Analysis of Lung Scan Imaging Using Combination of Image Processing Algorithms and Deep Multi-Task Learning Structure for Covid-۱۹ Disease

Background and aims: COVID-۱۹ is a global health challenge with over ۴۵۸ million confirmedcases and ۶.۱ million deaths worldwide as of March ۲۰۲۳. Accurate diagnosis of lung infectionsthrough computed tomography (CT) scans is crucial for effective treatment. However, manual diagnosisis time-consuming and subjective. Therefore, this study proposes an automated multi-taskdeep learning model for the segmentation and classification of CT scans to detect infected areas.To increase the efficiency of the model, CT scan images were enhanced using image processingalgorithms before entering the network.Method: In this study, an encoder-decoder model based on the U-net architecture was used. Inthe pre-processing phase, a median filter and mathematical morphology operation were appliedto the input images to improve their quality. The encoder was responsible for feature extraction,and the number of filters increased from ۶۴ to ۱۰۲۴ in the encoder. Skip connections were usedfollowing each convolutional block to preserve information. Then, the decoder level began witha sampling layer, followed by a convolution to decrease the number of features by a factor of ۲in the segmentation task to detect areas affected by COVID-۱۹ and other infections, as well ashealthy regions. In the classification task, a multilayer perceptron was used, and ۴ neurons wereconsidered for each class of the task in the last dense: COVID-۱۹, normal, other infections, andcombined infections. Since both classification and segmentation tasks used the same dataset, itwas necessary to use a dataset that had both masks and labels. The proposed model was trainedusing two-dimensional CT scans with allocated masks and labels obtained from the Italian Societyof Medical and Interventional Radiology, and by applying data augmentation techniques to thedataset, the number of slices increased to ۱۳۱۱ images. Images were segmented by a radiologist.However, in the segmentation task of the proposed model, infected areas infections and healthyregions were considered as segmentation labels.Results: The model achieved an accuracy of ۹۷.۱۶%, MSE of ۰.۰۲, and mean dice of ۸۸.۸۹±۰.۰۲in the segmentation task. In the classification task, the model used a combination of median filterand morphology operation to achieve an accuracy of ۹۷.۷۵% and AUC of ۰.۹۷. The median filteralone achieved an accuracy of ۰.۹۶ and mean dice of ۸۸.۷۸ ± ۰.۰۶ in the segmentation task and anaccuracy of ۰.۹۷ and AUC of ۰.۹۷ in the classification task. Similarly, the morphology operationachieved an accuracy of ۰.۹۶ and mean dice of ۸۸.۷۹ ± ۰.۰۴ in the segmentation task and an accuracyof ۰.۹۶ and AUC of ۰.۹۷ in the classification task. The model successfully identified infectedareas in lung CT scans and segmented them accurately.Conclusion: This paper proposes an efficient deep multi-task learning structure for Covid-۱۹disease, which uses image processing algorithms in the pre-processing phase. The model showedthe highest results among previous studies. The proposed model could be applied as a primaryscreening tool to help primary service staff better refer suspected patients to specialists.

کلیدواژه ها:

COVID-۱۹ ، Lung infections ، Computed Tomography scan ، Automated diagnosis ، Multi-task structure