Deep learning approach for automated hMPV classification

Abstract

Human metapneumovirus (hMPV) is a significant cause of respiratory illness, particularly in children, elderly individuals, and immunocompromised patients. Despite its clinical relevance, hMPV poses diagnostic challenges due to its symptom similarity with other respiratory illnesses, such as influenza and respiratory syncytial virus (RSV), and the lack of specialized detection systems. Traditional diagnostic methods are often inadequate for providing rapid and accurate results, particularly in low-resource settings. This study proposes a novel deep learning framework, referred to as hMPV-Net, which leverages Convolutional Neural Networks (CNNs) to facilitate the precise detection and classification of hMPV infections. The CNN model is designed to perform binary classification by differentiating between hMPV-positive and hMPV-negative cases. To address the lack of real-world patient data, simulated image datasets were used for model training and evaluation, allowing the model to generalize to various clinical scenarios. A key challenge in developing this model is the imbalance within the dataset, where hMPV-positive cases are often underrepresented. To mitigate this, the framework incorporates advanced techniques such as data augmentation, weighted loss functions, and dropout regularization, which help to balance the dataset, improve model robustness, and enhance classification accuracy. These techniques are crucial in addressing issues such as overfitting and generalization, which are common when working with limited datasets in medical imaging tasks. The dataset used for model training and testing consists of 10,000 samples, with an equal distribution of hMPV-positive and hMPV-negative cases. Experimental results demonstrate that the hMPV-Net model achieves a high test accuracy of 91.8%, along with impressive test precision, recall, and F1-score values around 92%. These metrics indicate that the model performs exceptionally well in classifying both hMPV-positive and hMPV-negative cases. Furthermore, the model exhibits superior computational efficiency, requiring only 3.2 GFLOPs, which is significantly lower than other state-of-the-art models such as ResNet-50 and VGG-16. This reduction in computational cost makes the model suitable for deployment in resource-constrained healthcare environments, where computing power and infrastructure may be limited.

Introduction

Human metapneumovirus (hMPV) is a significant respiratory pathogen that was first discovered in 2001. It is recognized for causing acute respiratory tract infections in individuals of all ages, but it can be especially severe in children, older adults, and those with weakened immune systems¹. Timely and accurate diagnosis of hMPV is essential for clinical decision-making and effective patient management². However, its clinical symptoms-ranging from cough and wheezing to bronchiolitis and pneumonia-often overlap with those of other viral infections such as influenza and respiratory syncytial virus (RSV), rendering differential diagnosis difficult^3,4. In many settings, the lack of rapid, reliable, and pathogen-specific diagnostic tools exacerbates delays in treatment. Recent progress in Artificial Intelligence (AI)-especially in the areas of Machine Learning (ML) and Deep Learning (DL)-holds significant promise for revolutionizing clinical diagnostics. These technologies are increasingly leveraged in diverse fields-ranging from communication systems and smart cities to cybersecurity and industrial automation^{5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20}-due to their ability to process vast datasets and learn complex representations. In healthcare, DL models have demonstrated superior capabilities in areas such as medical image classification, disease prediction, drug discovery, and robotic-assisted surgery^21,22.

Among DL techniques, Convolutional Neural Networks (CNNs) have achieved significant success in medical imaging applications, particularly in the detection of pneumonia and COVID-19 using chest radiographs^23,24. Their ability to learn spatial hierarchies of features makes CNNs highly effective for visual recognition tasks in medical diagnostics. However, despite their success in related domains, CNN-based approaches for hMPV detection remain notably underexplored^25,26. Existing models are often generalized for multiple respiratory pathogens, lacking the granularity required for high-accuracy hMPV-specific classification. A major technical barrier in developing such models is the issue of dataset imbalance. In most real-world clinical datasets, hMPV-positive cases are significantly underrepresented relative to healthy or negative cases, which can lead to biased learning, reduced sensitivity, and an inflated false negative rate²⁷. This poses a serious challenge for real-time inference and limits deployment in low-resource healthcare settings. To address these challenges, we propose hMPV-Net, a novel, lightweight, and computationally efficient CNN-based framework specifically designed for the detection and classification of hMPV infections from chest radiographic images. Unlike generic models for broad respiratory disease detection, hMPV-Net is tailored to the unique pathological characteristics of hMPV and integrates domain-specific enhancements to improve diagnostic performance. The key contributions of this study include:

• Development of a pathogen-specific DL architecture optimized for hMPV detection using chest X-ray imagery.

• Use of data augmentation and weighted loss functions to address class imbalance and improve model robustness.

• Incorporation of dropout regularization and hyperparameter tuning to minimize overfitting and enhance generalization.

• Design of a computationally lightweight model architecture suitable for deployment in real-time and resource-limited settings.

• Comprehensive evaluation using clinically relevant performance metrics^28,29,30.

In summary, this study aims to fill a critical gap in the application of deep learning to pathogen-specific diagnostics by introducing an end-to-end hMPV detection framework that is accurate, efficient, and deployable. By leveraging the strengths of CNNs and addressing the current limitations of class imbalance, overfitting, and resource constraints, hMPV-Net contributes to the growing field of AI-driven clinical decision support systems and offers a practical solution for improving respiratory disease diagnostics.

Literature survey on AI-based hMPV detection

Deep learning methods, especially CNNs, have proven to be highly effective in identifying viral diseases, including respiratory illnesses such as pneumonia^{31,32,33,34,35,36}. These models have demonstrated a strong ability to identify abnormalities in medical imaging, such as chest X-rays, with remarkable accuracy. For instance, Wahid et al. utilized an Enhanced Restricted Boltzmann Machine (ERBM) for pneumonia detection, highlighting deep learning’s capability to process complex medical datasets effectively³⁷. Similarly, Himeur et al. and Singh et al. investigated advanced CNN architectures and preprocessing methods to improve the diagnostic accuracy of chest X-ray analysis^38,39. These studies highlight the versatility of CNNs in tackling diagnostic challenges within the healthcare sector. However, most existing approaches are designed for broader respiratory diseases rather than focusing on specific pathogens like hMPV. Additionally, methods that combine CNNs with advanced techniques, such as Vision Transformers, have been explored to enhance performance and precision. Despite their success, these approaches often require substantial computational resources and a large amount of labeled data, which can restrict their implementation in resource-constrained environments. This highlights the need for lightweight, pathogen-specific models that are computationally efficient and capable of addressing these limitations in real-world clinical applications.

