Thread by @helenaw49365896

*Pelvic floor disorder clinic*

The artificial intelligence (AI) agent will see you now...

Does this sound like science fiction to you? Follow this #ThesisThread to see how this is more science than fiction! 1/19

To start: pelvic floor disorders (PFDs) are a group of conditions (i.e. incontinence, prolapse, sexual dysfunction..) caused by weakening of the pelvic floor. 25% of women experience at least 1 PFD.
(@tenawomen may advertise that peeing yourself is normal, but it is not!) 2/19

A doctor needs to investigate your symptoms👩‍⚕️🔎
But how? Gone are the days of physical inspection (no thank you forceps), ultrasound (US) imaging is here!
US allows us to see the pelvic floor function in real time and in 3D! It is cheap, accessible and comfortable.
3/19

The doctor will perform medical imaging analysis tasks on the US volume (i.e. manipulate the 3D image) to acquire standardised measurements that can lead to diagnosis🎉. The issue is, these manual tasks are prone to error and require a high level of expertise and time! 4/19

Therefore, we focused on creating algorithms to automate these manual tasks used in PFD assessment.
To lower the level of expertise required to perform PFD assessment, save time and reduce the error between doctors. 5/19

To automate PFD assessment we utilise advances in AI.
A convolutional neural network (CNN) is a deep learning AI algorithm that is used in many medical imaging analysis tasks. It is a data-based algorithm that learns features that are not always visible to the naked eye. 6/19

State of the art (SOTA) research showed 2D segmentation of the levator hiatus (the largest hernia portal in the human body) from a manually extracted 2D plane (ref. C-plane). The area is used to diagnosis prolapse. Unfortunately the C-plane has to be manually extracted🤷‍♀️. 7/19

So we developed a C-plane detection algorithm. The C-plane was extracted from an US volume by following the clinical workflow and detecting physical landmarks (the pubic bone and levator ani muscle) used by clinicians. CNNs were trained to detect the landmarks automatically. 8/19

We evaluated our automatic pipeline on 100 US volumes from the PFD clinic @UZLeuven, the algorithm's error was within the measured inter-rater error, and the time required to perform C-plane extraction significantly reduced from 120s to 20s⏲️. 9/19 link.springer.com/chapter/10.100…

We extended the pipeline to incl. 2D levator hiatus segmentation. The extracted C-plane was fed into a 2D CNN (lev. hi.). We evaluated on 73 US volumes, the algorithm's error was lower than inter-rater error, and time reduced from 2 mins 37s to 37s. 10/19 sciencedirect.com/science/articl…

Unfortunately, automatic segmentation is never 100% accurate, if they are to be used for diagnosis - it's really important doctors have liability & are able to correct poor segmentations intuitively. So we developed a 2D interactive segmentation tool with doctors in mind. 11/19

The tool takes an automatic segmentation, and represents it as an active contour which can be adapted by adding user-defined points. The contour evolves in real time, and the doctor can achieve a clinically acceptable contour. Check out the demo below! 12/19

The tool was evaluated on 30 C-planes and compared to 2 clinical tools and a SOTA tool called UGIR. The time taken and the perceived workload were measured. To measure workload the NASA-TLX survey was used, where a higher score=higher workload. 13/19 link.springer.com/chapter/10.100…

The proposed tool was less frustrating, required a lower mental, physical and temporal demand and had a better performance than the other tools. The NASA-TLX score was 23 cf 58-70 for the other tools and the time taken reduced significantly from 38-70s to 16s. 14/19

This concludes the analysis of the levator hiatus. We moved onto another structure that is often neglected in analysis due to its difficulty. The anal sphincter (AS) complex is a group of muscles supporting defecation, and anal incontinence is a highly prevalent PFD. 15/19

US can be used to assess the integrity of the AS by acquiring coronal-view slices across the total AS. These slices show tears (if present) and can indicate injury. We automated this clinical workflow by utilising 3D CNN segmentation of the external anal sphincter (EAS). 16/19

The pipeline segments the EAS and horizontalises it. We extract 8 slices across the AS comparable to the clinical workflow. We evaluated on 30 US volumes and achieved a clinical acceptability of 90%, and reduced the time required from 61s to 8s 🎉⏲️. 17/19 link.springer.com/chapter/10.100…

In this work we developed automatic medical imaging analysis tasks to aid diagnosis but not to replace a doctors diagnosis in PFD. For these tools to be clinically accepted it is paramount that they are understandable, intuitive and allow for correction by doctors. 18/19

To conclude, using AI we reduced the time and entry level of expertise required to perform several medical imaging analysis tasks for PFD assessment. Making difficult tasks more accessible to medical centres in the future, while reducing error between doctors. 19/19