Statistical Shape Modeling of Nulliparous, Pregnant, and Parous Female Pelvic Floor Muscle Complexes

Description

This data and associated research hypotheses, findings, and interpretations are described in detail in the associated/linked article. Briefly, we aimed to quantify pelvic floor muscle complex (PFMC) shape variation and compare the shapes of nulliparous (have never given birth), late pregnant (currently in the 3rd trimester at the time of imaging), and parous (have given birth but are not currently pregnant) women via statistical shape modeling (SSM). We hypothesized that late pregnant PFMCs would significantly differ from those of nulliparous and parous women. We also expected the nulliparous and parous shapes to differ, which would support the existence of long-term shape alterations following pregnancy and childbirth. PFMCs were segmented from pelvic MRIs of women age 20-49 collected retrospectively. PFMC segmentations included the coccygeus, levator ani, and superficial perineal muscles, the external anal sphincter, and surrounding connective tissues (the perineal body, perineal membrane, and annococcygeal ligament) and were segmented as one continuous structure. In total, 48 PFMCs were included in the shape analysis: 17 nulliparous, 14 late pregnant, and 17 parous PFMCs. A major step of a SSM is a principal component (PC) analysis, which outputs modes of variations (or PCs) each composed of an eigenvector and eigenvalue. PC scores are the projections of patient-specific data onto those eigenvectors and describe where each shape lies (i.e., how near to or far from the mean) along mode of variation. This statistical shape model had 7 significant modes of variation that described 34.2%, 18.1%, 6.2%, 5.5%, 4.7%, 3.6%, and 3.0% of the total shape variance. The average shapes were generated by averaging the PC scores across each of the 7 significant modes and inputting those scores back into the shape model. Average refers to the overall average, while the other shapes are group averages. While only 7 modes of variation were significant, 47 were generated. The PC scores file includes PC scores from all modes, but only the first 7 should be used to recreate this data. The first 3 modes in particular highlight significant differences between these 3 patient groups and are discussed extensively and visualized in the associated/linked article. It should be noted that while these differences are subtle (small differences involving many mesh vertices), they are statistically significant. Roughly 25% of the total shape variance corresponded with significant differences between nulliparous, late pregnant, and parous PFMCs, while an additional 34% describes significant differences between nulliparous and late pregnant PFMCs. Overall, the late pregnant group was the most distinct of the 3 patient groups, with many structures being in a relatively more inferior/posterior position and/or wider in late pregnant compared to nulliparous and parous women.

Steps to reproduce

The methods describing how these shapes and principal component scores were generated are explained in detail in the associated/linked article. Briefly, pelvic floor muscle complexes (including the cocccygeus, levator ani, and superficial perineal muscles) were segmented from pelvic MR images of nulliparous, third trimester pregnant (at the time of imaging), and parous (non-pregnant at the time of imaging) women collected retrospectively. These shapes served as the inputs for a statistical shape analysis. Segmentations were smoothed and corresponding points established using Deformetrica's atlas-registration function. Then Procrustes, principal component, and parallel analyses were carried out in Mathematica. These analyses removed differences due to translation, rotation, and scale; determined the principal components (PCs) and PC scores; and identified significant modes of variation. These PC scores are the projections of patient-specific shape data onto eigenvectors and describe where an individual shape lies along each mode of variation. PC scores were averaged across groups and served as inputs back into the shape model, which allowed for the generation of the overall and group average shapes (each defined by a surface mesh with 16,662 triangular elements and 49,986 vertices).

Institutions

Categories

Funding

Related Links

Licence