While CNN-based methods have shown promise for detecting respiratory illnesses, their application to hMPV detection remains limited. Studies such as those by Costa-Filho et al. have focused on clinical challenges associated with hMPV but lack computational frameworks tailored to address these issues¹. Additionally, research in pneumonia detection, including studies by Kaya et al. and Siddiqi et al., has largely neglected the issue of dataset imbalance, which often leads to biased predictions and reduced model accuracy^40,41. This is particularly significant for hMPV detection, where positive cases are frequently underrepresented in clinical datasets. Another critical gap in existing studies is the limited generalizability of CNN-based models across diverse populations and settings. Most models are trained and validated on specific datasets, making them less effective when applied to broader clinical environments⁴². Furthermore, the computational complexity of advanced approaches, such as those involving Vision Transformers, poses challenges for their adoption in low-resource healthcare settings. Addressing these limitations is essential for advancing hMPV diagnostics and ensuring reliable performance in diverse clinical scenarios.

Table 1 Literature survey on AI-based hMPV detection.

Several research studies have explored the use of DL, ML, and ensemble models to improve diagnostic accuracy. This literature survey, as shown in Table 1, provides insights into existing methodologies, focusing on key techniques, tools, and challenges associated with AI-based detection systems. Costa-Filho et al.¹ investigated the clinical challenges associated with hMPV detection in severe pneumonia cases. Their study focused on clinical analysis but highlighted limitations in generalizing findings to broader datasets. Alshanketi et al.⁴³ implemented DL and transfer learning for pneumonia detection from chest X-ray images. Their work utilized TensorFlow, data augmentation techniques, and transfer learning to handle imbalanced datasets. Alapat et al.⁴⁴ conducted a review on pneumonia detection using neural networks, emphasizing CNN-based preprocessing techniques. However, their study noted a lack of large-scale clinical validation, which affects model reliability in real-world applications. Siddiqi and Javaid et al.⁴¹ performed a comprehensive survey of deep learning architectures for pneumonia detection, highlighting various CNN architectures and identifying issues such as the lack of implementation details and dataset diversity.

Vision Transformers (ViTs) have also been explored as an alternative to CNNs. Singh et al.³⁹ proposed an efficient pneumonia detection model using ViTs, demonstrating improved accuracy but at the cost of high computational requirements. In another study, Kundu et al.⁴⁵ developed an ensemble model, incorporating CNNs, ResNet, and VGG to enhance pneumonia classification accuracy. Despite improved model performance, they identified increased computational complexity and longer inference times as significant drawbacks. Traditional machine learning approaches have also been applied to pneumonia detection. Lee and Wong et al.⁴⁶ used support vector machines (SVM) and feature extraction for classification. However, their model was found to be less accurate compared to DL-based methods. Hasan et al.⁴² examined advanced deep learning techniques for pneumonia detection, integrating CNNs and ensemble learning. Their study underscored the need for extensive labeled datasets to enhance model accuracy.

Mabrouk et al.⁴⁷ further explored ensemble CNN models for pneumonia detection, incorporating data augmentation techniques. While ensemble models improved accuracy, their computational overhead was noted as a challenge for deployment. Akhter et al.⁴⁸ provided a review of AI-based radio-diagnosis using chest X-rays, emphasizing CNNs and transfer learning. Their study pointed out the lack of specific use-case details, which limits the direct applicability of their findings.

The proposed framework distinguishes itself by addressing the limitations of existing methodologies with targeted innovations⁴⁹. Unlike previous studies that primarily focus on general respiratory diseases, this framework is specifically designed for the prediction and detection of hMPV. By leveraging CNNs for binary classification, the model effectively differentiates between hMPV-positive and hMPV-negative cases⁵⁰. It incorporates techniques like data augmentation and weighted loss functions to address dataset imbalances, enhancing the robustness and fairness of predictions compared to traditional CNN-based methods⁵¹. Unlike ensemble-based approaches, which often require significant computational overhead⁴⁵, this model employs a streamlined architecture with dropout regularization and hyperparameter tuning⁵³. These optimizations ensure reliable performance while addressing issues of dataset imbalance and limited generalizability⁵². As a result, the framework represents a significant advancement in AI-driven diagnostics for hMPV, providing an effective and practical solution to existing challenges in the field^48,54.

Summary table of recent key studies

Table 2 Summary of recent key studies on hMPV and respiratory virus detection.

Table 2 provides a summary of recent key studies related to hMPV and respiratory virus detection, highlighting their focus, techniques, and key findings. The table serves to provide insights into current challenges, methodologies, and the emerging trends in the detection of respiratory viruses, particularly hMPV, in clinical settings. The table summarizes key research efforts in the field of AI-based diagnostics for hMPV detection and related respiratory diseases, offering insights into the methodologies, challenges, and progress made in the area. As shown in Table 2, the studies contribute to various advancements such as the application of deep learning techniques, dataset expansion, and efforts to address diagnostic challenges such as sensitivity, dataset imbalance, and clinical applicability in different settings.

In summary, AI-based approaches for hMPV and pneumonia detection have evolved through deep learning, transfer learning, and ensemble models. CNN architectures remain dominant in medical image classification, while Vision Transformers provide an emerging alternative. Dataset imbalance, computational complexity, and lack of clinical validation are key challenges that need to be addressed for real-world implementation. Recent hMPV-focused studies from 2023–2025 have drawn attention to emerging epidemiological trends and ongoing diagnostic challenges. Studies have noted the continued global rise of hMPV infections, particularly post-COVID-19, and the diagnostic difficulties stemming from overlapping symptoms with other respiratory viruses. These findings underline the need for rapid, accurate, and scalable diagnostic tools. ML applications for hMPV detection have gained traction. Recent work demonstrates that techniques such as XGBoost and support vector machines perform effectively using clinical and synthetic datasets. Several studies employed strategies to tackle class imbalance (e.g., SMOTE), and focused on boosting sensitivity to minimize false negatives-crucial for clinical reliability. Moreover, integrating features like fever duration, patient age, and symptom profile improved both interpretability and diagnostic accuracy. Broader developments in deep learning for respiratory virus detection have seen the deployment of CNNs and vision transformers (ViTs) for conditions like pneumonia and COVID-19. Research has focused the importance of dataset diversity and the use of modern architectures (e.g., EfficientNet, DenseNet, InceptionResNetV2) for better generalization across populations and imaging settings. Regularization strategies, class-weighted loss, and augmentation have been particularly effective in mitigating overfitting. Increasing focus is also placed on explainable AI (XAI), where models incorporate tools like Grad-CAM and saliency maps to increase transparency and trust among clinicians. Additionally, several studies argue for rigorous clinical validation and seamless integration into existing healthcare workflows.

