Semi-Supervised Multi-Scale Dense Based Encoder-Decoder Architecture for Lung Computed Tomography Image Segmentation

ABSTRACT

Covid-19 pandemic continues to threat health of the global population, an efficient way to restrain the Covid-19 outbreak is timely screening suspected cases for quarantine and treatment. Despite of pathogenic laboratory testing is the gold standard to screen suspected cases, but it may obtain false negative results and consuming a lot of time. Computed tomography of chest can be an alternative diagnostic method to screen suspected cases that is based on radio graphical changes in lung area of Covid-19 confirmed case. Precisely delineate the lung area that is first and critical step for screening computed tomography image of chest by using deep learning method. In this paper, several related previous works will be introduced at first. Then, an improved encoder-decoder based segmentation framework is proposed, which is integrated with multi-scale densely connection-based convolution block and skip connection path. Moreover, in model training process, a semi-supervised manner is applied to train model which can reduce the demand of labeled training data. Finally, the proposed method is tested with a dataset of public X-ray image of chest. The experiment test proposed model in this paper with varieties of segmentation methods and result demonstrates promising performance of proposed model that against several other deep learning models. © 2022 SPIE