaria parasites, addressing common challenges in traditional methods and improving diagnostic performance.

3.2 Probabilistic extremely randomized trees classification

The PERT classifier is a robust and efficient ensemble learning method that constructs multiple decision trees and aggregates their predictions to enhance classification accuracy. It introduces randomness at various stages, including data sampling and feature selection, to improve model generalization and mitigate overfitting. Table 3 shows the proposed PERT classification algorithm. Figure 3 shows the PERT architecture diagram. A random subset of the training data is selected for each tree in the ensemble. It introduces diversity among the trees, making the ensemble more robust. At each node in a tree, a random subset of features is chosen from the feature set.

Figure 3.

Proposed probabilistic extremely randomized trees classifier.

Table 3.

Proposed probabilistic extremely randomized trees classification algorithm.

Input: PSA-CNN features

Output: Classified outcome

Step 1: Initialize Dataset: Define the entire PSA-CNN feature set S as per equation (11).

Step 2: Probabilistic Sample Subsets for Each Tree: For each tree t in the ensemble, randomly select a subset C from the dataset S as described in equation (12) using multi-level probabilities.

Step 3: Randomly Select Feature Subset: To introduce additional randomness, a random subset of features is selected at each decision tree node.

Step 3.1: Determine Best Split: Evaluate splits based on the selected feature subset C to find the best split that maximizes the separation of the classes based on information gain and the Gini impurity index, as presented in Equations (13) and (14).

Step 3.2: Split the Node: Split the node into child nodes based on the best-split criterion.

Step 3.3: Repeat: Recursively repeat the process for each child node until the stopping criteria (e.g., minimum node size or maximum tree depth) are met.

Step 4: Ensemble Aggregation: Aggregate the predictions from all trees in the ensemble to make the final classification decision as presented in equation (15).

Step 5: Training Phase: Train the ensemble of trees on the training dataset S by repeating steps 2 to 4 for each tree in the ensemble.

Step 6: Prediction Phase: For a new sample

X_{n e w}

, pass it through each tree in the ensemble to get individual tree predictions. Aggregate these predictions using majority voting to determine the final class label for

X_{n e w}

Step 7: Output: The final trained model, which consists of an ensemble of trees, was output. The model predicted whether new blood smear images are malaria infected.

3.2.1 Input PSA-CNN features

The classification begins by taking the feature vectors extracted by the PSA-CNN. These features, denoted as $x_{i}$ , contain critical information about blood smear images, including pixel intensities, texture metrics, and morphological descriptors. Each feature vector is paired with a corresponding class label, $y_{i}$ , indicating whether the sample is malaria infected. Equation (11) shows the complete feature set $(S)$ . It is given as follows

S = {(x_{i}, y_{i}), i = 1, 2, \dots, N}

(11)

Here, N represents the total number of training samples. This dataset serves as the input for the PERT classification process.

3.2.2 Probabilistic sample subsets for each tree

To introduce diversity among decision trees, PERT utilizes a bootstrap sampling approach, where each tree in the ensemble is trained on a randomly selected subset C of the entire dataset S. The subset C is generally smaller than S and is obtained via random sampling with replacement. This process ensures that different trees are exposed to varying portions of the dataset, improving the ensemble's robustness. The selected training subset is presented in equation (12).

C = {(x_{j}, y_{j}), j = 1, 2, \dots, c}

(12)

Here, c represents the number of samples in each subset.

3.2.3 Extremely randomly select feature subset

Each decision tree node selects a random subset of features to introduce additional randomness. Instead of evaluating all features, PERT considers only a small subset, ensuring that different trees explore different aspects of the data. It prevents overfitting and allows trees to specialize in different decision boundaries.

3.2.4 Determine best split

For each feature in the selected subset, the classifier identifies an optimal split point that maximizes information gain or minimizes impurity. The commonly used criteria for determining the best split include Gini impurity ( $G$ ), which measures the homogeneity of a node, and a lower Gini value indicates perfect values. Equation (13) estimates the Gini impurity for a node containing samples from different classes.

G = 1 - \sum_{i = 1}^{m} p_{i}^{2}

(13)

Here, $p_{i}$ is the probability of a sample belonging to class i, and m is the number of classes. The information gain ( $I G$ ) evaluates how much entropy is reduced by the split. Equation (14) computed the $I G$ .

