Hyperparameter-Tuned Deep Learning Framework for Accurate COVID-19 Detection Using Chest X-Rays and Computed Tomography Scans

The appearance of Coronavirus Disease 2019 (COVID-19), caused by the new Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2), has created big problems for healthcare systems, economies, and societies around the world [1]. It first appeared in early 2019 and quickly spread everywhere, causing millions of cases and many deaths [2]. Being able to diagnose the virus quickly and accurately is very important to stop its spread, give the right treatment, and reduce how much it spreads in communities [3]. Even though the Reverse Transcription Polymerase Chain Reaction (RT-PCR) test is usually the best way to detect COVID-19, this method has several issues. It can take a long time, mistakes can happen during sample handling, and it might not always find the virus if there's not enough virus in the sample [4]. Because of this, medical imaging tools like chest X-rays (CXR) and computed tomography (CT) scans have become important in helping doctors find and track lung infections caused by COVID-19 [5]. These imaging techniques can show signs of the virus, such as ground-glass opacities, lung infiltrates, and changes in the lungs, helping radiologists understand how serious the illness is and how it is getting worse. But looking at a lot of imaging data by hand during a pandemic takes a lot of time, is error-prone, and can vary between different people looking at the images, which shows how important it is to have fast, reliable, and easy-to-use diagnostic tools [6].

Recent improvements in artificial intelligence (AI), especially deep learning (DL), have exposed great promise in analyzing medical images. DL methods, like convolutional neural networks (CNN), are really good at finding complex patterns in images. They often do better than older techniques that depend on manually created features for tasks like finding diseases, separating parts of an image, and sorting images into categories [7, 8]. In the realm of COVID-19, researchers around the globe have investigated the application of DL systems to identify viral pneumonia characteristics directly from CXR and CT images, intending to present fast, cost-efficient, and reproducible diagnostic assistance [9].

These models can learn to separate COVID-19 infections from other pneumonia types and normal lung conditions by recognizing subtle radiographic indicators that could be overlooked by human observers [10, 11]. Additionally, by employing transfer learning approaches and utilizing pre-trained models like VGGNet, ResNet, DenseNet [12], and InceptionNet, researchers have achieved high diagnostic precision even when working with relatively small COVID-19 imaging datasets—an essential benefit during the pandemic's early days when well-annotated data were limited [13, 14]. Additionally, diagnostic systems utilizing DL might be implemented in real-time in clinical environments, enabling healthcare professionals to prioritize critical cases, enhance resource utilization, and elevate overall patient outcomes [15].

Even with these encouraging advancements, there are still considerable obstacles and areas of research that need to be tackled in order to move DL-based COVID-19 detection organizations from experimental environments into broad clinical practice. Variations in image collection methods across different hospitals, disparities in scanner types and patient positioning, along with the natural diversity in COVID-19 presentations across various demographic groups, introduce domain variances that can influence the generalizability and resilience of the models that have remained trained. Additionally, because the opaque nature of deep neural networks can hinder physicians' trust and acceptance, ensuring that the models are interpretable and understandable is crucial. Incorporating explainable AI methods such as Class Activation Mapping (CAM) and Grad-CAM can enhance and clarify clinical decision-making by visualizing the key areas that impact the model's predictions.

Additionally, the use of effective data augmentation, cross-domain training techniques, federated learning approaches, and multi-modal integration methods, which combine imaging data with clinical and laboratory information, can improve the efficacy, reliability, and practical utility of these diagnostic systems. This research introduces an improved DL framework aimed at accurately detecting COVID-19 through chest X-rays and CT scans, highlighting the simulation's precision, reliability, and interpretability. The suggested approach incorporates advanced CNN architectures, comprehensive pre-processing, and visualizations that elucidate the results, to assist radiologists in providing timely, consistent, and trustworthy diagnoses, ultimately contributing to enhanced management of COVID-19 and better readiness for possible future outbreaks. The core contribution of this paper is

• Developed a novel DL architecture that integrates CNN-Lattice with ResNet50 to improve feature extraction and classification of COVID-19 using chest X-ray and CT images.
• Applied an entropy-guided U-Net example to accurately segment lung regions, enhancing disease localization by removing irrelevant background information.
• Employed Bayesian optimization to optimize key hyperparameters with learning rate, batch size, and optimizer, leading to improved model performance and quicker convergence. 
• Assessed the model's efficacy using comprehensive evaluation metrics, verifying its dependability as a decision-support tool in real-world clinical scenarios, especially in areas with inadequate access to RT-PCR. 