Proposed methodology

The hMPV-Net framework is designed to automate the detection of hMPV using DL, specifically a CNN. The framework follows a structured pipeline comprising data preprocessing, feature extraction, model training, and performance evaluation, ensuring robust classification accuracy. The first stage of the framework involves image preprocessing, where raw datasets undergo resizing, denoising, and data augmentation techniques. Once preprocessing is complete, feature extraction is performed to highlight edge features, shape and texture characteristics, and complex patterns present in the images. This extracted feature set forms the foundation for classification, helping the CNN model distinguish between hMPV-positive and hMPV-negative cases effectively. In the next stage, the dataset is split into training and testing subsets to develop and validate the model. An optimal feature subset is selected to enhance learning efficiency and reduce computational complexity. The CNN-based classification model undergoes training, where it learns key distinguishing features from the images through multiple convolutional layers. The model then classifies new input samples into either hMPV-positive or hMPV-negative categories. Figure 1 illustrates the overall workflow of hMPV-Net, depicting the structured approach to leveraging deep learning for hMPV detection and classification. The combination of preprocessing techniques, optimized feature selection, and CNN-based classification ensures high performance and adaptability in medical image analysis. The CNN-Based hMPV Virus Detection Algorithm is shown in Algorithm 1 and leverages DL to automate the detection of hMPV infections in medical imaging. The structured pipeline consists of image preprocessing, feature extraction, model training, evaluation, and binary classification to ensure accurate and reliable detection. The final binary classification output predicts whether new samples are hMPV-positive or hMPV-negative, offering an efficient, AI-driven approach for early and accurate hMPV detection in medical imaging.

Algorithm 1

CNN-based hMPV virus detection algorithm.

Fig. 1

hMPV-Net: a CNN-based framework.

hMPV image dataset preparation and preprocessing

This subsection covers the preparation and preprocessing of the hMPV image dataset for DL classification. It includes 10,000 balanced images from fluorescence microscopy, CT scans, and chest X-rays. Key preprocessing steps, such as image resizing, normalization, noise reduction, contrast enhancement, and data augmentation, are applied to standardize the data, improve quality, and enhance model robustness. These steps ensure that the dataset is optimized for accurate hMPV detection using DL techniques.

Dataset overview and visual characteristics

In order to support the development of a robust DL-based classification model, a well-structured dataset is crucial. The dataset utilized in this study consists of 10,000 images, with an even distribution between hMPV-positive and hMPV-negative samples. This balanced dataset ensures that the model can learn effectively without bias toward any particular class. The images in the dataset are sourced from fluorescence microscopy, computed tomography (CT) scans, and chest X-ray modalities. Each image undergoes rigorous preprocessing to standardize and optimize the input for deep learning algorithms, thereby enhancing the model’s accuracy and generalization. Figure 2 illustrates a sample of fluorescence microscopy images from the dataset. The left panel displays a colored image where fluorescence markers have been used to stain infected cells⁶². The yellow and green regions indicate viral presence and highlight areas of intense replication. The grayscale conversion of the same image on the right demonstrates how the absence of color affects the way structural and intensity-based patterns are represented. Despite the lack of color, the grayscale image maintains important structural features that are crucial for hMPV detection. This visual comparison aids in understanding how both the color and texture of the images contribute to the diagnostic process.

Fig. 2

Sample hMPV-infected cell image (left) and its grayscale version (right).

Preprocessing pipeline for deep learning readiness

Preprocessing is a fundamental phase in any ML or DL model development. For hMPV detection, preprocessing ensures the quality of input images, mitigating issues such as image distortions, noise, and intensity variations. The preprocessing pipeline for the hMPV dataset includes multiple critical steps, each aimed at optimizing the data for training a CNN. These steps, ranging from image collection to data augmentation, enhance model performance by improving the quality and diversity of the data.

Image collection and curation

The first stage of dataset preparation involves image collection, ensuring that the images are correctly labeled and meet the quality standards required for effective training. The dataset consists of both infected (hMPV-positive) and non-infected (hMPV-negative) samples. These samples are collected from different imaging techniques, including fluorescence microscopy, CT scans, and chest X-rays, each providing valuable perspectives on cellular and tissue changes due to the viral infection. Prior to preprocessing, quality control measures are implemented to remove any low-quality images, images with incomplete or inconsistent labeling, and those that may contain distortions due to acquisition problems. Additionally, manual annotations are verified to ensure accuracy, and automated checks are employed to filter out defective or irrelevant samples. This step is essential to avoid bias in the dataset.

Image resizing

Medical images are often acquired at varying resolutions due to differences in imaging devices and protocols. To standardize these images and prepare them for processing by the CNN, all images are resized to a uniform dimension of $64 \times 64 \times 3$ pixels. Resizing is accomplished using bilinear interpolation, which ensures that the spatial relationships in the image are preserved and minimizes distortion. Bilinear interpolation is preferred as it helps maintain the visual integrity of the image while reducing computational overhead. This step ensures that all images, regardless of their original size or resolution, are standardized and ready for input into the CNN.

Normalization of pixel intensities

Normalization of pixel values is a critical preprocessing step to ensure that the model learns in a stable and efficient manner. In the case of the hMPV image dataset, Min-Max normalization is applied to scale pixel values to the range $[0,1]$. This transformation standardizes the intensity values across all images, enabling the CNN to process them more effectively. Normalization has several key benefits:

• It reduces the impact of lighting and contrast variations across images, ensuring that the model focuses on the structural features rather than on inconsistencies in pixel intensity.

• It accelerates the training process by improving the convergence of the optimization algorithm.

• It enhances the stability of the learning process by preventing issues such as exploding or vanishing gradients during backpropagation.

By applying Min-Max normalization, the pixel intensity distributions are made consistent across the entire dataset, which facilitates more efficient training and better model performance.

Noise reduction

Medical images are often prone to various forms of noise, which can obscure important structural information required for accurate classification. Noise can be introduced by multiple factors, including sensor limitations, image acquisition conditions, and biological artifacts. To address this issue, two types of noise reduction filters are applied:

• Gaussian Filtering: Gaussian filtering is applied to smooth the image and suppress high-frequency noise while retaining significant structural features. The Gaussian kernel helps in reducing random variations in the image, providing a cleaner representation of the underlying features.