I G = H (S) - \sum_{j = 1}^{m} \frac{| S_{j} |}{| S |} H (S_{j})

(14)

Here, $H (S)$ is the entropy of the parent node, and $H (S_{j})$ is the entropy of child nodes after splitting.

3.2.5 Split the node

Once the best split is determined, the node is divided into two child nodes containing samples that satisfy the splitting condition. This recursive partitioning process continues until a predefined stopping criterion is met, such as the maximum depth of the tree or the minimum number of samples per leaf.

3.2.6 Repeat for remaining nodes

The process of feature selection, determining the best split and node splitting, is repeated recursively for each newly formed node. The decision tree continues to grow until the stopping criteria are satisfied, resulting in a complete tree.

3.2.7 Ensemble aggregation

Once multiple decision trees are constructed, their predictions are aggregated to make a final classification decision. Each tree independently classifies a given test sample, and the final prediction is determined by majority voting as presented in equation (15).

y = mode ({y_{1}, y_{2}, \dots, y_{T}})

(15)

Here, $y_{t}$ is the prediction from tree t, and T is the total number of trees in the ensemble. An unseen blood smear image is passed through the PSA-CNN model to extract relevant features during testing. These features are input into the trained PERT classifier, where each decision tree provides a classification output. The final prediction is determined by majority voting among the trees.

4 Results and discussion

This section evaluated the performance of different malaria classification methods using the same dataset regarding primary metrics such as accuracy, precision, Recall, and the F1 Score. The robustness and accuracy of the models, such as Malaria-Net and others, are assessed to find the effective implementation of the models using consistent evaluative criteria. Each method is used to classify malaria in blood smear images, and the comparative analysis identifies which method performs best for malaria detection.

4.1 Simulation environment

The research uses Python programming language with a graphical processing unit environment. The hardware contains an Intel Core i7 10th Gen 10750H laptop with 16GB of RAM, 1TB SSD for storage, and an 8GB NVIDIA GeForce RTX 2070 graphics card to crunch the computing requirements. The working software environment is based on Python 3.7 and utilizes TensorFlow 2.x to build and train machine learning and deep learning models in the research. These resources handle the execution and processing of the simulation tasks, carefully selected to achieve maximum efficiency.

4.2 Dataset

The malaria classification dataset contains 27,558 blood smear images divided into two folders, i.e., Infected and Uninfected. The folder “Infected” includes images of red blood cells exhibiting malaria parasites, while “Uninfected” consists of images of healthy blood cells without infection. It is a microscopic image collection with a diverse and comprehensive helpful set for training and testing artificial intelligence models. Variations in quality, lighting, and staining, combined with the fact that the images are real-world mammalian bacterial detecting conditions and robust systems for these conditions are desired by both the biologist and industry, make the images well suited for training robust malaria detection systems. The available dataset from the National Institutes of Health gives a good opportunity to gain a step further in deep learning and image processing techniques to automate malaria diagnosis. This dataset is collected from a trusted source, guarantees its credibility, and is a key contribution to developing an AI-based model for malaria detection. Figure 4 shows the sample images. The breakdown of the malaria dataset into training and testing sets is given in Table 4. The dataset is split into 80% of images for training and 20% for testing. The ‘Infected’ and the ‘Uninfected’ classes have the same number of images in each class. The training set consists of 22,046 images, while the testing set contains 5512 images, leading to 27,558 images in the dataset. Here, the classes are balanced, which helps build a strong classification model.

Figure 4.

Sample images from the dataset.

Table 4.

Dataset distribution.

Class	Training Set	Testing Set	Total
Infected	11,023	2756	13,779
Uninfected	11,023	2756	13,779
Total	22,046	5512	27,558

4.3 Hyperparameters

Table 5 compares the hyperparameters of various deep learning networks, including the GNDO,¹¹ MOZZIENet,¹⁶ DLRFNet,²⁰ and the Proposed Malaria-Net. It highlights the key hyperparameters and settings used in each network, demonstrating the distinct configurations and advantages of the proposed Malaria-Net in classification.

Table 5.
Hyperparameters of various deep learning networks.

