Research team
Wireless capsule endoscopy based on Gaussian process latent variable models.
Abstract
Endoscopy plays a pivotal role in both diagnostic examinations and minimally invasive surgical procedures. A special type is wireless capsule endoscopy, where patients ingest a small pill-shaped camera. Despite its importance, the endoscopic images and videos exhibit serious drawbacks, such as substantial distortion, low resolution, missing frames, specular reflections, and so forth. In this project, I will tackle several of these challenges. In order to do so, I will first develop a novel endoscopic camera calibration procedure. Next, based on this, I will adopt approaches from the field of Gaussian process latent variable models and the world of generative AI in general to formulate models that construct an alternative latent space representation of the data. Probabilistic machine learning models, such as Gaussian processes, offer interpretability (no black-box), which is especially crucial in evidence-based medicine as it offers transparency and helps build trust with clinicians. Improved camera calibration and innovative perspectives on latent spaces hold the potential to revolutionise various techniques, including 3D trajectory estimation, mosaicking and many more. As a result, my research stands to significantly enhance the precision and efficiency of clinicians when interpreting endoscopic images. This, in turn, promises to elevate detection rates, enhance the accuracy of abnormality size measurements, and contribute to the advancement of minimally invasive surgery.Researcher(s)
- Promoter: Penne Rudi
- Fellow: De Boi Ivan
Research team(s)
Project type(s)
- Research Project
A general line variety model for sensors, allowing stable calibrations that meet the accuracy standards for medical applications.
Abstract
The popular pinhole model for imaging sensors and the associated calibration procedures appear to be inadequate for some of the new generation sensor technology. Even for classical RGB cameras, this standard model leads to unstable calibrations, with the need for an extra model to remove lens distortion. We propose line varieties as a unifying modelling for a broad set of sensors. As opposed to other previously published attempts in this direction, we identify the sub-varieties that correspond to real sensors. This enables us to extend interpolation techniques and Gaussian processes, to support sensor calibration from small samples of lines. We aim fundamental contributions to the fields of Line Geometry and Probabilistic Numerics. Our goal is to develop the framework for multi-sensor configurations (laser scanners, IR-cameras,…), providing measurement fusion, using the developed line models, and to achieve accuracy levels for sensor-supported Radiotherapy.Researcher(s)
- Promoter: Penne Rudi
- Fellow: De Boi Ivan
Research team(s)
Project type(s)
- Research Project