Automatic Diagnosis of COVID-19 Pulmonary Infection Based on Multi-Scale Feature Enhancement

Figure 1. COVID-19 pulmonary infection region extraction model network structure

2.1.1 Backbone network

This paper makes targeted improvements to the VGG-16 backbone network to ensure the model adapts to the fundamental characteristics of COVID-19 pulmonary infection images. Figure 2 shows the improved backbone network architecture. COVID-19 CT images typically contain numerous diffuse, small-sized ground-glass opacities and consolidation regions. These infection features are widely distributed and vary in shape within the image. The improved backbone network fixes the input size to 512×512 pixel RGB images, which balances retaining sufficient detail information with computational efficiency. By using a 3×3 small-sized convolution kernel with a stride of 1 and padding of 1, the network fully retains boundary information of the infection regions during the convolution process, avoiding the loss of fine infection features caused by size reduction. Additionally, the depth structure of 13 convolutional layers ensures that the network can progressively build a comprehensive understanding of the infection regions, from low-level texture features to high-level semantic features.

Figure 2. Backbone network architecture

In terms of feature extraction depth and efficiency optimization, this paper makes significant adjustments to the channel configuration of VGG-16. The number of output channels for the first downsampling layer is changed to 128, and three downsampling operations with 512 channels are implemented. The initial layers use relatively fewer channels to prioritize capturing basic visual features, such as edge and texture features, which are crucial for distinguishing infection regions from normal tissue. As the network depth increases, the number of channels gradually expands to 512, enabling the network to process multiple feature maps simultaneously and more comprehensively describe the complex infection patterns.

The abandonment of the fully connected portion of VGG-16 is another key design decision, which directly serves the unique needs of the image segmentation task. Traditional VGG-16 contains three fully connected layers, which disrupt the spatial structure information of the feature maps and introduce a large number of parameters into the model, increasing computational complexity and potentially causing gradient vanishing issues. For the dense prediction task of COVID-19 infection region segmentation, preserving the spatial information of feature maps is crucial, as pixel-level classification results are required. The improved backbone network exclusively uses convolution and pooling operations, gradually reducing the feature map size with five max pooling layers while increasing the level of feature abstraction. Max pooling operations are particularly suitable for medical image processing because they preserve the strong response of features, highlighting the contrast differences between infection regions and normal tissue, and providing rich spatial context for subsequent multi-scale feature enhancement modules. Specifically, assuming the width and height of the output feature map are represented by q_p and g_p, the width and height of the input feature map by q_u and g_u, the pooling kernel's width and height by q_o and g_o, and the stride by ST, the computation formula for max pooling is as follows:

${{q}_{p}}=\frac{\left( {{q}_{u}}-{{q}_{o}} \right)}{ST}+1$                (1)

${{g}_{p}}=\frac{\left( {{g}_{u}}-{{g}_{o}} \right)}{ST}+1$               (2)

In terms of nonlinear representation and gradient optimization, the improved VGG-16 backbone network uses the ReLU activation function to address the gradient vanishing problem in deep networks. The boundaries of COVID-19 infection regions are often blurred, with a gradual transition between infection and normal lung tissue, requiring the network to have strong nonlinear modeling capability. The ReLU activation function, with its one-sided suppression property, enhances the network’s nonlinear expression ability while maintaining training efficiency, enabling the model to learn the complex decision boundaries between infection regions and complex backgrounds. Although the final fully connected layer is removed, the network still retains sufficient nonlinear transformation ability at the end, and through progressive ReLU activations, the model can gradually refine high-level features representing infection regions.