• Median Filtering: Median filtering is effective in removing salt-and-pepper noise, which is often present in microscopy images. Unlike Gaussian filtering, median filtering preserves edges better, which is crucial for detecting cellular boundaries and other morphological features in hMPV-infected samples.

Both filtering methods collaborate to enhance image clarity by minimizing noise and making sure that essential features remain detectable for effective feature extraction.

Contrast enhancement

Contrast plays a crucial role in making features within medical images more visible. Subtle intensity differences are often essential for distinguishing between healthy and affected regions. To boost contrast and highlight important structural elements, Adaptive Histogram Equalization (AHE) is commonly used. This technique improves local contrast by splitting the image into smaller sections and performing histogram equalization independently within each segment. As a result, areas with minimal intensity changes become clearer, which helps the model better identify fine details. By enhancing contrast locally, AHE ensures that significant features-such as cell boundaries and viral inclusions-stand out more distinctly and are easier to differentiate.

Data augmentation

Expanding the diversity of training data through augmentation helps models become more adaptable by introducing previously unseen samples. This approach broadens the dataset by applying a range of transformation techniques, such as:

• Rotation: Images are randomly turned within a specified angle range (for example, $\pm 15^{\circ }$) to mimic various possible orientations of the target objects.

• Flipping: Both horizontal and vertical flips are performed to make the model less sensitive to the direction or position of features in the images

• Scaling: Random zoom-in transformations simulate different sizes of objects within the image, helping the model become invariant to size variations.

These transformations increase the dataset size and variability, ensuring that the model learns spatial invariance and becomes robust to changes in orientation, positioning, and scale that are commonly encountered in real-world clinical settings.

Each step of the preprocessing pipeline contributes to improving the CNN’s ability to detect and classify hMPV infections from medical images. By standardizing the input images, normalizing pixel values, reducing noise, enhancing contrast, and applying data augmentation, the preprocessing pipeline ensures that the CNN receives high-quality input data that is both consistent and representative of real-world variations. These improvements in data quality enable the model to learn more effectively, resulting in a robust and accurate hMPV detection model.

Feature extraction and feature set representation

Feature extraction is a cornerstone in DL-based human metapneumovirus (hMPV) detection. This process enables the model to learn spatial patterns and critical structural features necessary for accurate classification. In medical image analysis, feature extraction allows the model to identify and understand complex visual structures, such as edges, textures, and shapes, which are essential for distinguishing hMPV-positive samples from hMPV-negative samples. CNNs, due to their hierarchical structure, are highly effective in learning such patterns automatically from raw pixel data. This section details the CNN architecture employed for hMPV detection and describes the role of each layer in the feature extraction process, followed by the importance of feature set representation and selection for classification. The CNN model used for this task is composed of five convolutional layers, interspersed with ReLU activation functions and max-pooling layers. These layers are followed by a fully connected layer and an output layer. This structured architecture ensures efficient feature extraction while maintaining computational efficiency. Below is a detailed explanation of each layer in the CNN and its contribution to feature extraction.

Convolutional layers for hierarchical feature learning

The layered structure of CNNs enables them to automatically identify spatial structures and extract relevant features from visual data. These networks utilize adaptable matrices-often called filters-that are systematically applied to the incoming data to recognize both basic and advanced spatial arrangements. In medical image processing, these operations focus on identifying critical properties such as boundaries, textures, and intricate configurations, which are essential for distinguishing between different types of clinical images. Each convolutional stage shifts its set of weights across the input, performing a dot product between the filter and the corresponding section of the data, and generates a transformed output that highlights the detected patterns. For instance, the initial convolutional stage might employ 32 filters, each of dimension $4 \times 4$, to extract fundamental features like contours and gradients. Following this, a nonlinear activation function-such as Leaky ReLU-is applied to the output of the convolution, introducing nonlinearity and enabling the network to capture more sophisticated relationships beyond simple linear combinations. The convolution process can be mathematically described as:

$$\begin{aligned} Q(a, b) = \sum _{u} \sum _{v} D(a+u, b+v) \cdot M(u, v), \end{aligned}$$

(1)

where $D(a, b)$ is the input image, $M(u, v)$ is the filter, and $Q(a, b)$ is the resulting feature map. This operation is performed at every location as the filter moves over the input, producing a spatial representation of learned features. The second convolutional layer (Convolutional Layer 2) uses another 32 filters of size $3 \times 3$. This layer refines the representations captured by Layer 1 and begins to detect more complex features such as corners and cellular edges. The ReLU activation function is applied again to capture non-linear patterns. At this level, the network becomes capable of recognizing more intricate visual cues indicative of hMPV infection. Max Pooling Layer 1 is applied, using a pooling window of size $2 \times 2$. This down-sampling operation reduces the dimensions of the feature map $Q(a, b)$ by half. The pooling function selects the maximum value from each $2 \times 2$ region:

$$\begin{aligned} P(a, b) = \max \left\{ Q(a+i, b+j) \right\} , \quad i,j \in \{0,1\}, \end{aligned}$$

(2)

resulting in translation invariance and reduced computational complexity. The third convolutional layer (Convolutional Layer 3) applies 64 filters of size $3 \times 3$, capturing more abstract and high-level features from the feature maps produced earlier. These include patterns related to cell morphology and possible viral structures relevant to hMPV. The fourth convolutional layer (Convolutional Layer 4), also with 64 filters of size $3 \times 3$, further enhances the discriminative ability of the network. It detects subtle morphological changes such as variations in nuclear shape or cytoplasmic boundaries that are critical in medical diagnostics. Another max-pooling operation (Max Pooling Layer 2) with window size $2 \times 2$ is applied afterward to compress the data and reduce spatial redundancy. Finally, the fifth convolutional layer (Convolutional Layer 5) uses 128 filters of size $3 \times 3$, learning the most abstract patterns necessary for the classification task. These patterns are vital for distinguishing hMPV-infected cells from non-infected ones. The extracted features are then flattened and passed to fully connected layers for classification.