Hyperparameter GNDO¹¹ MOZZIENet¹⁶ DLRFNet²⁰ Proposed Malaria-Net

Learning Rate 0.0005 0.009 0.001 to 0.0001 0.0001

Batch Size 128 64 64 32

Maximum Epochs 200 150 100 30

Optimizer Adagrad Adadelta Stochastic Gradient Adam

Weight Initialization Normal (Gaussian) LeCun initialization Xavier initialization He initialization

Activation Function Tanh Leaky ReLU Sigmoid ReLU

Dropout Rate 0.9 0.9 0.5 0.5

Learning Rate Scheduler Exponential decay Cosine Annealing Cyclical Step decay (drop every 10 epochs)

Number of Layers 5 Conv layers + 2 FC layers 4 Conv layers + 2 FC layers 6 Conv layers + 3 FC layers 2 Conv + 2 Maxpooling + 2 FC layers

Attention Mechanism SoftMax Tanh sigmoid PSA

Loss Function Hinge loss Kernal divergence Sparse categorical cross-entropy Categorical cross-entropy

Early Stopping Minimum Validation Scheduled Learning Rate Validation Loss Monitored Class

Hyperparameter	GNDO¹¹	MOZZIENet¹⁶	DLRFNet²⁰	Proposed Malaria-Net
Learning Rate	0.0005	0.009	0.001 to 0.0001	0.0001
Batch Size	128	64	64	32
Maximum Epochs	200	150	100	30
Optimizer	Adagrad	Adadelta	Stochastic Gradient	Adam
Weight Initialization	Normal (Gaussian)	LeCun initialization	Xavier initialization	He initialization
Activation Function	Tanh	Leaky ReLU	Sigmoid	ReLU
Dropout Rate	0.9	0.9	0.5	0.5
Learning Rate Scheduler	Exponential decay	Cosine Annealing	Cyclical	Step decay (drop every 10 epochs)
Number of Layers	5 Conv layers + 2 FC layers	4 Conv layers + 2 FC layers	6 Conv layers + 3 FC layers	2 Conv + 2 Maxpooling + 2 FC layers
Attention Mechanism	SoftMax	Tanh	sigmoid	PSA
Loss Function	Hinge loss	Kernal divergence	Sparse categorical cross-entropy	Categorical cross-entropy
Early Stopping	Minimum Validation	Scheduled Learning Rate	Validation Loss	Monitored Class

4.3.1 Learning rate

The learning rate determines how much the model's weights are adjusted concerning the loss gradient during each step of the optimization process. The proposed Malaria-Net uses a learning rate of 0.0001, slightly lower than GNDO (0.0005) and MOZZIENet (0.009). With a lower learning rate, the chance of overshooting the optimal point is lower, especially in medical image classification tasks such as malaria detection. The learning rate of the proposed Malaria-Net is optimal when considering the balance between slow convergence speed and high prediction accuracy.

4.3.2 Batch size

Batch size is the number of samples the model works over before it updates its internal parameters. Using a smaller batch size of 32 is beneficial for better generalization and avoiding overfitting. The proposed Malaria-Net has utilized this. GNDO employs a large batch size of 128 while reducing the batch size. It helps the model converge more, especially when training data is noisy or imbalanced, as in the malaria classification example.

4.3.3 Number of epochs

The term is several epochs or how the training during the training as a whole number of times this dataset will pass through the model. As for the number of epochs (maximum 30 for the proposed Malaria-Net), it is less than for the GNDO (200 epochs) and MOZZIENet (150 epochs). It will also help prevent overfitting and reduce the training time, especially when the architecture is sophisticated with attention mechanisms and similar techniques like dropout, as seen in the proposed Malaria-Net.

4.3.4 Optimizer

Parameters are adjusted according to the gradient of the loss function. The Adam optimizer helps adapt the learning rate for each parameter independent of others to perform better in complex models and is used by the proposed Malaria-Net. In contrast, the Adagrad learning rate is adapted based on the previous gradient, which GNDO uses. In contrast, MOZZIENet uses Adadelta, a variant of Adagrad with bounds on past gradient accumulation.