The next parts of the document are organized like this: related work is covered in section 2, the proposed model is explained in section 3, the experiment results are in section 4, and the conclusion is in section 5.

The proposed framework is an actual process for identifying COVID-19 CXR and CT images, which includes steps for pre-processing, segmentation, and classification. Initially, the images are resized, normalized, and enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE) to improve contrast. A U-Net model based on entropy then segments the lung regions by highlighting areas with high information content while filtering out irrelevant features. The segmented data is analyzed utilizing a CNN-Lattice ResNet50 model to enhance the capability to extract features across various scales and carry out classification. Significant hyperparameters are optimized through Bayesian techniques to develop the prototype's efficacy. Ultimately, the organization is demonstrated in Figure 1.

Figure 1. Block diagram illustrating the proposed approach

3.1 Pre-processing

In order to organise the input image for further processing, it is first pre-processed. CLAHE, normalization, and image scaling are all included in the processing. Think about the $\triangle C V I D$ dataset. $\triangle C V I D \in \left\{C V I D_1, C V I D_2, C V I D_3, \ldots, C V I D_n\right\}$, where $n$ is the number of images, and each image's vector function is provided in the database of chest CT and CXR images. The associated feature values come in a variety of formats, including numeric, binary, and nominal. It is challenging to handle various formats of CT and CXR image data for the categorization procedure. As a result, we pre-process the datasets of CT and CXR pictures. During the pre-processing phase, values are resized, normalized, and pixel values are enhanced. Following preprocessing, the output data is sent for additional processing, which is described in detail below:

3.1.1 Resizing and normalizing images

CXR and CT images were scaled to a standard determination of $256 \times 256$ pixels. This scaling procedure preserves spatial relevance while allowing compatibility with the ResNet50 framework. In mathematical terms, resizing involves mapping pixel coordinates from the original space $(i, j)$ to the designated grid $\left(i^{\prime}, j^{\prime}\right)$, using bilinear interpolation. The pixel value ($i^{\prime}, j^{\prime}$) at the new location is determined using Eq. (1) as a weighted combination of the four nearest neighboring pixels in the original image:

$C V I D(i, j)=\sum_{m=0}^1 \sum_n^1 w_{m n} * I\left(x_m, y_n\right)$    (1)

where, $x=\frac{i^{\prime}}{H^{\prime}} * H, y=\frac{j^{\prime}}{W^{\prime}} * W$ and $w_{m n}$ are the weights determined by distance. Subsequent to the resizing procedure, the pixel values of the image were normalized to a standard range to enhance the convergence through the model training phase. Initially, a min-max normalization method was utilized, scaling pixel intensities after the unique 8-bit range of [0,255] down to $[0,1]$ using Eq. (2):

$C V I D_{\text {norm }}(p)=\frac{C V I D(p)-C V I D_{\min }}{C V I D \min _{\max }}$     (2)

where, $\mathrm{CVID}_{(\mathrm{p}) \rightarrow}$ the original value for sample/pixel p, $\mathrm{CVID}_{\min } \rightarrow$ the minimum CVID value in the dataset, $\mathrm{CVID}_{\max } \rightarrow$ the maximum CVID value in the dataset, CVID_norm(p): the normalized value after scaling. To further adapt to the pre-trained ResNet50 model, channel-wise standardization as per Eq. (3) was carried out based on the statistics from the ImageNet dataset:

$I_{s t d}(p)=\frac{C V I D_{\text {norm }(p)-\mu}}{\sigma}$    (3)

Here, μ = [0.485, 0.456, 0.406] and σ = [0.229, 0.224, 0.225] denote the mean and standard deviation for the RGB channels, respectively. This pre-processing step guarantees that the input data is both scale-invariant and numerically stable, which is crucial for optimizing DL models.