2.1.2 Multi-scale feature enhancement

To address the key challenges of multi-scale characteristics, blurred boundaries, and adhesion to normal tissue in COVID-19 pulmonary infection regions in CT images, this paper designs a multi-scale feature enhancement module aimed at achieving precise feature extraction and differentiation through architectural innovation. The specific architecture is shown in Figure 3. This module consists of a foreground feature enhancement process and a context information enhancement module working in tandem. Due to the significant variation in infection lesion sizes, ranging from tiny ground-glass nodules to large consolidation areas, and the often infiltrative blurring of boundaries, the module builds a context information enhancement component by integrating dilated convolutions with different dilation rates to systematically expand the model’s receptive field. The main reason for adopting this multi-scale receptive field strategy is that it effectively distinguishes infection regions with similar appearances from background structures such as blood vessels and interstitium, reducing misjudgments. At the same time, the foreground feature enhancement process dynamically adjusts the weights between foreground and background features using learnable parameters, actively strengthening the response to infection features and suppressing noise interference caused by the complex lung parenchyma background.

Figure 3. Multi-scale feature enhancement module structure

(1) Foreground Enhancement

The foreground feature enhancement module is based on the core challenges of low contrast, blurred boundaries, and complex background interference between infection regions and normal tissue in COVID-19 pulmonary CT images. This module uses a Sigmoid-based soft attention mechanism to achieve precise separation and weight adjustment between foreground and background features. Specifically, the module first applies the Sigmoid function to the high-level prediction results, generating foreground regions with values close to 1 and background regions with values close to 0. The probabilistic treatment effectively handles the gradual transition between the ground-glass-like blurred shadows of infection regions and normal lung tissue, which is common in COVID-19 images. By multiplying the background probability map from the Sigmoid output with the upsampled high-level features, the module can clearly identify background noise features, thus providing the foundation for subsequent foreground enhancement. Specifically, assuming the upsampling process is denoted by US( ), the context information enhancement process by XXZQ( ), the upsampled high-level features by $D_g^{U P}$, the background noise features by $D_m^y$, the input low-level features by D_m, the high-level features D_g, the high-level prediction results by O_g, and the output fused features by D_RE, the more refined segmentation result by O_m, the calculations for each layer’s features are as follows:

$D_{g}^{UP}=US\left( Con{{v}_{7\times 7}}\left( {{D}_{g}} \right) \right)$            (3)

$O_{g}^{UP}=US\left( Sigmoid\left( {{O}_{g}} \right) \right)$            (4)

$D_{m}^{y}=XXZQ\left( \left( 1-O_{g}^{UP} \right)\times {{D}_{m}} \right)$            (5)

${{D}_{RE}}=RELU\left( BN\left( D_{g}^{UP}-\beta \times D_{m}^{y} \right) \right)$            (6)

$O_m=\operatorname{Conv}_{3 \times 3}\left(D_{R E}\right)$         (7)

Based on the preliminary separation of foreground and background features, the module further explores and optimizes these features through a dual-branch context information enhancement architecture, which separately processes foreground and background features. These multi-scale characteristics include both scattered small ground-glass opacities and large, merged consolidation areas. Two independent context information enhancement modules process the foreground and background features, respectively, using dilated convolutions with different dilation rates to construct multi-scale receptive fields and capture infection features of different sizes. Figure 4 shows the context information enhancement module structure. For the foreground branch, the module focuses on extracting internal texture features and boundary morphology of the infection region; for the background branch, it focuses on identifying normal tissue structures that are easily misjudged as infection. The dual-path processing approach ensures that the model can learn discriminative features of the infection region from both the forward and reverse dimensions. In particular, it effectively improves the model's recognition ability for small infection foci and blurred boundaries, which are caused by partial volume effects in COVID-19 pulmonary images.

To achieve dynamic weight optimization for foreground and background features, the module introduces learnable parameters α and β, and continuously improves segmentation results through an incremental multi-level refinement architecture. This parameterized fusion mechanism allows the model to automatically adjust the contribution of foreground and background features based on different infection patterns. For example, in diffuse ground-glass opacity regions, the foreground features are given higher weight, while in normal lung parenchyma regions, the importance of background features is enhanced. To accommodate the complex manifestations of COVID-19 infection, where infection regions present different feature patterns at different scales, this paper uses a cascading structure with four multi-scale feature enhancement modules, forming a feature optimization process from coarse to fine: high-level modules provide global infection distribution semantic information, while low-level modules gradually incorporate more spatial details to refine boundary localization. Finally, through the collaboration of batch normalization and the ReLU activation function, the module enhances the nonlinear expression ability while maintaining feature distribution stability, enabling the model to progressively suppress background noise, strengthen infection features, and ultimately output infection region segmentation results with clear boundaries and accurate spatial positioning.

