Pelvic floor dysfunctions represent a spectrum of disorders affecting a significant number of women, manifesting through various symptoms such as pelvic organ prolapse, and urinary and fecal incontinence. To decipher the complexities of these symptoms, both biomechanical simulations and clinical investigations have been employed, underlining the indispensable role of imaging technologies in the accurate diagnosis of pelvic floor disorders. The detailed depiction of pelvic anatomy is crucial for this purpose, although it faces challenges like noise interference and the partial volume effect, owing to the intricate nature of pelvic structures.
Magnetic resonance imaging defecography (MRI defecography) emerges as a specialized MRI technique for examining the pelvic floor and rectal area during the process of defecation. This method provides detailed and dynamic images of the rectum and adjacent regions. Furthermore, the advancement towards an automated segmentation approach is set to revolutionize the analysis of pelvic floor disorders by providing objective data, thereby improving the speed, efficiency, and consistency of segmentation-based biometric analyses. This study is dedicated to developing an automated, rapid, and dependable method for vaginal canal segmentation and biometric measurement extraction from MR defecography images.

We utilized the nnU-Net segmentation model to accurately delineate various pelvic structures, including the vaginal canal, pubic symphysis, sacrum, bladder, and rectum within T2-weighted MRI defecography images. These images, in DICOM format from MRI defecography, were imported into the 3D Slicer software. Here, we used orthogonal projection images to carry out segmentation, meticulously annotating the vagina, bladder, symphysis pubis, sacrum, and rectum on individual layers using manual drawing tools.
We annotated a collection of 47 three-dimensional grayscale MRI images for vaginal region segmentation, conducted by an expert in the field. The resultant segmentation files are comprised of 3D binary data, with the value "1" signifying the vaginal area and "0" representing other tissues or background. Both the annotated images and their corresponding segmentation data were stored in the NRRD file format. We organized the dataset into five groups or folds for analysis (for folds 1-2: 37 images were used for training and 10 for testing; for folds 3-5: 38 images were allocated for training and 9 for testing).
The nnU-Net model was trained separately on each of these folds using a 5-fold cross-validation strategy, operating in a 3D full-resolution setting. The model configurations included a batch size of 2 and a training duration set to 1000 epochs. For training purposes, images within each fold were further split into training and validation groups, adhering to a standard ratio of 4:1 (for folds 1-2: 30 training and 7 validation images; for folds 3-5: 30 training and 8 validation images).
To assess the accuracy of our medical image segmentation, we calculated both the Dice similarity coefficient (DSC) and the Intersection over Union (IoU) coefficient on the test images across all folds. These metrics are widely recognized for their effectiveness in evaluating the precision of segmentation in medical imaging.

Our nnU-Net segmentation model outperformed other tested models, achieving a notably higher average Dice similarity coefficient (DSC). The overall mean DSC was recorded at 0.9714, indicating exceptional segmentation accuracy. The model also demonstrated a robust performance in terms of the Intersection over Union (IoU) score, averaging at 0.9675, with scores spanning from a low of 0.7254 to a high of 0.9896.

Remarkably, our algorithm was capable of processing an MRI defecography sequence in under one second, showcasing its efficiency. In comparison, manual segmentation of the vaginal canal, conducted by an experienced technician, required between 30 to 45 minutes per image, not including the additional time required for training.

The developed algorithmic pipeline significantly enhances both the efficiency and reproducibility of image segmentation for diagnosing pelvic floor disorders when compared to traditional manual methods. This pipeline supports more accurate biomechanical analysis of pelvic floor disorders and facilitates the creation of personalized treatment plans tailored to the unique anatomical and biomechanical characteristics of individual patients.

Continence 12S (2024) 101477
DOI: 10.1016/j.cont.2024.101477