OBJECTIVES: Lung segmentation in CT images represents a fundamental process for quantitative evaluations of changes in lung parenchyma density and volume as well as for radiomics investigations, in order to assess the frame, extent, and severity of diffuse lung pathologies. A relevant limitation of commonly used segmentation software is the difficulty or inability to properly detect the lung/chest-wall interface in the case of pathologically increased parenchymal density (e.g. ARDS or COVID-19) adherent to the chest-wall. In order to overcome such limitation and, at the same time, to avoid time-consuming manual segmentation we developed an innovative semi-automatic algorithm.
MATERIALS & METHODS: The actual lung parenchyma volume is identified by modelling lung edges with appropriate spline functions calculated by considering shape and position of lung neighboring anatomical districts and local density patterns (pixel-based radiomics). Thereafter the internal high-density pathological regions are segmented with proper thresholds.
RESULTS: The algorithm segmentation accuracy was compared to the one of experienced radiologists showing performances at least not inferior to that of their manual segmentation.
CONCLUSIONS: A new algorithm, (international patent pending) was developed using an innovative approach to accurately segment lung parenchyma and, in particular, consolidative tissues, even in cases where commercial algorithms tipically fail, such as when these tissues adhereto the lung wall.