3.1.2 CLAHE enhancement

To enhance the visual quality and highlight subtle disease-related features in CXR and CT images, CLAHE was applied as part of the pre-processing pipeline. Unlike global histogram equalization, CLAHE operates on insignificant regions within the image, improving local contrast while avoiding noise amplification. The image is initially segmented into M × N sections, and histogram equalization is applied separately to each section. To avoid excessive enhancement, a contrast limiting threshold T is utilized to trim the histogram and uniformly redistribute any clipped pixels. The mathematical transformation function for a pixel intensity p contained within a section is expressed as follows:

$p_{e n h}=C D F(p) *(L-1)$       (4)

where, CDF is the increasing delivery role of the trimmed histogram, and L is the number of gray levels. Bilinear interpolation is used to merge the enhanced tiles, providing seamless transitions at the edges. By greatly improving image clarity, our method aids in identifying radiological features, similar lung consolidations, and ground-glass opacities, which are necessary for exact diagnosis.

To boost the local contrast without over-amplifying the noise in CXR and CT images, a clip limit value of 2.0 and a tile grid size of 8 × 8 were used in the CLAHE pre-processing stage. The Entropy-based U-Net model uses a patch size of 256 × 256 pixels for lung segmentation, to balance spatial information and maintain computational efficiency and compatibility with the ResNet50 backbone.

3.2 Segmentation: Entropy-based U-Net

To enhance the effectiveness of disease localization and segmentation, the entropy-driven U-Net segmentation framework is developed to accurately distinguish lung regions in CXR and CT images. By integrating entropy metrics into the U-Net architecture, the model emphasizes regions with high uncertainty, which often indicate potential clinical issues.

This methodology enables the segmentation network to dynamically concentrate on areas with differing intensities and textures, resulting in better boundary definition between healthy and affected lung tissues. The segmented lung areas provide refined inputs for subsequent classification models, ultimately improving diagnostic performance for various conditions. Figure 2 illustrates the segmentation process.

Figure 2. Segmentation using Entropy-based U-Net

3.2.1 Patch-wise entropy map computation

Patch-wise entropy map computation is a critical pre-processing step in the Entropy-based U-Net that helps identify informative regions, typically areas with greater uncertainty or variation, such as infected lung tissues. Here is a detailed breakdown:

The input prepressed chest image $\operatorname{CVID}(x, y) \in R^{H * W}$ is divided into small overlapping or non-overlapping patches, usually of size $k \times k$ as mentioned in Eq. (5):

$p_{i, j}=\{\operatorname{CVID}(x, y) \mid x \in[i, i+k], y \in[j, j+k]\}$   (5)

For each patch $p_{i, j}$, compute the histogram of pixel intensity values. Normalize it using Eq. (6) to obtain the probability distribution $p_i$ for each intensity bin.

$p_i=\frac{n_i}{\sum_{i=1}^N n_i}$    (6)

where, $n_i$ is the count of pixels with intensity level i, and N is the number of bins. Apply the Shannon entropy formula to compute the entropy for each patch as in Eq. (7):

$H(i, j)=\sum_{i=1}^N p_i \log _2\left(p_i+\epsilon\right)$   (7)

In this context, ϵ is a small constant introduced to prevent log(0). A higher entropy value suggests greater texture, which is frequently associated with lung infections or abnormalities. The entropy value H(i,j) is assigned to the central pixel of each patch and is replicated throughout the image using Eq. (8) to create the entropy map EM(x,y):

$E M(x, y)=H(i, j)$ for $(x, y) \in$ center of $p_{i, j}$    (8)

Normalize the values of EM(x,y) to the range [0, 1] for consistency using Eq. (9):

$E M^{\prime}(x, y)=\frac{E M(x, y)-\min (E M)}{\max (E M)-\min (E M)}$    (9)

The normalized entropy map is subsequently active to adjust the input image within the U-Net framework, steering the model's attention towards normal or abnormal regions of the lung.

3.2.2 Weighted input to U-Net encoder

The entropy map plays a role in the Weighted Input stage of the Entropy-based U-Net, improving feature learning by directing the model’s focus towards regions of uncertainty or anomalies, typically linked to lung areas impacted by illness. The input image is adjusted based on the entropy map to emphasize significant (high-entropy) areas. The weighted input CVID ′(x,y) is divided as per Eq. (10):

$\operatorname{CVID}^{\prime}(x, y)=\operatorname{CVID}(x, y) *\left(1+\lambda * E M^{\prime}(x, y)\right.$   (10)

where, $\lambda \in R^{+}$is a hyperparameter controlling the influence of entropy, high values $E M^{\prime}(x, y)$ increase the pixel intensity in uncertain (informative) regions, and low values $E M^{\prime}(x, y)$ preserve normal regions. The weighted image CVID $^{\prime}(x, y)$ is passed into the encoder path of U-Net, and each encoder block applies as per Eq. (11).

