STUDENT PROJECT IDEAS

Multispectral Image Registration

Registering multiple images of a scene acquired at different illumination wavelengths is a difficult challenge because the image appearance varies with diffent lights and because deformation may occur between image aquisition points.

 

 

Details and prerequisite skills: Existing code written in C++ will be a starting point for the project and other implementations for Matlab are available online. The project will require programming in Matlab and/or C++ and would benefit from experience and interest in image processing or computer vision.

Further reading:

Learning Local Propagation Strategies

Recovering motion and 3D shape from videos usually requires solution to the correspondence problem. For surgical videos local approaches are effective for correspondence search as they can easily incorporate strategies to handle specular highlights, instruments and shadows. In this project the aim will be to investigate and potentially use machine learning techniques in order to improve best-first correspondence growing. An example of growing correspondence with this scheme is shown in the video on the left where we compute 3D structure from stereo starting with a set of sparse seed points. Improvements could be defined in terms of computational optimization, better reasoning or theoretically justifying growth decisions.

Details and Prerequisite Skills: Existing code written in C++ will be a starting point for the project and other implementations for Matlab are available online. The project will require programming in Matlab and/or C++ and would benefit from experience and interest in image processing or computer vision.

Segmentation and Grouping Using Motion Coherence

Automatic understanding of surgical images requires the segmentation of the image into different categories identifying the surgical tools and different tissues or anatomical structures. The problem can be approached by using image information alone and performing classification based on colour or texture features. However, with recent advances in recovering motion and shape at the surgical site this information can provide additional constraints to classification strategies. The aim of this project will be to explore how motion and shape can be incorporated or grouped to identify uniform tissue regions, the surgical tools and areas affected respiratory or cardiovascular motions. We will explore methods for obtaining motion firstly in 2D, with the possibility of looking at 3D motion, and then clustering the motion vectors for example using k-means or similar techniques. The project will require programming in Matlab and/or C++ and would benefit from experience and interest in image processing or computer vision.

Multispectral Image Deblurring

Images acquired at different illumination wavelengths during surgery can potentially provide information about changes in the concentration of oxy and deoxy-hemoglobin (HbO and HbR) in the blood. Multispectral imaging can be performed with lightweight equipment that is easily integrated in the operating theatre, does not require contact with the exposed tissue and does not interfere with the surgical instruments. However, acquiring images at low wavelengths (<500nm) requires the camera shutter to remain open for a significant period of time and when the tissue is undergoing respiratory or cardiac motion the resulting images appear blurred. The aim of this project will be do work with a custom trinocular channel endoscope and to develop algorithms for deblurring multispectral images using motion computed from white light stereo images or using priors on the tissue shape. Development of the algorithms and mathematics can be carried out in Matlab with potential for real-time implementations using C++ and/or GPU accelleration.

Scene Flow and Non-Rigid Structure-from-Motion

The focus of this project will be to develop algorithms for sparse and quasi-dense structure from motion from stereo sequences in robotic surgery. In robotic surgery, recovering the motion of the imaging device within a deformable surgical environment is an important aspect of providing accurate localisation. This problem is highly challenging as rigidity constraints typically underpin localisation methods based on vision. With a stereo-laparoscope it is possible to relax these explicit constraints on certain areas of the operating field where there is local rigidity or it is possible to model the motion of deformable regions with specific rhythmic motions due to respiration or the heart beat. Additional constraints can be derived from the motion of specular highlights, which will be informative given the special arrangement between camera and light source in laparoscopes. Development will be in Matlab where algorithms can be prototyped quickly with real-time C++ implementation following that if there is sufficient time.

Catheter Detection and Tracking

Tracking guide wires (GW) and catheters during endovascular procedures has many applications, including interventional navigation for accurate localisation within the vasculature and adaptive image enhancement of the GW for clearer visualisation during procedures. Accurate positioning of the GW and catheter with regard to the vasculature is a prerequisite for the successful deployment of stent grafts in procedures for aneurism repair and the treatment of atherosclerotic plaque. However, fluoroscopic image quality of the GW is usually restricted by restricting the radiation exposure of the patient, the use of contrast agent and the inherent physiological motion artefacts due to the heart cycle and respiration. Therefore GW detection, tracking and enhancement is an important but challenging task. This project will focus on the detection and tracking of catheters and guidewires from existing data of surgical simulators, phantom models and actual interventional procedures. Methods can be develop in Matlab or C++ and the most recent algorithms in this field use machine learning techniques so some background in the area would be helpful.

Skill Evaluation in TOE Simulation

This project will focus on analysing instrument motion for understanding and objectively measuring operator learning curves and skill in Transesophageal Echocardiography (TOE) procedures. The tool motion will be recorded from an advanced simulation environment for developed by HeartWorks (http://www.heartworks.me.uk). TOE is now routine for monitoring cardiac function and surgical treatment in the operating theatre, as well as on the cardiac ITU. TOE requires the insertion of an ultrasonic transducer into the gastroesophageal tract and the correct manipulation of the transducer to visualise multiple planes through the heart. The HeartWorks simulator allows trainees to develop the core skills required for performing TOE examinations but currently the learning process and the relationship between the motion of the transducer and the skill of the operator is poorly understood. The aim of this project will be to investigate the rich information provided by the simulation environment in order to develop evaluation metrics that can provide real-time feedback during training and quantitative assessment methods. Development will initially be performed using Matlab and will involve a combination of signal processing techniques such as Dynamic Time Warping (DTW), dimensionality reduction algorithms like Principal Component Analysis (PCA) and machine learning techniques for classification.