(2) Context Information Enhancement

COVID-19 infection regions exhibit significant scale diversity in CT images: In the early stages, they often appear as scattered small ground-glass opacities, while in the progressive stage, large consolidation areas may merge. Standard convolutional networks inevitably lose spatial details as they downsample to expand the receptive field, leading to decreased detection ability for small lesions. To address the contradiction between the multi-scale distribution characteristics of COVID-19 pulmonary infection regions and the need to retain details in CT images, this paper designs a context information enhancement module. This module innovatively adopts a multi-branch dilated convolution architecture, setting four branches with dilated convolutions at rates of 2, 4, 8, and 2, combined with 1×1, 3×3, 5×5, and 7×7 convolution kernels of different sizes. For the characteristic patterns of peripheral distribution and multi-lobe involvement in COVID-19, the module uses smaller dilation rates with 3×3 convolution kernels to finely capture the subtle density changes of ground-glass opacities, while larger dilation rates with 7×7 convolution kernels understand the spatial distribution relationships between multiple infection foci.

The complex manifestations of COVID-19, such as the "stepping stone sign," include increased ground-glass density and thickened interlobular septa. These require different scales of context information for accurate interpretation. To address the spatial distribution characteristics of COVID-19 infection regions, where infection regions are not isolated but show spatial correlation within the lung, the four branches of the module adopt an increasing dilation rate design. Small ground-glass opacities may merge as the disease progresses, forming more dangerous consolidation areas. By setting increasing dilation rates, the model constructs a hierarchical receptive field system: lower dilation rate convolution layers capture internal features and clear boundaries of individual infection foci, while higher dilation rate convolution layers integrate broader region information, identifying potential relationships between multiple foci.

In terms of feature fusion, the context information enhancement module integrates multi-scale features from the four branches using a 1×1 convolution to address the heterogeneity and complexity of COVID-19 pulmonary infection regions. The features extracted from different branches represent infection region information at different scales: the local branch retains edge and texture details crucial for detecting small lesions, while the large receptive field branches provide the contextual information necessary for recognizing infection region distribution patterns. The final fusion process, through batch normalization and ReLU activation functions, ensures the coordinated distribution and nonlinear expression ability of features from different scales. Specifically, assuming the input feature D, the four branches’ calculation processes are as follows:

${{D}_{BR1}}=Conv_{3\times 3}^{f=1}\left( Con{{v}_{1\times 1}}\left( Con{{v}_{1\times 1}}\left( D \right) \right) \right)$          (8)

${{D}_{BR2}}=Conv_{3\times 3}^{f=2}\left( \left( Con{{v}_{1\times 1}}\left( D \right)+{{D}_{BR1}} \right) \right)$            (9)

${{D}_{BR3}}=Conv_{3\times 3}^{f=4}\left( Con{{v}_{5\times 5}}\left( Con{{v}_{1\times 1}}\left( D \right)+{{D}_{BR2}} \right) \right)$          (10)

${{D}_{BR4}}=Conv_{3\times 3}^{f=8}\left( Con{{v}_{7\times 7}}\left( Con{{v}_{1\times 1}}\left( D \right)+{{D}_{BR3}} \right) \right)$                 (11)

The module’s output is:

${{D}_{XXZQ}}=Con{{v}_{1\times 1}}\left( Concat\left( {{D}_{BR1}},{{D}_{BR2}},{{D}_{BR3}},{{D}_{BR4}} \right) \right)$            (12)