Output layer for binary classification

After hierarchical feature extraction, a single neuron output layer is used for binary classification of hMPV infection. It employs the sigmoid activation function to map the output to a probability value. The decision rule is:

$$\begin{aligned} P(\text {hMPV}) \ge 0.5 \Rightarrow \text {hMPV-Positive}, \quad P(\text {hMPV}) < 0.5 \Rightarrow \text {hMPV-Negative}. \end{aligned}$$

(3)

Mathematical representation of CNN feature extraction

Feature extraction via CNNs involves convolution, activation, and pooling operations. After convolution, non-linearity is introduced using:

$$\begin{aligned} A(a, b) = \max (0, Q(a, b)). \end{aligned}$$

(4)

Subsequently, max pooling reduces spatial dimensions:

$$\begin{aligned} P(a, b) = \max \left\{ A(a+i, b+j) \right\} , \quad i,j \in \{0,1\}. \end{aligned}$$

(5)

This process builds hierarchical and abstract representations crucial for effective classification.

Feature set representation and selection

After CNN-based extraction, the next step is constructing a representative and optimized feature set for classification. The CNN-derived features are grouped into three primary categories:

• Edge Features: Capture strong intensity transitions at cellular boundaries, extracted by early filters from $D(a, b)$.

• Shape Features: Encode morphological descriptors-like roundness and elongation-identified by deeper convolutional layers using complex compositions of $M(u, v)$.

• Texture Features: Capture pixel intensity variations. Texture is often quantified using Gray-Level Co-occurrence Matrix (GLCM) metrics:

  $$\begin{aligned} P(i, j) = \sum _{a,b} \delta \left( D(a,b)=i, D(a+\Delta a, b+\Delta b)=j \right) , \end{aligned}$$

  (6)

  where $\delta$ is an indicator function, $i, j$ are intensity levels, and $(\Delta a, \Delta b)$ denote pixel displacement.

The final feature vector combines these descriptors:

$$\begin{aligned} \phi (x) = [E, S, T, C], \end{aligned}$$

(7)

where $E$, $S$, and $T$ denote edge, shape, and texture features, respectively, and $C$ includes deeper abstract features learned through convolution layers. This multidimensional representation enables the classifier to accurately differentiate between hMPV-positive and hMPV-negative samples. By combining low-, mid-, and high-level features through deep hierarchical learning, the CNN efficiently captures the relevant spatial and contextual patterns in medical images, making it a robust solution for automated hMPV detection and diagnosis.

Model training and evaluation for hMPV image dataset

The training and assessment procedure represents a vital phase in developing an effective CNN architecture that can accurately identify distinctive characteristics and demonstrate strong generalization capabilities when processing novel data. This stage encompasses data partitioning, neural network training, addressing class distribution imbalances, and performance evaluation through suitable metrics. An effectively organized training methodology enhances the model’s capacity to distinguish between hMPV-positive and hMPV-negative samples while reducing overfitting and enhancing classification precision. The neural network architecture follows a systematic training approach where it acquires feature representations through successive convolutional operations. The training process utilizes an adaptive optimization method, a loss function designed for binary classification tasks, and a mini-batch learning strategy. The Adam optimization algorithm is employed for effective gradient-based learning, combining the benefits of momentum-driven updates (RMSprop) and adaptive learning rates, which enhances convergence speed and stability. Adam is selected for its capability to adjust learning rates dynamically, which assists in maintaining training stability across various feature dimensions. The loss function is formulated as:

$$\begin{aligned} \mathcal {L} = -\frac{1}{B} \sum _{k=1}^{B} [t_k \log (\hat{t}_k) + (1-t_k) \log (1-\hat{t}_k)], \end{aligned}$$

(8)

where $t_k$ represents the actual class label, with a value of 1 for hMPV-positive samples and 0 for hMPV-negative samples. The term $\hat{t}_k$ indicates the predicted probability that a specific sample belongs to the hMPV-positive class. Finally, B represents the total number of samples in a batch, used to calculate the average loss across the batch during training. Binary cross-entropy loss penalizes incorrect predictions more severely when confidence is high, making it particularly suitable for deep learning-based classification applications. Training is performed using a batch size of 32 images per batch, which maintains a balance between computational efficiency and stable gradient updates. This methodology enables the model to process a portion of the dataset simultaneously, enhancing memory utilization and convergence stability. The training procedure involves 30 complete iterations, or epochs, across the entire dataset. Each epoch allows the model to iteratively improve its learned feature representations, gradually enhancing its capability to distinguish between hMPV-positive and hMPV-negative samples. The integration of batch-wise learning and multiple epochs ensures gradual model improvement while preventing overfitting.

Simulation results and analysis

The Simulation evaluation of the proposed CNN-based hMPV-Net framework demonstrates its effectiveness in detecting hMPV cases with high accuracy and reliability. The dataset used in the experiments consists of 10,000 image samples, evenly distributed between hMPV-positive and hMPV-negative cases, ensuring a balanced evaluation of the model. Each sample underwent essential preprocessing techniques, including normalization, data augmentation (rotation, flipping, scaling), and feature extraction. The hMPV-Net framework was trained using an adaptive learning optimization technique to facilitate optimal convergence. The training process was conducted with a batch size of 32 and over 30 epochs. The model achieved a training accuracy of 99.2 percent, validation accuracy of 93.5 percent, and test accuracy of 91.8 percent, with a final loss value of 0.2154 and a ROC-AUC score of 0.91. These results highlight the model’s superior generalization ability, minimizing overfitting and significantly improving over previous small-scale experiments.

Test set performance analysis

Fig. 3

Confusion matrix for test set.

Table 3 Summary of model performance metrics on test set.

The visualization in Fig. 3 presents the confusion matrix of the hMPV detection model on the test set. Specifically, the model correctly identified 462 cases as hMPV-positive (true positives) and 456 cases as hMPV-negative (true negatives). There were 38 instances where hMPV-negative cases were incorrectly classified as positive (false positives), and 44 hMPV-positive cases were missed (false negatives). The testing set comprised 1,000 samples, with a nearly balanced distribution of 506 actual hMPV-positive and 494 actual hMPV-negative cases. The model achieved excellent results across all major performance metrics on the testing set. The overall accuracy was 91.80%, with 918 out of 1,000 cases correctly classified. Macro-averaged precision, recall, and F1-score were all approximately 91.8%, indicating that the model performs consistently well across both classes. This balanced performance suggests minimal bias toward either hMPV-positive or hMPV-negative predictions, which is crucial for reliable clinical application. A closer examination of class-specific metrics reveals strong results for both hMPV-positive and hMPV-negative cases. For the hMPV-positive class, the precision was 92.40%, meaning that when the model predicted a positive case, it was correct 92.4% of the time. The recall (sensitivity) for this class was 91.30%, reflecting the model’s ability to detect actual positive cases. The F1-score, which balances precision and recall, was 91.85%. For the hMPV-negative class, the precision was 91.20%, recall was 92.31%, and the F1-score was 91.75%.