4.3.5 Weight initialization

Weight Initialization is important to start faster and prevent vanishing exploding gradients. The proposed Malaria-Net mitigates gradient flow issues, which arise with ReLU activation functions, using He initialization, which is tailored explicitly for layers using the ReLU activation function. In contrast to GNDO's Normal (Gaussian) initialization or MOZZIENet's LeCun initialization, which is less suited for deep networks, He initialization is preferred.

4.3.6 Activation function

An activation function will produce an output for all the layers. The proposed Malaria-Net uses ReLU, a simple and deep learning network. It works around the vanishing gradient problem that can occur with other activation functions like Tanh used in GNDO and Sigmoid used in DLRFNet, making the training faster and more efficient.

4.3.7 Regularization technique

One of the regularization techniques is dropout, which helps prevent overfitting by randomly dropping units during training. The proposed Malaria-Net uses a 0.5 dropout rate, which is suitable for keeping the model shape regular yet not too complex. Here, MOZZIENet and GNDO underfit due to a higher value and tested with a higher dropout rate of 0.9, while the proposed Malaria-Net has a dropout rate of 0.5, like DLRFNet.

4.3.8 Learning rate

A learning rate scheduler is a method for adjusting the learning rate in training based on certain predefined conditions. The Proposed Malaria-Net uses a step decay scheduler (decreases the learning rate every 10 epochs). The model will fine-tune weights towards the end of the training process. Unlike GNDO, MOZZIENet uses cosine annealing, and DLRFNet uses cyclical scheduling, but step decay is well-tuned for fine-tuning large neural networks with complicated architectures.

4.3.9 Depth of network

The depth of the network stands by the complexity and learning of intricate features. Using a relatively simple architecture consisting of 2 convolutional, two max-pooling layers followed by one fully connected layer. This simplicity is highly efficient for learning, combined with more sophisticated techniques such as attention mechanisms. GNDO and MOZZIENet are architectures with more layers than the current state of the art, representing an increased computational complexity and a slight improvement in malaria detection performance.

4.3.10 Attention mechanism

The attention mechanism makes it easy for the network to focus on relevant regions in an image, which is helpful for tasks such as malaria parasite detection. The proposed Malaria-Net uses the PSA-CNN mechanism to utilize features specific to malaria parasites to enhance diagnostic accuracy. Instead of SoftMax, MOZZIENet uses Tanh, and DLRFNet uses Sigmoid, which is not as specific to where they focus on the image.

4.3.11 Loss function

It controls the process of optimizing the model. Categorical Cross entropy is suitable for multi-class classification problems like Malaria detection, and the proposed Malaria-Net uses it. It is more suitable for multi-class tasks than hinge loss and is used in GNDO and Kernal Divergence in MOZZIENet.

4.3.12 Early stopping

Early stopping helps prevent overfitting by stopping training when performance stops improving. The Proposed Malaria-Net uses monitored class-based early stopping, which stops training based on the performance of each class separately. It ensures the network does not continue training when performance becomes suboptimal for any class. It is a more nuanced approach than GNDO's minimum validation stopping or MOZZIENet's scheduled learning rate stopping.

4.4 Prediction results

In Figure 5, Output Predicted Probabilities (OPP) values represent the predicted outcomes on sample blood smear images. In Figure 5(a), the OPP values are {0.01, 0.99}, where 0.99, the higher probability, is associated with the “Uninfected” class. Similarly, Figure 5(b) has OPP = {0.02, 0.98}, and Figure 5(c) has OPP = {0.12, 0.88}, both showing higher probabilities for the “Uninfected” category, confirming the exact prediction. In Figure 5(d), the OPP values are {0.82, 0.18}, and in Figure 5(e), {0.76, 0.24}, both show higher probabilities for the “Infected” class. Finally, Figure 5(f) shows OPP = {0.95, 0.05}, with 0.95 indicating a strong prediction for the “Infected” category. Based on the OPP values assigned to each sample image, the values suggest high confidence in the respective predictions, whether infected or uninfected.

Figure 5.

Predicted outcomes on sample images. (a), (b), (c) predicted as uninfected. (d), (e), (f) predicted as infected.

