LDNP-enhanced CNN improves COVID-19 lung CT segmentation in small dataset

A lightweight convolutional neural network that combined directional texture information with original chest CT images outperformed six comparison methods for COVID-19 lung infection segmentation in a computational study of 100 annotated axial grayscale images. The approach added a Local Directional Number Pattern (LDNP) representation to each original image before dividing it into overlapping patches for classification as background or infected tissue. On the studied dataset, the complete method improved on the second-best reported result by 5.5% in Dice score, 0.6% in sensitivity and 3.8% in specificity, while reducing mean absolute error by 47.6%.

The study measured image-segmentation performance only; it did not evaluate clinical diagnosis, treatment decisions or patient outcomes. The research, published on 5 September 2026 in Discover Artificial Intelligence, addressed a central difficulty in COVID-19 CT image analysis: infected lung regions may resemble neighboring tissues, have unclear boundaries or occur in scattered patterns. The authors designed their method to retain the anatomical and intensity information in the original image while making local directional structures more visible to the neural network.

They used LDNP, a compact texture descriptor originally introduced for face and expression recognition, as a preprocessing operation rather than as an independent feature-extraction branch. The experimental scope was limited to 100 axial lung CT grayscale images, 512 by 512 pixels each, from an open-access dataset associated with the Italian Society of Medical and Interventional Radiology. LDNP is constructed by convolving an image with eight Kirsch compass masks, which capture oriented edge responses at 45-degree intervals.

At each pixel, the mask producing the largest response supplies three bits of a six-bit code, while the mask producing the smallest response supplies the remaining three bits. The resulting LDNP image therefore encodes the strongest positive and negative directional responses around each location. In the COVID-19 application, this local gradient information was intended to emphasize texture transitions and boundary-related patterns associated with heterogeneous infection appearances.

Because the descriptor represents direction as well as local contrast, it provides structural distinctions that may be less explicit in the original grayscale image. Rather than sending the original CT image and LDNP representation through separate encoders or combining their learned features in a multiscale, multibranch or multistage architecture, the researchers used input-level hybridization. A pixel-wise sum of the two images formed one composite input to a single-stream convolutional neural network.

This design differs from the more complex feature-fusion systems reviewed in the paper, including approaches based on dual paths, attention modules, transformers, capsule networks, dilated fusion and separate decoder structures. The simpler strategy was intended to capture complementary information without adding the computational burden of later-stage feature fusion. It also preserved the original CT data rather than replacing it with a handcrafted representation, although the composite image changed the numerical values presented to the network.

The network itself was deliberately compact. It accepted grayscale patches measuring 15 by 15 pixels and contained four convolutional layers with 3-by-3 kernels and respectively 8, 16, 32 and 64 filters. Batch normalization and rectified-linear-unit activation followed each convolution, while max-pooling reduced the spatial dimensions of the feature maps.

Four fully connected layers with 100, 100, 20 and 2 neurons converted the extracted features into a final softmax classification between background and COVID-19 infection. The architecture was selected through a bottom-up sizing process in which network complexity was increased until performance stabilized, balancing the risk that a model might be too small to represent complex patterns against the risk of overfitting in a small medical-imaging dataset. Local patches also served as a data-expansion mechanism.

Each 512-by-512 image was divided into 15-by-15 windows with a stride of 5 pixels, and 20 pixels were removed from the horizontal and vertical borders because useful lung information was considered unlikely near those edges. In the example given by the researchers, this configuration produced 9,025 patches from each image, compared with 961 without the overlapping-patch scheme. The extracts were drawn from both infected and healthy lung regions so that the classifier could learn both classes.

Training augmentation expanded each patch into eight versions through rotations of 90, 180 and 270 degrees, combined with the original orientation, together with added salt-and-pepper noise. At inference, the model labeled each patch and assigned its prediction to the corresponding spatial window in a full segmentation map. Overlapping predictions were aggregated to reconstruct the image, with a smaller stride increasing overlap and segmentation resolution at the cost of additional computation.

A stride of 5 was selected empirically to balance boundary continuity against processing efficiency. The final map underwent morphological post-processing: erosion with a disk-shaped structural element of radius 1 refined boundaries, objects smaller than 40 pixels were removed to suppress noise, and dilation restored the shape of retained regions. The authors reported that this sequence produced output visually closer to the reference masks than the unprocessed prediction.

Evaluation used 45 images for training, five for validation and 50 for testing. The infection masks had been annotated by a radiology specialist. The study assessed Dice similarity coefficient, sensitivity, specificity and mean absolute error at the pixel level.

Training ran for 20 epochs, and the selected 15-by-15 patch size performed best among the tested alternatives of 13 by 13 and 17 by 17 pixels. The full model was compared with U-Net, attention-gated U-Net, U-Net++, another attention-gated architecture, H-DenseUNet and Inf-Net. Based on the reported table, it achieved the strongest results across all four metrics, although the article did not provide confidence intervals or statistical significance tests for the differences.

An ablation analysis examined whether the reported gains could be attributed to individual design choices. Replacing LDNP with a basic Local Binary Pattern was less effective than the full configuration, while removing LDNP altogether produced reported differences of 8.7% in Dice, 12% in sensitivity, 1.3% in specificity and 12% in mean absolute error relative to the complete method. Removing the last convolutional layer, omitting the first fully connected layer or adding another convolutional layer with 128 filters also altered performance.

These comparisons support the contribution of the texture representation and cleanup step within this experimental setup, but they do not establish how the method would behave across different hospitals, scanners, patient groups or disease stages. The findings position LDNP-assisted, patch-based learning as a potentially efficient approach for pixel-level COVID-19 infection segmentation, particularly where annotated data and computing resources are limited. The experiments ran in MATLAB 2023 on a computer with a 9th-generation Intel Core i5 processor, 16 GB of memory and an Nvidia GeForce GTX 1050 GPU, and the authors describe the resulting single-stream network as relatively lightweight and suitable in principle for resource-constrained settings.

That deployment potential remains a design goal rather than a demonstrated clinical workflow. The study used one relatively small, open-access dataset, did not report external validation across independent cohorts, and tested segmentation accuracy rather than whether clinicians make better decisions with the model. The authors identify evaluation on additional datasets, different imaging conditions and other handcrafted features as important directions for future work.

Subject of Research: Computational segmentation of COVID-19 lung infection regions in chest CT images using an LDNP-enhanced patch-based CNN

Article Title: Hybrid deep features for COVID-19 lung CT image segmentation

Article References: Parandakh, P., & Fotouhi, A. M. (2026). Hybrid deep features for COVID-19 lung CT image segmentation. Discover Artificial Intelligence, 6(1), Article 1062. https://doi.org/10.1007/s44163-026-02145-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02145-8

Keywords: COVID-19, lung CT, image segmentation, deep learning, convolutional neural network, LDNP, directional texture, medical artificial intelligence, pixel-level analysis, Hybrid, deep, features