COVID-19 Volumetric Pulmonary Lesion Estimation on CT Images Based on Probabilistic Active Contour and CNN Segmentation
Conference Paper
Overview
Research
Identity
Additional Document Info
View All
Overview
abstract
Quantifying the affected lung areas in a COVID-19 patient is important to assess the disease progression and treatment response. It also helps with phenotype differentiation and prognosis prediction. In a clinical setting, this quantification is performed manually by an expert radiologist, based on their experience and knowledge, however it is a time-consuming task with inter-observer variability, creating thus the need to develop automatic algorithms to delimitate diseased areas. Convolutional neural networks (CNNs) have been widely used for this purpose. However, these architectures require a substantial amounts of labeled data and time to be trained, imposing a limitation to their usage.In the present study, an alternative two step-methodology is proposed to obtain a volumetric estimation of the COVID-19 related lesions on CT images. The first, and critical step, consists of performing a lesion segmentation using a non-supervised approach based on a probabilistic active contours algorithm. The second step uses a pre-trained CNN to segment the lung parenchyma. Such neural network is trained with non COVID-19 images; its only purpose is to obtain a whole-lung volume estimation. Finally, the resultant masks of lesion and parenchyma are used to compute the percentage of the lesion in the lungs.Validation of the proposed segmentation approach was made on a publicly available dataset with 20 CT COVID-19 labeled images. The median dice coefficient value for these 20 images was 0.66; a value comparable with results found in the literature.The proposed approach was applied on other 295 low and high-resolution CT images from COVID-19 patients, acquired at the Instituto Nacional de Ciencias Médicas y Nutrición Salvador Zubirán (Mexico City, Mexico). The dataset was used to compare the amount of lesion on survived and deceased patients. The results show a significant difference on percentages of lesion between the groups with a p value of 9.1 x 10-4 in low resolution CT and 5.1 x 10-5 for high resolution images. Furthermore, an average difference of 10%25 on the percentages of lesion between high and low resolutions images was found. This difference could be important to take into account in those places where high resolution CT is difficult to access.