$f^{(l)}=\sigma\left(W^{(l)} * f^{(l-1)}+b^{(l)}\right)$   (11)

where, $f^{(l)}$ is the feature map at layer l, σ is the activation function, W and b are learnable weights and biases. This process ensures that feature maps are high in relevant disease features, especially in boundary regions that are often difficult to segment accurately.

3.2.3 U-Net segmentation

The U-Net segmentation module forms the backbone of the Entropy-based segmentation pipeline. It is a symmetric encoder-decoder architecture specifically designed for pixel-wise segmentation tasks, such as extracting lung regions from CXR or CT images.

The input $\operatorname{CVID}(x, y) \in R^{H * W}$ image carries enhanced focus on informative lung regions and is passed into the U-Net model. The encoder's function is to diminish spatial resolution while preserving hierarchical traits. Each encoder block consists of.

Two convolutional layers as represented in Eq. (12):

$f^{(l)}=\operatorname{Re} L u\left(w^{(l)} * f^{(l-1)}+b^{(l)}\right)$    (12)

where, f^(l) denotes the output feature map at the l^thconvolutional layer, while f^(l−1) represents the input feature map obtained from the previous layer. The parameter W^(l) that extracts important spatial features such as edges, textures, and infection-related patterns from the input image. Max pooling for down-sampling is given in Eq. (13).

$f_{\text {pool }}^{(l)}=\operatorname{MaxPool}\left(f^{(l-1)}\right)$    (13)

In Eq. (13), $f_{\text {pool }}^{(l)}$ represents the pooled feature map after the max-pooling is performed at l^th layer, while f ^(l−1). The feature map of the previous convolutional layer is represented as MaxPool(⋅), which is a spatial downsampling operation in which the maximal value in a user-defined pooling window (usually 2 × 2) will be retained. This operation will lower the dimensionality of the feature map but keep the most important and discriminative features. Max-pooling also aids in reducing the complexities of the computation, overfitting, and provides translational invariance, which makes the network able to capture predominant abnormalities of the lungs and infection-related patterns in the CXR and CT images. To create a segmentation mask, the decoder upsamples the feature maps back to the original resolution. Each decoder block performs. Transposed convolution (upsampling) is done using Eq. (14):

$f_{u p}^{(l)}=U p \operatorname{Conv}\left(f^{(l-1)}\right)$    (14)

In Eq. (14), $f_{u p}^{(l)}$ is the upsampled feature map resulting at the l^th. The decoder layer will be the following one. f^(l−1) represents the input feature map of the previous decoder stage. The UpConv(⋅) operation denotes transposed convolution (up-convolution or deconvolution), which is the operation performed during the decode stage of the U-Net architecture to upsample the feature maps. This operation can be used to restore fine spatial detail and original image size when downsampling is performed in the encoder path. The decoder fully recovers lung boundaries and infection regions, allowing for accurate segmentation of COVID-19 affected tissues by gradually expanding on the feature maps. Concatenation with corresponding encoder features via skip connection using Eq. (15).

$f_{\text {Concat }}^{(l)}=\operatorname{Cat}\left(f_{u p}^{(l)}, f_{\text {encode }}^{(l)}\right)$   (15)

In Eq. (15), $f_{\text {Concat}}^{(l)}$ represents the concatenated feature map at the $l^{\text {th }}$ decoder layer. The term $f_{u p}^{(l)}$ denotes the corresponding downsampled feature map for this upsampled feature map $f_{\text {encode}}^{(l)}$ the feature map corresponds to the feature map in the ‘encoder path'. A Cat(⋅) operation is a feature concatenation operation performed by skip connections in U-Net architecture. At the final layer of the decoder, apply a 1 × 1 convolution to reduce feature maps to one channel (binary mask), and use a sigmoid activation to generate the segmentation probability mask as in Eq. (16):

$S(x, y)=\frac{1}{1+e^{-z(x, y)}}$   (16)

The output is a binary mask of the same size as the input, where pixel value ≈ 1 → lung region, and pixel value ≈ 0 → background. This mask is then used for further classification or visualization.