4.5 Performance evaluation

Table 6 comprehensively compares the performance of various malaria classification methods. The accuracy of a classification model indicates the proportion of correct predictions out of all predictions made. The proposed Malaria-Net achieves an impressive accuracy of 99.937%. Compared to GNDO,¹¹ which has an accuracy of 96.796%, this represents an enhancement of approximately 3.243%. Similarly, compared to MOZZIENet,¹⁶ with an accuracy of 97.740%, the proposed Malaria-Net shows an improvement of 2.243%. Finally, compared to DLRFNet,²⁰ which achieves an accuracy of 98.805%, the proposed Malaria-Net's accuracy is 1.143% higher. These improvements underscore the robustness of the Proposed Malaria-Net in accurately classifying instances.

Table 6.
Performance comparison of various classification methods.

Proposed

Metric GNDO¹¹ MOZZIENet¹⁶ DLRFNet²⁰ Malaria-Net

Accuracy (%) 96.796 97.740 98.805 99.937

Precision (%) 96.232 97.852 98.240 99.669

Recall (%) 96.324 97.951 98.822 99.337

F1-Score (%) 96.924 97.265 98.848 99.539

				Proposed
Accuracy (%)	96.796	97.740	98.805	99.937
Precision (%)	96.232	97.852	98.240	99.669
Recall (%)	96.324	97.951	98.822	99.337
F1-Score (%)	96.924	97.265	98.848	99.539

Figure 6.

Confusion matrices. (a) GNDO.¹¹ (b) MOZZIENet.¹⁶ (c) DLRFNet.²⁰ (d) Proposed Malaria-Net.

Figure 7.

Accuracy and Loss Graphs. (a) Validation accuracy. (b) Training accuracy. (c) Training loss. (d) Validation loss.

The precision is the ratio of the correctly predicted positive to the total predicted positives. The Malaria-Net proposed attains a precision of 99.669%, which is higher than 96.232% reported in GNDO¹¹ and is an improvement of 3.574%. The proposed Malaria-Net has a more significant improvement of 1.862% than MOZZIENet,¹⁶ which has a precision of 97.852%. Furthermore, compared to a precision of 98.240% in DLRFNet,²⁰ the proposed Malaria-Net has been enhanced by 1.453%. These striking increases suggest that the proposed Malaria-Net can better remove false positives and thus improve the credibility of optimistic predictions.

The recall or sensitivity ratio is the number of correctly predicted positive observations to the number of observations in the actual class. The proposed Malaria-Net recall is 99.337%, which is an improvement of 3.013% compared to the Recall of GNDO¹¹ (96.324%). Compared to MOZZIENet,¹⁶ Malaria-Net improves recall by 1.413% (97.951%). Last, the proposed Malaria-Net achieves 0.515% more than the Recall of 98.822% of DLRFNet.²⁰ The Proposed Malaria-Net has improved since these three methods, and these improvements indicate that the Proposed Malaria-Net is more effective in identifying positive instances than the other three.

The F1-Score is the harmonic mean of precision and recall, fitting well between the two. In addition, the proposed Malaria-Net achieves an F1-Score of 99.539%. It represents an improvement of 2.612% from the F1-Score of 96.924% for GNDO.¹¹ The F1 Score obtained for the proposed Malaria-Net compared to MOZZIENet,¹⁶ which has an F1 Score of 97.265%, is 2.338% enhancement. In the last, we achieve the F1-Score of 98.848% using DLRFNet²⁰'s, which is 0.691% higher than Malaria-Net's F1-Score of 98.188%. The improvements in the Proposed Malaria-Net's balanced performance are precision and recall.

Figure 6 shows the confusion matrices of four different malaria detection models, such as GNDO,¹¹ MOZZIENet,¹⁶ DLRFNet,²⁰ and the proposed Malaria-Net. Each model's confusion matrix will assess the classification's performance depending on the true positives, true negatives, false positives, and false negatives present, giving details of the accuracy and the possible misclassification errors the models have made. Although the GNDO¹¹ model provides strong performance with an accuracy of 96.796%, several misclassifications exist—particularly to correctly identify some malaria parasite types. Detecting targets using the MOZZIENet model¹⁶ with a slightly increased accuracy of 97.740% bests the case of general image quality inconsistency, which leads to a few false negatives. Although the DLRFNet obtains an even higher accuracy of 98.805%,²⁰ this still produces some errors due to long training time and increased complexity. The most accurate model is the proposed Malaria-Net, which achieves an accuracy of 99.937%, ensuring the highest ability to accurately classify malaria parasites with very low misclassification, confirming its success in dealing with the related issues of its previous models like overfitting, variability in image quality, and computational requirements.

