Otimização experimental e avaliação de métodos de segmentação de micro imagens usando luz síncrotron: uma aplicação na herpetologia

Abstract

Synchrotron radiation X-ray microtomography is a high-resolution imaging technique that allows visualizing and quantifying structures inside samples. In the field of herpetology, micro-CT has been bringing innovations to research due to the facility that this technique allows for the study of reptile and amphibian morphology. In the present work, samples of the tadpole of the species Thoropa miliaris at developmental stages TM28, TM32, TM36, TM40 and TM46 were used as a biological model for the study of microtomographic images obtained by propagation-based phase-contrast imaging (PBI). The samples were measured in liquid medium (alcohol) and positioned at different sampledetector distances (SDD = 5 cm, 15 cm, 25 cm). The experimental setup was set up on the SYRMEP beamline of the synchrotron laboratory ELETTRA/Italy. Three regions of interest (ROIs) were chosen to be analyzed using the image quality parameters of contrast (C), contrast-to-noise ratio (CNR), signal-to-noise ratio (SNR), and edge enhancement to noise ratio (EEN). The ROIs in focus were the lens, muscle tissue, and notochord. ANOVA analysis of variance (significance level α = 0.1) was performed to investigate how the chosen SDDs affected the quality parameters. For all samples, the calculation of quality parameters C, CNR, SNR, and EEN in ROIs was significantly different (p << 0.1) as a function of increasing SDDs, as expected, but no great effect was observed from increasing SDD from 15 cm to 25 cm on average. To analyze the effects of chemical staining on soft tissue, the TM28, TM32 and TM36 specimens underwent the chemical staining process where they were immersed in 1% metal iodine solution for 2 hours and 12 hours and new micro-CT images were processed. Quantitative analysis was performed using the quality parameters on this new data focusing on the three ROIs. For the lens and muscle tissue of the TM28 sample, the differences between the contrast, SNR and CNR of the unlabeled and labeled images were statistically significant (p << 0.1). Regarding the immersion time of the TM28 sample in the iodine solution, for the lens, only SNR was statistically different (p << 0.1), for C and CNR no statistically significant differences were observed (p > 0.5 and p > 0.4, respectively). For muscle tissue, C, SNR and CNR were statistically different (p << 0.1). It was observed that the interior of the notochord was not iodine stained as expected and, as a consequence, impaired the visualization of the internal structures that make up the notochord. It can be concluded that iodine staining was efficient in increasing the contrast of the lens and muscle tissue structures. However, increasing the immersion time of the sample in the iodine solution does not seem to have been an important parameter in this case. To study the post-processing of the micro-CT data, which includes the identification of the tadpole structures T. miliaris using image segmentation, an experiment was carried out on the IMX beamline of the synchrotron light laboratory LNLS/Brazil. From the obtained images, methods 1 and 2 of semiautomatic segmentation (Interpolation and Watershed, respectively) and method 3 of automatic segmentation (U-Net architecture algorithm) were applied to the following ROI: the lens. The evaluation of each method was performed by volumetric and statistical analysis of the rendered volume. The results obtained showed that the greatest difference in the volumetric quantification was between methods 1 and 3 with a reduction of 4.24% in the rendered volume. In the statistical analysis performed by the Dice Similarity Coefficient (DSC), methods 1, 2, and 3 had averages of 97.02%, 95.41%, and 89.29%. Although method 3 obtained the lowest DSC, this average is compatible with other results in the literature. Another important factor was the time consumed for segmentation of the lens: semiautomatic methods were around 1h 30min and the automatic method was between 8 and 26 seconds. In addition, it was possible to establish a baseline for future evaluations of DNNs (Deep Neural Networks) applied to PBI volumes.