This thesis focuses on the automatic image analysis of transperineal ultrasound (TPUS) data. 2D, 3D, and 4D TPUS are used to investigate female pelvic floor problems. These problems have a high prevalence, but the understanding of pelvic floor (dys)function is still limited. The full potential of TPUS analysis of the pelvic floor is limited by the fact that this analysis is still done manually; Making it time-consuming and observer-dependent. This hinders both the research into understanding TPUS and the clinical use of TPUS. To overcome these problems we use automatic image analysis to ease the use of TPUS in clinical practice and broaden our understanding of the pelvic floor mechanics.
Currently, one of the main methods used to analyze the TPUS in both research and clinical practice is manually selecting the slice of minimal hiatal dimensions (SMHD) and measuring the dimensions of the urogenital hiatus and puborectalis muscle in this slice. In the first chapter of this thesis, we show that reliable automatic segmentation of the urogenital hiatus and the puborectalis muscle in the SMHD can be successfully implemented, using deep learning. In the next chapter, we show that deep learning can also be used to successfully automate the full process of selecting and segmenting the SMHD, with human-level performance.
4D TPUS is available in the clinical practice but by the aforementioned method of selecting the SMHD the information used is reduced to length, distance, and gray-value measurements. Therefore, information about the volume appearance of the pelvic floor muscles and muscle functionality is not quantifiable. In the third chapter of this thesis, we propose a reproducible manual 3D segmentation protocol of the puborectalis muscle. Furthermore, we show that the manual segmentations resulting from this protocol can be used to train active appearance models that can be used for reliable automatic 3D segmentation. Since deep learning is the state of the art for image segmentation we show its success in this automatic segmentation task in the next chapter.
During this study data from a newer TPUS machine became available. On this data, it is possible to identify all subdivisions of the main pelvic floor muscle group, the levator ani muscles. The protocol to identify and segment these subdivisions is presented in the fifth chapter of this thesis.
All segmentations methods presented in this thesis rely on supervised learning, which means that they require manual labels in order to learn their segmentation tasks. To investigate patterns within the TPUS data that are available without requiring manual labels we used unsupervised deep learning in the last chapter.
Using a convolutional auto-encoder we compress 3D TPUS frames to a latent feature vector of 128. Using these feature vectors, we show that it is possible to discriminate with 90% accuracy between TPUS movies containing pelvic floor muscle contraction or a Valsalva maneuver.
The segmentation results presented in this thesis are an important step to reduce the TPUS analysis time and will therefore ease the study of large populations and clinical TPUS analysis. The 3D identification and segmentation of the levator ani muscle subdivisions help us to identify if they are still intact. This, therefore, is an important step to better informed clinical decision-making.