Figure 7 shows the networks presented, such as Malaria-Net, DLRFNet,²⁰ GNDO,¹¹ and MOZZIENet,¹⁶ and comparative accuracy and loss graphs for these networks. Figure 7(a) illustrates the validation accuracy of each model, where the proposed Malaria-Net reaches the highest accuracy compared with DLRFNet, MOZZIENet, and GNDO. The proposed Malaria-Net generalizes well to unseen data, with less overfitting than the others. Figure 7(b) illustrates the training accuracy, where the proposed model again leads, indicating that it learns the training data most effectively. Figure 7(c) displays the training loss curves, where the proposed Malaria-Net shows the lowest training loss, indicating that it converges more quickly and efficiently than the other models, which exhibit higher and more fluctuating training losses. Figure 7(d) depicts the validation loss, where the proposed Malaria-Net maintains the lowest loss throughout the training process, reflecting its superior ability to balance accuracy with generalization. In comparison, the other models show higher validation losses, with GNDO¹¹ exhibiting the highest, suggesting overfitting issues or less effective optimization.

4.6 Ablation study

Table 7 presents an ablation study of the proposed Malaria-Net, illustrating the impact of excluding key components such as preprocessing and PSA-CNN feature extraction on the performance metrics. When preprocessing is excluded from the Proposed Malaria-Net, the accuracy drops to 97.324%, reducing 2.613% from the overall accuracy of 99.937%. Similarly, precision falls to 97.926%, a decrease of 1.743% compared to the overall precision of 99.669%. Recall also experiences a decline, reaching 97.618%, which is 1.719% lower than the 99.337% achieved by the complete methodology. The F1-Score, balancing precision and recall, decreases to 97.988%, marking a reduction of 1.551% from the overall F1-Score of 99.539%. These figures highlight the significant role of preprocessing in enhancing the model's overall performance.

Table 7.
Ablation study of proposed Malaria-Net.

Metric Malaria-Net without Preprocessing Malaria-Net without PSA-CNN feature extraction Overall Malaria-Net

Accuracy (%) 97.324 98.099 99.937

Precision (%) 97.926 98.708 99.669

Recall (%) 97.618 98.245 99.337

F1-Score (%) 97.988 98.724 99.539

Metric	Malaria-Net without Preprocessing	Malaria-Net without PSA-CNN feature extraction	Overall Malaria-Net
Accuracy (%)	97.324	98.099	99.937
Precision (%)	97.926	98.708	99.669
Recall (%)	97.618	98.245	99.337
F1-Score (%)	97.988	98.724	99.539

Excluding PSA-CNN feature extraction impacts the methodology's efficacy, albeit to a lesser extent than omitting preprocessing. Without PSA-CNN feature extraction, the accuracy is 98.099%, 1.838% lower than the overall accuracy. Precision drops to 98.708%, a decrease of 0.961% from the complete methodology. The recall is 98.245%, reflecting a reduction of 1.092% compared to the overall recall. The F1-Score without PSA-CNN feature extraction is 98.724%, 0.815% lower than the overall score. These reductions underscore the importance of PSA-CNN feature extraction in optimizing the model's performance, though its absence has a slightly lesser impact than the absence of preprocessing. Together, these components significantly contribute to the high performance of the proposed Malaria-Net, as evidenced by the overall metrics.

5 Conclusion

The proposed Malaria-Net framework presents a novel and effective solution for automated infected and uninfected cell classification from blood smear images, addressing the limitations of existing methods, such as manual examination errors, variability in slide quality, and inefficient feature extraction. The method significantly improves diagnostic accuracy and reliability by combining advanced data preprocessing techniques, the PSA-CNN for targeted feature extraction, and PERT for probabilistic classification. This hybrid approach not only overcomes the drawbacks of traditional methods but also enhances the automation and scalability of malaria diagnostics. This methodology was expanded in the future by integrating additional parasite-specific attention mechanisms to detect various malaria species or other blood-borne pathogens. Furthermore, more extensive and diverse datasets and real-time processing capabilities could extend their applicability to remote and resource-constrained regions. Incorporating advanced explainability techniques for medical professionals to interpret results and combining the framework with telemedicine platforms could revolutionize global malaria surveillance and management systems.

