Towards a Clipless-Workflow in Ocular Proton Therapy

Abstract

Ocular proton therapy (OPT) is a special niche of proton therapy. It was originally developed at the Massachusetts General Hospital (MGH) in Boston and was adapted and improved at different facilities ever since. The number of patients treated at PSI lies at 200-250 per year. An important aspect of the current treatment workflow is the use of radiopaque markers (clips) which need to be surgically attached to the surface of the eye surrounding the tumor. These clips fulfill three different tasks. First they are used to create an eye model, by which a spheroid is fitted to the spatial positions of the clips as determined by x-ray imaging. Second they serve as landmarks for defining the tumor outline and therefore create the target volume. Finally they are used to position (navigate) the patient in the treatment room at each treatment fraction. However, the surgical procedure for placing the clips is expensive and uncomfortable for the patient. In this thesis therefore we investigate whether OPT be performed without the use of these clips. As such, alternatives need to be found and evaluated for all of the above mentioned tasks of the clips. In particular, the use of MR imaging for defining the eye model, tumour and critical structures has been investigated (chapter 3), an automated treatment planning system for OPT, also capable of supporting the MR model, developed (chapter 4) and two prototype optical eye tracking systems implemented and tested to provide clip-less navigation of the eye on a daily basis (chapters 5 and 6). Finally, in chapter 7, some first investigations into the problem of mapping fundus type eye images (sometimes used to more accurately define the tumour positions and shape) to a spherical eye model have been performed . For a clipless modeling of the eye, an eye model has been created based on three-dimensional information as can be extracted from MR imaging, with the eye globe, the lens and the tumor being contoured by hand. These structures were geometrically compared with the standard (clip guided) model, by calculating overlap and volume ratios between the models, and assessing the dosimetric consequences of moving to a MR based appraoch. Geometrical differences between the eye globes have been found to not affect the target coverage, whereas the tumour defined on MR was typically found to be smaller than that defined using the standard approach, for a small proportion, MR defined tumours were found to be larger. Using the standard treatment planning system for OPT (EyePlan), the consequences of these volume differences were tested by substituting the tumour models. For the EyePlan based plans, the larger target volumes were most conservative enough to also cover the MR defined tumors. On the other hand, for plans based on the MR defined tumor, even when the MR defined tumour was larger than the standard volume, this did not guarantee coverage of the standard tumor volume. To create a treatment plan which is based on the geometry of the MR model, and which is able to compare treatment plans, a new planning engine (ATP) for OPT has also been developed. The dose calculation was designed to be as similar as possible to the one of EyePlan which, due to its simplicity allows for the creation and comparison of multiple plans is a short time. In ATP a number of possible treatment plans is pre-calculated and presented to the user in a dedicated user-interface the quality of the plans in respect to the previously defined planning constraints. This user interface allows the user to navigate through the plans and adapt the constraints, allowing for a quantified comparison and selection of an optimal plan. ATP is a first step to an completely automated treatment planning system for OPT. For the task of patient navigation, an optical eye tracking system (ETS) was installed in the treatment room. This enabled the monitoring of the eye position and orientation in a clinical setting. Tests of a first generation tracker showed promising results by comparing its capability to determine the clip position with the x-rays. However, weak spots in the original design could be identified. As such, a second ETS prototype was designed, aiming at a non-invasive and automatic eye localization and patient navigation system. This was also integrated into the OPTIS treatment room at PSI. The visibility of reflections on the surface of the eye, the main drawback of the first prototype, could be successfully increased. Such ETS use the center of the cornea curvature (C-point) and the center of the pupil (P-point) as optical landmarks to determine the localization as well as the orientation of the eye. These two points function as surrogates however, and are not guaranteed to be similar to their anatomical equivalents. Due to the use of an MR model, the necessity of a patient specific correction regarding the optical landmarks became obvious. In terms of patient navigation by the ETS, the use of the MR based model was not significantly different from the EyePlan one, but a patient specific correction needed to be applied in each case. Despite the achieved improvement and progress, the ETS based navigation has still been found to be less accurate than the clip based approach, with clinically acceptable positioning accuracy only being achieved in about 88% of tested treatment fractions. As such, although, the ETS is a promising alternative regarding patient navigation, further developments are required to compete with the current navigation approach. As part of the MR model work, it became obvious that, in order to accurately define the tumor volume, other imaging information is required, one of the most important being fundus imaging. The fundus image is a photograph of the inner surface of the eye, visualizing the extent of the tumor as well as structures of the eye (macula and optic disc). In order to use this information however, the eye model must be unfolded into a two dimensional plane and registered with the fundus image, typically using the macula and optic disc as landmarks for position and scaling. This approach has two weak drawbacks however. Firstly, the two registered images result from different mathematical transformations. Secondly, in case of the unfolded eye model, the landmarks used for the registration are not patient specific, but follow a geometrical model. Therefore, an alternative approach has been investigated in this work, where the model is not unfolded, but a virtual photography of the eye model is instead created. For this, the required camera parameters are determined by simulating the optical properties of the fundus image, using the implanted clips as reference landmarks. Using a simple optical model (a pin-hole camera approach), successful registration could only be achieved in about 40% of the tested cases however, indicating that such an approach would require much more work before it could be reliably used in a clipless OPT procedure.