2.1.3 Upsampling

COVID-19 infection regions in CT images often manifest as a gradual transition between ground-glass opacities and normal lung tissue, with blurred boundaries and a lack of clear edge contours. To address this, this paper adopts bilinear interpolation for upsampling in the decoder section. Bilinear interpolation computes a weighted average of the surrounding four pixels to generate smooth transition pixel values, which is particularly suited to handling the blurriness of infection region boundaries. Unlike transposed convolution, which may produce "checkerboard artifacts," bilinear interpolation preserves the natural smooth transition of boundaries, avoiding the introduction of unnatural hard edges at the infection region boundaries. Additionally, the computational efficiency of this method is an important consideration for its selection in this study. COVID-19 lung CT images typically have high spatial resolution, and large volumes of image data need to be processed in practical diagnosis. As a lightweight interpolation method, bilinear interpolation only requires simple arithmetic operations to perform the upsampling, which significantly reduces the model's computational complexity and speeds up inference. Particularly in cases where the model requires four consecutive upsampling operations, the efficiency advantage of bilinear interpolation becomes more apparent, ensuring the practicality of the entire segmentation process.

To build an efficient feature recovery path, this paper cleverly combines bilinear interpolation with a skip connection mechanism. Through four bilinear interpolation operations, each time doubling the feature map resolution, and by integrating features from the corresponding layers of the encoder, the model can progressively recover the spatial details lost during deep feature extraction. For example, when identifying scattered small ground-glass opacities, high-resolution features from the shallow network can provide the necessary texture details, while deep features upsampled using bilinear interpolation offer semantic context. The combination of both ensures accurate detection and localization of small infection foci. Specifically, given known W₁₁(a₁,b₁), W₁₂(a₁,b₂), W₂₁(a₂,b₁), W₂₂(a₂,b₂), the interpolated pixel value at coordinate O, denoted as E₁(a₁, b₁), is first calculated based on W₁₁ and W₂₁. Then, the pixel value E₂(a₂, b₂) is calculated using W₁₂ and W₂₂, and finally, the pixel interpolation at point O is obtained using E₁ and E₂ as follows:

$\phi \left( {{E}_{1}} \right)=\left( \frac{{{a}_{1}}-a}{{{a}_{2}}-{{a}_{1}}} \right)\phi \left( {{W}_{11}} \right)+\left( \frac{a-{{a}_{1}}}{{{a}_{2}}-{{a}_{1}}} \right)\phi \left( {{W}_{21}} \right)$            (13)

$\phi \left( {{E}_{2}} \right)=\left( \frac{{{a}_{2}}-a}{{{a}_{2}}-{{a}_{1}}} \right)\phi \left( {{W}_{12}} \right)+\left( \frac{a-{{a}_{1}}}{{{a}_{2}}-{{a}_{1}}} \right)\phi \left( {{W}_{22}} \right)$             (14)

Performing linear interpolation based on φ(E₁) and φ(E₂) along the vertical axis:

$\phi \left( O \right)=\left( \frac{{{b}_{2}}-b}{{{b}_{2}}-{{b}_{1}}} \right)\phi \left( {{E}_{1}} \right)+\left( \frac{{{b}_{1}}-b}{{{b}_{2}}-{{b}_{1}}} \right)\phi \left( {{E}_{2}} \right)$                  (15)

Combining the above three formulas, there is:

$\begin{aligned} & \phi(O)=\left(\frac{b_2-b}{b_2-b_1}\right)\left(\frac{a_2-a}{a_2-a_1}\right) \phi\left(W_{11}\right) +\left(\frac{b_2-b}{b_2-b_1}\right)\left(\frac{a-a_1}{a_2-a_1}\right) \phi\left(W_{21}\right) \\ & +\left(\frac{b-b_1}{b_2-b_1}\right)\left(\frac{a_2-a}{a_2-a_1}\right) \phi\left(W_{12}\right) +\left(\frac{b-b_1}{b_2-b_1}\right)\left(\frac{a-a_1}{a_2-a_1}\right) \phi\left(W_{22}\right)\end{aligned}$             (16)