Table 1. Model parameters and hyperparameter settings

The proposed framework uses 5 randomly initialized starting points and optimizes the hyperparameters for 100 iterations of Bayesian Optimization to ensure good exploration of the hyperparameter search space. A Gaussian Process (GP) based surrogate model and the Expected Improvement (EI) acquisition function were used to achieve a tradeoff between exploration and exploitation during optimization. The optimization process looked for the optimal setting of the hyperparameters as listed in Table 1.

3.3 Model architecture

ResNet50 architecture integrated with CNNs-Lattice, combining DL's residual feature extraction with lattice-based transformations for improved classification, as exposed in Figure 3.

Figure 3. Proposed ResNet50 architecture integrated with CNNs-Lattice

3.3.1 ResNet50 encoder

To improve diagnostic accuracy in medical imaging, segmented input images typically generated using entropy-based U-Net are used as input to the ResNet50 network. The segmented image, denoted as $S(x, y)$, contains only the lung regions with the background removed, focusing the model on diagnostically relevant features. The ResNet50 model, made up of convolutional layers and residual blocks, takes the image as its input. To obtain low-level features, the initial convolutional layer utilizes a 7 × 7 kernel with a stride of 2 as per Eq. (17):

$F_1=\operatorname{Re} L U\left(B N\left(W_1 * S+b_1\right)\right)$    (17)

This is followed by a max-pooling operation as given in Eq. (18).

$\left.F_2=\operatorname{MaxPoo} 1\left(F_1\right)\right)$    (18)

The resulting features are then passed through a series of residual blocks from Conv2_x to Conv5_x. Each residual block computes as per Eq. (19):

$\begin{aligned} F^{(l)}=\sigma\left(W_1^{(l)} *( \right. & \left.\left.\sigma\left(W_1^{(l)} * F^{(l-1)}+b_1^{(l)}\right)\right)+b_2^{(l)}\right) \\ & +F^{(l-1)}\end{aligned}$     (19)

where, σ is the ReLU activation, W₁and W₂are weight tensors of convolution layers, $F^{(l-1)}$ is the input to the block, and is also used in the skip connection. These residual connections help maintain gradient flow and preserve spatial detail across many layers. After passing through the deepest block (Conv5_x), the network outputs a high-dimensional feature map in Eq. (20):

$F_{\text {ResNet }} \in R$    (20)

This $F_{\text {ResNet}}$ feature map encodes rich hierarchical information about the segmented lung region, including shapes, textures, and structural abnormalities in Figure 4. It serves as the deep feature representation and can be further processed using global average pooling as per Eq. (21):

$f=G A P\left(F_{\text {ResNet }}\right) \in R$     (21)

where, GAP(⋅): Global Average Pooling, which averages each feature channel across spatial dimensions. f the resulting feature vector after pooling. Advanced modules, CNNs-Lattice, were directly applied for non-linear transformations and classifications. With segmented input, ResNet50 fully concentrates on the pertinent lung anatomy, enhancing its generalization and reliability in identifying conditions like COVID-19 or pneumonia.

(a) Chest X-rays (CXR) images

(b) Computed tomography (CT) images

Figure 4. Feature map

3.3.2 Lattice layer transformation

The CNNs-Lattice layer applies a trainable piecewise-linear function to the input features. Each output dimension is a function of multiple input dimensions using a lattice function as mentioned in Eq. (22):

$y_k=L_k\left(f ; \theta_k\right) \in R$      (22)

where, $L_k$ is the lattice function for output node $k, \theta_k$ are the learnable lattice parameters, the transformation is nonlinear, monotonic, and interpretable. The CNNs-Lattice layer models structured decision boundaries by learning interpolated outputs over a multi-dimensional grid (the lattice), which is effective for high-variance medical features.

3.3.3 Final prediction layer

To decrease the dimensionality of the feature map $f$, we apply GAP, which summarizes each feature channel into a single value using Eq. (23):

$f_i=\frac{1}{H \cdot W} \sum_{x=1}^H \sum_{y=1}^W y_k(x, y, i)$    (23)

The final prediction is obtained by spreading over a softmax function depending on the task, as in Eq. (24).

$\hat{y}=\frac{e^{z_i}}{\sum_{j=1}^C e^{z_i}} \quad$ for $i=1,2,3, \ldots \ldots, C$     (24)

The final output is a vector of probabilities for each class for a 3-class classification COVID/Pneumonia/Normal.

