Abstract
Purpose: :
A challenging topic in ophthalmology is the treatment of intraocular tumors by irradiation. Choroidal melanoma and other tumors are commonly treated by charged particle beam irradiation or brachytherapy. For charged particle beams, it is possible to shape the irradiation volume in a very precise and sharply delineated way. This allows for an optimized tumor control rate, at the same time sparing as much as possible eye structures at risk. In order to take advantage of charged particle radiotherapy a precise treatment planning is crucial. State of the art treatment planning is based on a parametric eye model, which is adapted to the patient’s anatomy based on Euclidean distances, extracted from different image modalities e.g. CT, MRT, US, etc. This approach is highly limited in accuracy and strongly depends on the surgeon. In this study we evaluate the feasibility of a Statistical Shape Model as a method for modeling the human eye to improve treatment planning accuracy. An increase in planning accuracy will lead to higher treatment efficacy and safety for the patient. At a later stage we plan to fuse the different image modalities into a single patient specific eye model to enable multi-modal treatment planning.
Methods: :
Statistical Shape Models (SSM) have been shown to be a powerful tool to capture the large, but limited variability of organ shape in a compact manner. In this study a 3D SSM of the human eye including Cornea, Lens, Sclera as well as the position of the Optic Nerve Head was built. 18 manual segmented head CT scans with a resolution of 0.39 x 0.39 x 0.6 mm has been used as training set for the SSM construction as well as for validation purposes. In a first step an atlas is constructed representing the mean eye shape. From this atlas, landmarks are extracted in a second step and propagated to each shape in the training set via a volumetric non-rigid registration technique. This approach enables the construction of a SSM in presence of large shape variability and multiple objects. In order to automatically fit the SSM into patient specific CT slice stacks an Active Shape Model (ASM) has been built.
Results: :
To measure the performance of the SSM and ASM a Leave-One-Out cross validation on the training set was carried out and similarity with the manual segmentation done by an expert was measured by the Dice Similarity Measure (DSM). The cross validation revealed a Dice coefficient of 95.2±2.2% for the Sclera and Cornea and 91.4±2.1% for the Lens. An average model fit took 30s on a regular desktop PC.
Conclusions: :
The presented method to model a patient’s specific eye has proven to be highly accurate. Furthermore it has been shown that this model-based method can be used to segment eye structures on CT scans with a high amount of robustness and with a short execution time.
Keywords: radiation therapy • tumors • image processing