From a clinical perspective, the model’s sensitivity (true positive rate) of 91.30% indicates a high likelihood of correctly identifying patients with hMPV infection. The specificity (true negative rate) of 92.31% shows that the model is also effective at correctly excluding non-hMPV cases. The positive predictive value (PPV) of 92.40% and negative predictive value (NPV) of 91.20% further confirm the model’s reliability in both positive and negative predictions, supporting its potential utility in diagnostic workflows.Analysis of the model’s errors reveals a false negative rate of 8.70% (44 out of 506 actual positive cases missed) and a false positive rate of 7.69% (38 out of 494 actual negative cases incorrectly flagged). The slightly higher false negative rate suggests a minor tendency toward under-detection of hMPV cases. In clinical settings, this may be preferable to a high false positive rate, as it reduces the risk of unnecessary treatments for non-infected individuals. When comparing the testing set results to those obtained on the training set, the model demonstrates consistent generalization capability with minimal evidence of overfitting. Both datasets achieved an accuracy of approximately 91.8%, indicating robust performance across different data distributions. This consistency underscores the model’s reliability and suitability for deployment in real-world diagnostic scenarios. Table 3 presents a summary of the model’s performance metrics on the testing set, all indicating strong classification performance. In summary, the hMPV detection model exhibits high effectiveness for diagnostic applications, with balanced and clinically acceptable performance across both positive and negative classes.

Fig. 4

Accuracy vs epochs for the proposed model.

Figure 4 presents the accuracy trend of the proposed model over multiple training epochs. The graph showcases both training and validation accuracy, illustrating the model’s learning progression. The training accuracy curve demonstrates a steady increase, reflecting effective feature learning by the model. Similarly, the validation accuracy follows an upward trajectory, suggesting strong generalization capabilities. As the epochs progress, the accuracy of both training and validation improves, with the validation accuracy reaching close to 90 percent. The relatively small gap between the two curves indicates that the model is well-optimized, with minimal overfitting. The smooth and consistent rise in accuracy highlights the model’s stable learning process, reinforcing its effectiveness in distinguishing between hMPV-positive and hMPV-negative cases. This visualization further validates the robustness of the proposed CNN-based framework for reliable medical image classification.

Fig. 5

Loss vs epochs for the proposed model.

Fig. 6

ROC-AUC score vs epochs for the proposed model.

Figure 5 shows how the loss values for the developed model evolve during several rounds of training. The graph displays both training and validation losses, with the horizontal axis indicating the number of training cycles and the vertical axis representing the loss values. As the number of training cycles increases, both training and validation losses drop considerably, which demonstrates the model’s capability to reduce classification mistakes. The small difference between the two curves indicates that the model maintains its accuracy on unseen data. The consistent decline in loss values underscores the stability of the training process and supports the reliability of the proposed detection framework. Figure 6 depicts how the ROC-AUC score changes throughout the training cycles, highlighting the model’s capacity to distinguish between positive and negative cases. The graph shows a steady rise in the ROC-AUC score, reflecting ongoing improvements in the model’s ability to separate the two classes. In the early training phases, the ROC-AUC score is low but increases quickly as the model learns to extract and recognize significant features. As training continues, the ROC-AUC score exceeds 0.90, confirming the model’s strong performance in distinguishing between the two categories. The upward trend of the curve indicates that the model achieves stable learning and improved predictive power. The final ROC-AUC score, which approaches 0.97, further confirms the effectiveness of the proposed CNN-based detection method, suggesting its potential usefulness in real-world diagnostic applications.

Statistical validation of model performance metrics

To rigorously evaluate the statistical significance of the model’s classification performance and learning behavior over time, we conducted a series of ANOVA tests, the results of which are summarized in Tables 4, 5, and 6.

One-way ANOVA for confusion matrix metrics

To validate the model’s classification capability, we analyzed the confusion matrix derived from the testing set, as shown in Fig. 3. A one-way ANOVA was applied to compare the classification frequencies of True Positives (TP = 462), False Positives (FP = 38), and False Negatives (FN = 44). The results indicated highly significant differences between:

• TP and FP: $F(1,498) = 719.1, p < 0.001$

• TP and FN: $F(1,504) = 690.6, p < 0.001$

Post-hoc comparisons revealed substantial mean differences of 424 between TP and FP (95% CI [380.2, 467.8], $p < 0.001$), and 418 between TP and FN (95% CI [373.9, 462.1], $p < 0.001$). These results, summarized in Table 4, confirm that the model exhibits a statistically significant ability to correctly identify true positive cases while maintaining low false positive and false negative rates. The large effect sizes and F-values further validate the robustness and discriminatory power of the model’s classification performance.

Table 4 One-way ANOVA results for confusion matrix metrics from testing set.

Two-way ANOVA for accuracy and loss trends over epochs

To further assess the evolution of model performance, we conducted two-way ANOVA tests using the accuracy and loss values recorded across epochs for both training and validation phases, as visualized in Figs. 4 and 5. This analysis aimed to examine the main effects of training condition (training vs. validation), epoch number, and their interaction.

Accuracy Analysis:

• Training Type: $F(1,90) = 124.7, p < 0.001, \eta ^2 = 0.58$; Epochs: $F(9,90) = 892.3, p < 0.001, \eta ^2 = 0.90$; Interaction: $F(9,90) = 8.42, p < 0.001, \eta ^2 = 0.46$

Loss Analysis:

• Training Type: $F(1,90) = 89.2, p < 0.001, \eta ^2 = 0.50$; Epochs: $F(9,90) = 1247.8, p < 0.001, \eta ^2 = 0.93$; Interaction: $F(9,90) = 6.78, p < 0.001, \eta ^2 = 0.40$

The ANOVA results, detailed in Tables 5 and 6, indicate that both training type and epoch number significantly affect model accuracy and loss. Moreover, the significant interaction effects suggest that the trends of training and validation performance evolve differently over epochs. These findings confirm the model’s stable learning trajectory and generalization ability, supporting the visual interpretations and suggesting minimal risk of overfitting.