By considering the weights of the four points affecting point O, denoted by μ₁₁, μ₂₁, μ₁₂, and μ₂₂, the equation simplifies to:

$\begin{aligned} & \phi(O)=\mu_{11} \phi\left(W_{11}\right)+\mu_{21} \phi\left(W_{21}\right) +\mu_{12} \phi\left(W_{12}\right)+\mu_{22} \phi\left(W_{22}\right)\end{aligned}$                      (17)

Finally, through cascaded bilinear upsampling and feature fusion, the model can output a high-quality segmentation result that matches the input image size. In the final layer of the network, after four upsampling and feature fusion operations, the feature map undergoes a 1×1 convolution to reduce the channel number to the number of target classes, followed by the application of the Softmax function to generate the probability of each pixel belonging to a class. Figure 5 shows the model execution flow.

Figure 5. Model execution flow

2.1.4 Loss function

COVID-19 infection regions in lung CT images typically appear as scattered, discontinuous lesions, with infection pixels occupying a small proportion of the entire image, leading to significant class imbalance. Furthermore, as COVID-19 pulmonary infections often manifest as gradually increasing ground-glass densities, the boundary between the infection region and normal tissue is usually unclear, which introduces a certain level of uncertainty in the annotated data itself. Therefore, this paper selects cross-entropy loss as the loss function for the segmentation task. This loss function, through its logarithmic calculation, sensitively penalizes the model’s predicted probabilities: when the model misclassifies the actual infection region as background, the loss function generates large gradient signals, forcing the model to quickly adjust its parameters to correct such severe errors. Additionally, by independently evaluating the prediction probability of each pixel, this loss function adapts better to this situation of boundary fuzziness, without applying excessive penalty to boundary pixels. Specifically, let L denote the total number of samples, V the number of classes, the probability of the u-th sample belonging to the k-th class be o_uk, and the probability predicted by the model for the u-th sample belonging to the k-th class be w_uk, the loss function formula is as follows:

$L O S S=-\frac{1}{L}\left[\sum_{u=1}^L \sum_{k=1}^V o_{u k} \log \left(w_{u k}\right)\right]$             (18)

2.2 Automatic diagnosis of COVID-19 pulmonary infection

Based on the precise infection region segmentation results extracted by the multi-scale feature enhancement model, the core strategy for the automatic diagnosis of COVID-19 in this paper is to convert pixel-level segmentation results into clinically valuable quantitative indicators and classification decisions.

The model first calculates key quantitative diagnostic indicators based on the high-precision segmentation results. These include the total volume ratio of the infected region in the entire lung, the distribution of infection across different lung lobes, and the volume and ratio of ground-glass opacities and consolidation areas. These indicators are not isolated; they are integrated into a comprehensive evaluation framework. For example, the total infection volume ratio combined with the consolidation region ratio can serve as an important basis for grading disease severity; the distribution characteristics of the infection region across the upper, middle, and lower lung regions help determine the typical patterns of the disease. This strategy directly serves clinical decision-making, allowing doctors to quickly obtain objective and repeatable quantitative data without manually delineating the infection regions, significantly improving diagnostic efficiency and consistency.

Furthermore, the automatic diagnostic system combines these quantitative features with multi-scale deep features such as texture and morphology of the infection regions and inputs them into a lightweight classifier to achieve the final diagnostic classification. The model not only extracts the scope of the infection region but also includes its intrinsic radiological features. For instance, subtle texture changes captured by the multi-scale feature enhancement module can be used to distinguish active inflammation from later-stage fibrosis; morphological features such as the clarity of infection region boundaries and irregularity of shapes serve as key indicators to differentiate COVID-19 from other pulmonary infections. This strategy of integrating "segmentation result quantification" with "deep feature analysis" elevates the automatic diagnosis from simple region detection to a comprehensive judgment of disease type and progression stage, ultimately outputting clinically meaningful diagnostic conclusions such as "suspected," "confirmed mild," "confirmed moderate," or "confirmed severe," thereby fully achieving the closed loop from image analysis to auxiliary diagnosis.