Footnotes

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Author Biographies

Chityala Nandini is an Undergraduate in Computer Science and Engineering (Data Science) department at Malla Reddy Engineering College and Management Sciences, Kistapur, Medchal, Telangana.

T Rajesh is a Professor in Computer Science and Engineering at G. Narayanamma Institute of Technology and Science, Hyderabad. His research interests are medical image processing, and artificial intelligence.

Ramulakollu Nikil Kumar is an Undergraduate in Computer Science and Engineering department at Malla Reddy Engineering College and Management Sciences, Kistapur, Medchal, Telangana.

Pasupuleti Shyam Srikar is an Undergraduate in Computer Science and Engineering department at Malla Reddy Engineering College and Management Sciences, Kistapur, Medchal, Telangana.

R Sai Samanth is an Undergraduate in Computer Science and Engineering (Data Science) department at Malla Reddy Engineering College and Management Sciences, Kistapur, Medchal, Telangana.

Nimmana Deepak Sai is an Undergraduate in Computer Science and Engineering (Data Science) department at Malla Reddy Engineering College and Management Sciences, Kistapur, Medchal, Telangana.

Maram Sai Kumar Reddy is an Undergraduate in Computer Science and Engineering (Data Science) department at Malla Reddy Engineering College and Management Sciences, Kistapur, Medchal, Telangana.

References

Acherar

Tannier

Tantaoui

, et al. Evaluating plasmodium falciparum automatic detection and parasitemia estimation: a comparative study on thin blood smear images. PLos One 2024; 19: e0304789.

Sukumarran

Hasikin

Khairuddin

ASM

, et al. Machine and deep learning methods in identifying malaria through microscopic blood smear: a systematic review. Eng Appl Artif Intell 2024; 133: 108529.

Grignaffini

Simeoni

Alisi

, et al. Computer-aided diagnosis systems for automatic malaria parasite detection and classification: a systematic review. Electronics (Basel) 2024; 13: 3174.

Rismayanti

. Segmentation and feature extraction for malaria detection in blood smears. IJAIMI 2024; 2: 18–29.

Mujahid

Rustam

Shafique

, et al. Efficient deep learning-based approach for malaria detection using red blood cell smears. Sci Rep 2024; 14: 13249.

Dhevisha Sukumarran

ESL

Khairuddin

ASM

Ngui

, et al. Automated identification of malaria-infected cells and classification of human malaria parasites using a two-stage deep learning technique. IEEE Access 2024; 12: 135746–135763.

Midasala

Prabhakar

Krishna Chaitanya

, et al. MFEUsLNet: skin cancer detection and classification using integrated AI with multilevel feature extraction-based unsupervised learning. Eng Sci Technol Int J 2024; 51: 101632.

Moysis

Brown

Shokunbi

, et al. Leveraging deep learning for detecting red blood cell morphological changes in blood films from children with severe malaria anaemia. Br J Haematol 2024; 205: 699–710.

Chintamaneni

Krishna

Suresh

, et al. Deep learning-based diagnostic model for automated detection of monkeypox: introducing MonkeypoxNet. Trait Signal 2024; 41: 493–502.

10.

Shashikiran

Sunitha

. Malaria cell identification using improved machine learning and modified deep learning architecture. Indones J Electr Eng Comput Sci 2024; 34: 2078–2086.

11.

Amin

Anjum

Ahmad

, et al. Microscopic parasite malaria classification using best feature selection based on generalized normal distribution optimization. PeerJ Comput Sci 2024; 10: e1744.

12.

Raju

Naveen Reddy

Keerthi

, et al. Deep learning-based detection of malaria infection through blood sample analysis for malaria diagnosis. Hist Med 2024; 10: 1–10.

13.

Muhammad

Iqbal

Majeed

, et al. Malaria cell classification through exercising deep learning algorithms. JCBI 2024; 7: 53–61.

14.

Hoyos

. Supporting malaria diagnosis using deep learning and data augmentation. Diagnostics 2024; 14: 690.