Table 5 Two-way ANOVA results for model accuracy over epochs.

Table 6 Two-way ANOVA results for model loss over epochs.

Comparison of existing vs proposed system

Table 7 presents a performance comparison between the proposed hMPV-Net model and existing deep learning systems, such as ResNet-50 and VGG-16. The results demonstrate that hMPV-Net outperforms these established models in several key aspects, including accuracy, computational efficiency, and loss minimization. The proposed hMPV-Net model achieved a test accuracy of 91.8%, significantly higher than the 81.4% test accuracy achieved by both ResNet-50 and VGG-16. Furthermore, the model’s ROC-AUC score of 0.91 indicates superior classification performance. In terms of computational cost, hMPV-Net requires only 3.2 GFLOPs, which is a significant reduction compared to the 15.6 GFLOPs required by ResNet-50 and VGG-16. This reduction in computational cost makes hMPV-Net more suitable for deployment in resource-constrained clinical environments, where high computational power may not be available.

The experimental findings validate that hMPV-Net is a lightweight, pathogen-specific model that incorporates advanced techniques such as class-weighted loss and targeted data augmentation. These features address common challenges, such as dataset imbalance, which are often overlooked in other deep learning models. The proposed framework’s ability to generalize effectively, while maintaining high computational efficiency, emphasizes its potential for real-world applications in AI-powered diagnostics for early hMPV detection. Future work will focus on expanding the dataset for improved robustness, optimizing hyperparameters to boost performance further, and validating the model using real-world patient data. These efforts will contribute to enhancing the clinical applicability of the model and further advance the field of AI-driven medical diagnostics.

Table 7 Performance comparison of proposed hMPV-net and existing DL models.

The performance of hMPV-Net has several notable advantages. First, in terms of Performance, hMPV-Net demonstrates superior accuracy and ROC-AUC scores compared to ResNet-50 and VGG-16, while also significantly reducing computational cost. This makes hMPV-Net more suitable for deployment in resource-constrained clinical environments. In terms of Methodological Innovation, hMPV-Net is specifically designed for hMPV detection, incorporating advanced techniques like class-weighted loss and targeted data augmentation to address dataset imbalance-a challenge often overlooked by existing frameworks. This targeted approach improves the model’s accuracy and ensures better performance with imbalanced datasets. Regarding Generalizability, hMPV-Net utilizes a balanced and diverse dataset combined with robust preprocessing techniques, enhancing its ability to generalize across different imaging modalities and patient populations. This generalization is a critical improvement over existing models, which often struggle to perform well across diverse clinical settings.

The Strengths of the model are evident in its high Computational Efficiency, with a lightweight architecture (3.2 GFLOPs) supporting real-time applications in clinical settings with limited hardware resources. Furthermore, the model’s Robustness to Imbalance, achieved through class-weighted loss and data augmentation techniques, helps mitigate bias from underrepresented hMPV-positive cases, ensuring improved diagnostic fairness. The high Clinical Applicability of the model is demonstrated by its high sensitivity and specificity, positioning it as a valuable tool for early hMPV detection, which can potentially enhance patient outcomes. However, there are some Limitations. First, while simulated data was used to train the model, real-world clinical data is needed to better capture the variability of clinical images, and further validation with patient data is required. Additionally, the model is currently optimized for hMPV detection, but expanding the framework to include multi-pathogen detection could further enhance its clinical utility and make it applicable to a broader range of diagnostic scenarios.

Clinical integration of high-performance models

A 10–15% increase in diagnostic accuracy represents a substantial advancement for medical AI, particularly in the context of hMPV detection, where conventional assays often struggle with limited sensitivity and high false-negative rates. Such improvements directly reduce the incidence of diagnostic errors, leading to enhanced patient safety, better treatment outcomes, and fewer unnecessary invasive procedures. The hMPV-Net model, achieving 91.8% test accuracy with 91.8% testing precision and recall, demonstrates a significant reduction in both false positives and false negatives compared to existing methods. These gains are especially meaningful in clinical practice, as diagnostic errors contribute to a considerable proportion of patient morbidity and mortality, and improvements in sensitivity and specificity are closely linked to tangible clinical benefits.

Analysis of model scaling and dataset size indicates that expanding the hMPV-Net training set by several thousand images would likely yield only minor additional gains in test accuracy-typically 1–3%-due to diminishing returns once the model has learned the primary data patterns from a balanced, diverse 10,000-image dataset. Beyond this point, further improvements depend more on the diversity and clinical relevance of new samples rather than sheer volume. Large-scale dataset expansion also increases computational demands, which may not be justified by marginal accuracy improvements. Learning curve analysis and the small gap between training and test accuracy suggest that the current dataset size is well-matched to the model’s complexity, supporting robust generalization without overfitting. Heuristics such as maintaining at least ten times more data points than model parameters further validate the adequacy of the dataset for the chosen CNN architecture.

From a deployment perspective, hMPV-Net’s computational efficiency (3.2 GFLOPs, substantially lower than models like ResNet-50 or VGG-16) makes it particularly suitable for resource-limited settings where both computational power and large datasets may be unavailable. The model’s strong generalization and efficiency suggest that further accuracy gains through dataset expansion may not justify the associated costs. Instead, future efforts should focus on increasing data diversity and quality, particularly by incorporating rare case types and varied clinical scenarios, to further enhance model generalizability across different patient populations and imaging conditions.

In summary, the performance improvements achieved by hMPV-Net are clinically meaningful and align with standards for impactful AI in medical imaging. While incremental dataset expansion may offer slight gains, the most effective strategy for further advancement is to prioritize high-quality, diverse data and clinical validation with real-world patient populations. This approach will ensure that hMPV-Net remains robust, generalizable, and relevant for deployment in both well-resourced and resource-limited healthcare environments.

Principal challenges and prospective research avenues

Key challenges

Dataset limitations present a significant obstacle for the development and application of hMPV-Net. One of the primary challenges is the reliance on simulated data. Due to the scarcity of large-scale, real-world human metapneumovirus (hMPV) imaging datasets, the model was trained and evaluated primarily on simulated data. While this approach enabled model development, it may not fully capture the variability and complexity that exist in clinical images. As a result, the model’s performance might not be representative of real-world clinical scenarios, which can differ substantially in terms of image quality, patient characteristics, and infection types.