3.4 Hyperparameter tuning

To improve the classification effectiveness of the ResNet50 model integrated with the CNNs-Lattice approach, Bayesian Optimization was employed for tuning hyperparameters. This technique effectively navigates the hyperparameter landscape by building a surrogate probabilistic model, often a Gaussian Process (GP), to approximate the target function, such as validation accuracy. The main hyperparameters that were optimized include the learning rate (η), batch size (B), and type of optimizer (O). The search space AD is defined as in Eq. (25):

$A D=\left\{\begin{array}{c}\eta \approx \log \left(1 X 10^{-5}, 1 X 10^{-2}\right. \\ B \in\{16,32,64,128\} \\ O \in\{\text { Adam }\}\end{array}\right.$     (25)

Bayesian Optimization recursively chooses the next set of hyperparameters θ= (η, B, O) ∈ AD by maximizing an acquisition function a(θ), which assesses both anticipated performance and uncertainty derived from the surrogate model using Eq. (26):

$\theta^*=\underset{\theta \in A D}{\operatorname{argmaxa}}(\theta)$    (26)

The performance objective is to maximize validation accuracy, or equivalently, minimize validation loss, denoted as in Eq. (27):

$\theta^*=\underset{\theta \in A D}{\operatorname{argmax}} A C C_{\text {val }}(\theta)$    (27)

where, θ: a candidate set of model parameters or hyperparameters. AD: the search space/domain of allowable configurations. $A C C_{v a l}(\theta)$ validation accuracy achieved using configuration θ. argmax: “the argument that maximizes,” meaning the value of θ producing the largest accuracy. $\theta^*$ the best/optimal configuration found. By continuously updating the surrogate model based on observed outcomes, Bayesian Optimization efficiently converges to an optimal or near-optimal hyperparameter configuration, achieving better generalization performance with significantly fewer model evaluations compared to exhaustive search techniques.

The proposed CNN-Lattice integrated ResNet50 framework shows good classification and segmentation accuracy, whereas it does not extensively discuss the robustness under small-sample conditions and noisy medical images. The quality of the images acquired by CT and CXR may differ in the practical clinical setting, as image acquisition settings, imaging devices, motion artifacts, and noise levels may differ and impact the model performance and its ability to generalize. While preprocessing methods like CLAHE enhancement and entropy segmentation using U-Net have been employed to enhance the image quality and minimize the presence of noise and irrelevant artifacts, further testing on more diverse and challenging clinical datasets is still required to validate the robustness and reliability of the proposed method in practical settings.

This research introduces a dependable and accurate DL designed for the automatic detection of COVID-19, pneumonia, and healthy subjects using chest CXR and CT scans. The proposed architecture integrates ResNet50 with enhancements derived from CNNs-Lattice, resulting in a significant boost in feature representation and classification performance. A significant advantage of this framework lies in its organized pre-processing pipeline, where image quality is enhanced through CLAHE, and irrelevant areas are eliminated via entropy-based U-Net segmentation, allowing focus on the lung region. The research employs two balanced datasets, COVIDx CXR-4 and COVIDx CT, ensuring equitable training and assessment. Comprehensive testing and comparative evaluations against leading models such as DenseNet121, EfficientNetB0, and VGG16 indicate that the proposed methodology consistently outperforms the others regarding accuracy, precision, recall, F1-score, and AUC. Significantly, it attains enhanced segmentation outcomes, as demonstrated by higher DSC and IoU metrics for both CT and CXR images. Furthermore, the use of Bayesian Optimization for hyperparameter adjustment, including learning rate, batch size, and optimizer, has led to faster convergence and enhanced generalization. An examination using a confusion matrix also reveals that the proposed model has the lowest misclassification rate, highlighting its dependability in clinical applications. In summary, the proposed CNNs-Lattice enhanced ResNet50 model offers an efficient, interpretable, and scalable solution for the prompt and precise identification of COVID-19 and related pulmonary conditions, making it particularly well-suited for real-time application in medical diagnostic systems, especially in resource-limited environments where RT-PCR testing may be restricted. The proposed framework is a very high performing one, but some aspects of this are still limited. The model was tested primarily on publicly available datasets and could not be optimized for smaller datasets, images captured by different imaging devices, and real clinical settings. Future work will focus on multi-center clinical validation and improving model generalization for real-time deployment.