Microtomografia por luz síncrotron usando recuperação de fase aplicada ao estudo morfológico do Aedes aegypti

Abstract

Synchrotron X-ray microtomography is a technique capable of generating high-resolution and high-contrast images, enabling the study of both external and internal structures of various types of samples. When applied to biological samples, phase contrast is used as a tool to enhance image contrast. In the present study, a male specimen of Aedes aegypti was employed as the object of study to determine the impact of the phase recovery parameter (δ/β) proposed by Paganin et al. (2002) on the image of the mosquito. The specimen was analyzed as part of a scientific commissioning proposal for the MOGNO beamline at the Brazilian Synchrotron Light Laboratory (LNLS), using the SIRIUS source. It was scanned in a liquid medium (isopropanol) with a sample-to-detector distance of 1.7 meters and a pixel size of 1.142 μm. During the phase recovery process, six different ratios (δ/β = 1000, 250, 100, 10, 5, and 1) were used, and the brain was chosen as the region of interest (ROI). To quantify the differences between the images, two methods were employed: intensity profiles and segmentation. Preliminary tests excluded the three highest ratios due to image blurring, which impaired the sharp visualization of internal structures within the ROI. For the intensity profiles, three sub-regions of interest containing internal brain structures were defined. In both sub-regions, the ratio 10 showed better contrast between the analyzed structures. Deep learning tools were utilized for brain segmentation using the U-net network, and six models with different frame sets (containing 12, 24, 36, 36 – retrained, 48, and 72 frames) were evaluated. A comparison between the segmentations obtained was performed using the Dice Similarity Coefficient (DSC) and ANOVA with Tukey’s post-hoc test. The model consisting of 12 frames (for the three evaluated ratios), ratio 1 (for the models containing 12, 36 – retrained, and 72 frames), and ratio 10 (for the 24-frame model) exhibited the lowest DSC values. Additionally, it was observed that if the data are normalized beforehand, a model trained on one ratio can be applied to the segmentation of another, simulating the segmentation of the same structures in different samples.