Another critical challenge associated with datasets is class imbalance. In medical imaging, particularly in the case of hMPV, positive cases (i.e., those showing evidence of infection) are often underrepresented. This underrepresentation can lead to biased model training, where the model becomes more attuned to detecting non-infected cases while struggling with accurate identification of infected cases. Although techniques such as data augmentation and class-weighted loss functions were employed to address this issue, class imbalance remains a persistent challenge in medical imaging and must be dealt with more effectively in future iterations of the model.

Generalizability and clinical validation are additional challenges that require attention. While hMPV-Net has shown promising results in initial testing, the model’s performance has not yet been validated on diverse, real-world clinical datasets. This lack of real-world testing raises concerns about the model’s robustness across different patient populations, imaging modalities, and healthcare settings. Without validation on clinical data from various institutions, it remains unclear whether the model can generalize effectively outside of controlled environments.

Moreover, variability in imaging protocols introduces another layer of complexity. Different image acquisition techniques (such as fluorescence microscopy, CT scans, and X-rays) can introduce inconsistencies in the data, which may affect the model’s ability to generalize across different imaging modalities. Variations in imaging protocols-such as resolution, contrast, and noise-are commonplace in clinical environments and can impact the model’s performance, further emphasizing the need for better adaptability and validation across a broader range of imaging systems.

Computational and practical constraints also present challenges. While hMPV-Net is designed for computational efficiency, with a reduced computational cost of 3.2 GFLOPs, deploying AI models in real-world, resource-constrained environments still presents practical hurdles. These include hardware limitations, integration with existing clinical workflows, and ensuring that the model can function effectively on low-cost machines with limited computational power. Additionally, data privacy and accessibility are significant concerns in healthcare. Due to strict privacy laws and regulations, access to real patient data is often restricted. This limits opportunities for large-scale validation and ongoing model improvement, further hindering the model’s development and widespread adoption.

Deep learning models, such as CNNs, are frequently labeled as “black boxes” due to their complex and opaque decision-making processes. This lack of transparency can hinder their acceptance in clinical environments, as healthcare professionals may be hesitant to rely on predictions that cannot be clearly explained. To promote wider clinical adoption and foster trust among medical practitioners, it is crucial to create tools that clarify how these models reach their conclusions.

Future research directions

Clinical validation and dataset expansion will be a priority in future research. The next step is to validate hMPV-Net on real-world clinical datasets collected from multiple healthcare institutions. This real-world testing will be critical in assessing the robustness and generalizability of the model across diverse patient populations and different imaging modalities. Additionally, expanding the dataset is necessary to improve model performance. Collaboration with healthcare providers to collect and curate larger, more diverse datasets, including rare and atypical hMPV cases, will help ensure that the model can handle a broad spectrum of clinical scenarios and improve its overall diagnostic accuracy.

Multi-pathogen and multimodal detection is another area for future work. Extending hMPV-Net to detect multiple respiratory pathogens (e.g., influenza, RSV) within a single model could broaden its clinical utility. By incorporating a multi-pathogen detection framework, hMPV-Net could become a more comprehensive diagnostic tool for respiratory infections, making it highly beneficial in clinical practice. Furthermore, the integration of multimodal data-combining different imaging modalities like CT, MRI, and X-rays-along with clinical metadata (e.g., patient demographics, clinical history) could enhance the model’s diagnostic accuracy. This would allow the model to make more informed decisions, considering multiple factors that influence diagnosis.

Advanced model optimization is essential for improving performance and ensuring that the model adapts to evolving clinical needs. One promising approach is transfer learning, which involves leveraging pre-trained models from other domains to enhance the performance of hMPV-Net, especially when dealing with limited data. This technique can help improve the model’s performance on smaller datasets and facilitate adaptation to new imaging domains. In addition, further exploration of regularization methods and hyperparameter tuning will be necessary to prevent overfitting and optimize the model’s performance across various settings.

Interpretability and explainability will also be a focus in future research. The integration of explainable AI (XAI) tools, such as saliency maps and Grad-CAM, will help provide visual explanations for model predictions, offering transparency and increasing clinician confidence in the model’s outputs. These tools will make the decision-making process more understandable, thereby improving the trustworthiness of the model in clinical practice.

Finally, deployment and integration of hMPV-Net into real-world healthcare settings will require addressing practical issues. Future research will focus on seamlessly integrating the model into existing clinical workflows. This includes developing user-friendly interfaces, real-time decision support, and ensuring that the model can function effectively in real-time clinical environments. In addition, continued efforts will be directed towards optimizing the model for resource-constrained settings, ensuring that it remains accessible and scalable, even in areas with limited computational resources.

Conclusion

The hMPV-Net framework has demonstrated significant advancements in detecting and classifying hMPV-positive and hMPV-negative cases using DL particularly CNNs. The model has achieved 91.8% test accuracy, 91.8% test precision, and 91.8% test recall, showcasing its ability to reliably identify hMPV infections. Its computational efficiency, with a low cost of 3.2 GFLOPs, makes it feasible for deployment in resource-constrained healthcare environments, an essential factor for real-world applications in under-resourced settings. Despite these promising results, there are several opportunities for optimization. Expanding the dataset to increase robustness and generalization across diverse populations is a key area for improvement. This will help balance the representation of hMPV-positive and hMPV-negative cases, as well as address any dataset biases. Moreover, exploring different regularization techniques or advanced CNN architectures could help prevent overfitting and improve performance for clinical use. Clinical validation using real-world patient data is necessary to ensure the practical applicability of the model in diverse clinical settings. Real-world validation will help confirm the model’s robustness across different patient conditions, imaging techniques, and environmental factors. Additionally, integrating AI-powered diagnostic systems like hMPV-Net into routine clinical practice has the potential to revolutionize early hMPV detection, facilitating timely medical intervention and improving patient outcomes by automating and accelerating the detection process. Future advancements could focus on incorporating multimodal imaging (e.g., CT scans, chest X-rays, MRI) to enhance the model’s ability to work with various imaging modalities and improve diagnostic reliability across different systems. The application of transfer learning could also improve the model’s generalization to new or rare hMPV cases, reducing training time and computational costs. This would allow the model to adapt to a broader range of medical conditions and further improve diagnostic accuracy. In conclusion, while the hMPV-Net framework has shown promising results in detecting and classifying hMPV infections, future work involving dataset expansion, hyperparameter optimization, clinical validation, and the incorporation of multimodal imaging and transfer learning will be essential in improving its real-world applicability.