Abstract
Abstract Worldwide, the coronavirus disease 2019 (COVID-19) led to a critical public health situation. While automatic segmentation of lung infections of COVID-19 in Computed Tomography (CT) scan images leads to enhance treatment strategies and patient care, it faces many challenges, including similarity to adjacent tissues, unclear boundaries, and scattered infections. This paper investigates hybrid deep features to enhance segmentation accuracy and reduce uncertainty. The study explores and synthesizes raw CT images with Local Directional Number Pattern (LDNP) encoded images within a novel deep learning framework. Hybrid deep features are extracted from overlapping image patches in a refined Convolutional Neural Network (CNN) architecture. Using LDNP to transform CT lung images into a new representation with gradient-based features along to original image, improves segmentation accuracy. Also, in the proposed method, each pixel of the image is classified into normal and infected tissues, allowing for the extraction of more complex infection patterns. The patch training approach results in a lightweight single-stream CNN, valuable in resource-constrained environments, and expands training data, reducing overfitting and making it efficient even with small datasets. Experiments conducted on a standard COVID-19 CT segmentation dataset show the effectiveness of the proposed scheme respect to the second-best state-of-the-art method 5.5%, 0.6%, 3.8%, and 47.6%, for Dice, Sensitivity, Specificity, and MAE evaluation metrics, respectively.