Computer-guided ureteroscopy and laser lithotripsy

Thesis

Ureteroscopy with laser lithotripsy has evolved into a minimally invasive routine treatment for stones in the urinary tract. Automated segmentation can assist the urologists during the procedure and is also the primary step towards the estimation of stone size, one of the most important parameters to decide if the stone requires further fragmentation. Continuous irrigation is maintained during the procedure for optimal visualisation and to reduce the excess heat generated from the laser. The irrigation flow is increased when the visualisation is impaired by stone debris. However, this increases the intrarenal pressure which can potentially lead to several complications. This highlights the need for computer vision algorithms that can virtually flush out the debris whilst keeping the irrigation flow low. The ureteroscopy and laser lithotripsy datasets suffer from challenges such as small field of view, occlusions from stone debris, intra-operative bleeding, motion blur, specular highlights, dynamic background, varying illumination conditions, stone heterogeneity and artefacts from the turbid fluid. In addition, there is both inter-patient and intra-patient variability in terms of tissue appearance, illumination conditions, stone variability, and different viewpoints with respect to the camera. This makes it even more challenging to develop generalisable machine-learning models for ureteroscopy and laser lithotripsy. To this date, there are no comprehensive computer vision studies that explore or solve the real-world challenges of ureteroscopy and laser lithotripsy.

In this thesis, novel datasets have been acquired and annotated to facilitate the development of computer vision algorithms for ureteroscopy and laser lithotripsy. This thesis presents novel algorithms that leverage motion information between frames for improved segmentation of stone fragments and laser fibre. This thesis also presents frameworks for unsupervised domain adaptation and improved generalisation of segmentation models for various endoscopy datasets. Further, this thesis investigates turbidity mitigation in lithotripsy datasets for an improved understanding of the challenging problem and proposes a framework to effectively enhance image quality.

The original contributions are as follows: (1) Classical approach to stone segmentation in ureteroscopy and laser lithotripsy datasets. (2) Deep learning-based segmentation of stone and laser fibre in ureteroscopy and laser lithotripsy datasets. (3) Improving domain adaptation and generalisation of segmentation models using Variational Autoencoders (VAEs) in various endoscopy datasets. (4) Turbidity mitigation for improved visualisation in laser lithotripsy. (5) New synthetic, animal and clinical ureteroscopic laser lithotripsy datasets acquired and annotated to facilitate the development of computer vision algorithms.

Authors: Gupta, S

• DOI: 10.5287/ora-bmp9k4zxa
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: kidney stones; deep learning; laser lithotripsy; turbidity mitigation; ureteroscopy; segmentation
• Subjects: Image processing; Artificial intelligence; Computer vision