15.

Dev

Fouda

Kerby

, et al. Advancing malaria identification from microscopic blood smears using hybrid deep learning frameworks. IEEE Access 2024; 12: 71705–71715.

16.

Asif

Ur Rehman Khan

Zheng

, et al. Mozzienet: a deep learning approach to efficiently detect malaria parasites in blood smear images. Int J Imaging Syst Technol 2024; 34: e22953.

17.

Kumar

Muchahari

Poonkuntran

, et al. Application of hybrid capsule network model for malaria parasite detection on microscopic blood smear images. Multimed Tools Appl 2024; 84: 3459–3485.

18.

Sukumarran

Hasikin

Khairuddin

ASM

, et al. An optimised YOLOv4 deep learning model for efficient malarial cell detection in thin blood smear images. Parasit Vectors 2024; 17: 188.

19.

Choudhary

Nausheen

Khan

, et al. The design of an efficient bioinspired CNN model for automated malaria detection in blood smear images. AIP Conf Proc 2024; 3167: 030020.

20.

Murmu

Kumar

. Dlrfnet: deep learning with random forest network for classification and detection of malaria parasite in blood smear. Multimed Tools Appl 2024; 83: 63593–63615.

21.

Kundu

Anguraj

Bhattacharyya

. Utilizing image analysis with machine learning and deep learning to identify malaria parasites in conventional microscopic blood smear images. Trait Signal 2024; 41: 343.

22.

Ramos

Vieira

IHP

Rocha

, et al. A transfer learning approach to identify plasmodium in microscopic images. PLoS Comput Biol 2024; 20: e1012327.

23.

Özbilge

Güler

Özbilge

. Ensembling object detection models for robust and reliable malaria parasite detection in thin blood smear microscopic images. IEEE Access 2024; 12: 60747–60764.

24.

Chaharou

IML

Lawani

Dagba

, et al. Image cropping for malaria parasite detection on heterogeneous data. J Microbiol Methods 2024; 225: 107022.

25.

Swaminathan

Thanuja

Krishna

, et al. A deep learning-based malarial parasite detection using blood smear images for healthcare techniques. In: Approaches to human-centered AI in healthcare. IGI Global, 2024, pp.214–235.

Deep learning-based classification of malaria through parasitized and uninfected blood smear image analysis

Abstract

Keywords

1 Introduction

2 Literature review

3.1.2 CNN for spatial-temporal feature extraction

3.2.4 Determine best split

3.2.6 Repeat for remaining nodes

3.2.7 Ensemble aggregation

4.1 Simulation environment

4.2 Dataset

4.3.2 Batch size

4.3.3 Number of epochs

4.3.4 Optimizer

4.3.5 Weight initialization

4.3.6 Activation function

4.3.7 Regularization technique

4.3.8 Learning rate

4.3.9 Depth of network

4.3.10 Attention mechanism

4.3.11 Loss function

4.3.12 Early stopping

4.4 Prediction results

Table 6. Performance comparison of various classification methods. Proposed Metric GNDO 11 MOZZIENet 16 DLRFNet 20 Malaria-Net Accuracy (%) 96.796 97.740 98.805 99.937 Precision (%) 96.232 97.852 98.240 99.669 Recall (%) 96.324 97.951 98.822 99.337 F1-Score (%) 96.924 97.265 98.848 99.539

Table 7. Ablation study of proposed Malaria-Net. Metric Malaria-Net without Preprocessing Malaria-Net without PSA-CNN feature extraction Overall Malaria-Net Accuracy (%) 97.324 98.099 99.937 Precision (%) 97.926 98.708 99.669 Recall (%) 97.618 98.245 99.337 F1-Score (%) 97.988 98.724 99.539

Footnotes

Funding

Declaration of conflicting interests

Author Biographies

References

Table 7.
Ablation study of proposed Malaria-Net.

Metric Malaria-Net without Preprocessing Malaria-Net without PSA-CNN feature extraction Overall Malaria-Net

Accuracy (%) 97.324 98.099 99.937

Precision (%) 97.926 98.708 99.669

Recall (%) 97.618 98.245 99.337

F1-Score (%) 97.988 98.724 99.539