Hypothesis / aims of study
In patients who are unable to void on a uroflowmetry device, urethral resistance cannot be determined during urodynamic study. Videoflow is an algorithm that can derive uroflowmetry curves from standard care video urodynamic (VUDS) images [1]. In recently published versions of Videoflow the bladder was manually segmented on the VUDS images. The current study describes the validation of a fully automated machine learning (ML) process for segmentation and converting VUDS images into a uroflowmetry curve.
Study design, materials and methods
A multi-stage ML approach was chosen for segmentation of the bladder on VUDS images. Initially, a nnUNet model was trained on a random sample of 769 images taken from a dataset of 99218 VUDS images [2]. Subsequently, a second video object segmentation model (Cutie) was added to the analysis [3]. Cutie implements so-called memory frames, which are earlier segmentations of the object. This enables the ML algorithm to track an object over time, providing a starting point, which was given by the nnUNet segmentation. nnUNet and Cutie segmentation quality was assessed by the Dice Similarity Coefficient (DSC). Finally, the constructed Videoflow curve was compared with the standard uroflowmetry curve.
Interpretation of results
The use of a multi-stage ML model for segmenting the bladder has a significant advantage over the single stage nnUNet. Especially for the smaller bladder areas, the sequential object tracking of Cutie resulted in a significantly better segmentation of the bladder. This better segmentation has a direct effect on the performance of the Videoflow algorithm, as visualized by the accuracy of the maximum flow.
An interesting dependency of the segmentation accuracy on the bladder size was noted for automatic segmentation by the nnUNet algorithm. Especially for small bladder areas at the end of the voiding, the nnUNet did not recognize the bladder contour correctly, which resulted in huge flow spikes ‘constructed’ by the Videoflow algorithm directly on the nnUNet results. This dependency was also found in earlier studies implementing nnUNet, and is a generally acknowledged downside of UNet like ML algorithms. Due to the inherent down sampling process of the UNet architecture, small regions may be lost, resulting in poorer performance in smaller bladder areas.
To overcome this, we used Cutie as a secondary segmentation stage, which implements object tracking over time. Use of Cutie resulted in a significantly better performance after the second stage, especially for the smaller bladder areas on the VUDS images. Although the combined ML model still showed dependency on the segmentation accuracy of the segmented bladder area, this dependency was significantly reduced compared to using the nnUNet model alone.