The aim of this work is to delineate the lungs in three-dimensional image sequences of the same subject, with decreasing inhomogeneous contrasts, and to assess the applicability of the method to quantify the aeration in subjects diagnosed with acute respiratory distress syndrome (ARDS).

Assuming that the first, and most-contrasted, image of the sequence can be successfully segmented by an existing method, we propose to align the remaining images with the first one, using a cascade of successive registrations, and then to deform the initial segmentation, using a cascade of transformations. Because the lungs slide against the rib cage, the registration requires the use of motion masks that separate the moving organs from less moving ones. Nevertheless, while such masks can be relatively easily obtained in well-contrasted images, obtaining them in images locally lacking contrast is challenging. Our main contribution is the method proposed to generate appropriate motion masks for the entire image sequence despite the decreasing contrasts. This method also uses the principle of cascade registrations and transformations starting from an initial mask obtained in the most-contrasted image of the sequence.

The entire segmentation method was applied to CT images of 16 piglets with induced ARDS, 20 through 35 images per piglet, acquired at varying mechanical-ventilation conditions.

The processing time was less than 15 minutes/image, on average. Dice similarity index with an independent standard was of 0.963 globally, but clearly worse overlap was locally measured in the bottom-most region of the lungs (average Dice of 0.810), where the largest displacements and contrast variations were observed.

The proposed method allows lung segmentation despite the loss of contrast, and can be used to quantify the lung aeration in subjects with ARDS. An improvement in the bottom-most region may be foreseen by combining gray-level based registration and anatomical-landmark matching.