Semi-automatic stereotactic radiosurgery with neural network-based multi-modal image segmentation and geometric analysis of the lesion
Mehmed Özkan · 1992
Stereotactic Radio Surgery (SRS) has been employed in the treatment of various brain tumors since 1968. Recently, Linear Accelerator-based radiosurgical systems have been developed with good mechanical accuracy and better treatment planning capabilities due to advances in computer technology. It is now possible to treat lesions located deep within the brain while sparing surrounding, healthy tissue. Computerized Photon Knife (CPK) is a linear accelerator coupled to a localization subsystem. Collimators of various sizes are used to direct photon beams precisely. This system's five degrees of freedom in spatial coordinates makes it possible to tailor a lethal dose of radiation to irregularly shaped tumors. Flexibility of the computer software allows more accurately sculpted treatment plans than were previously possible with the Gamma Unit. The drawback of having such flexibility, however, is the multiplicity of options to consider in a relatively short time. The number of planning parameters are too varied. Therefore, a less optimum plan may be acceptable. However, the goal is to find the perfect plan in a short time. This dissertation studied two aspects of the stereotactic radiosurgery process; lesion localization and CPK treatment planning. Software was written in an attempt to speed up the overall treatment process and produce more efficient plans. Lesion localization was partly automated using artificial neural networks on multi-modal tomographic brain images. The application of the technique is not limitted to stereotactic radiosurgery, but can be utilized in many other medical applications that depend on quantitative tissue information. Planning was performed by geometrical analysis of the segmented lesion. Computed Tomography (CT), Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) are used for quantization of images to provide multi-dimensional information for each pixel. A statistical pattern recognition technique, the Bayes maximum likelihood classifier, and artificial neural networks (ANN) are tested and compared as pattern recognition techniques for each of the image modalities. The contribution of each modality is evaluated. The partitioning of the parameter space is studied for various ANN architectures. An adaptive learning scheme is proposed to overcome intensity inhomogeneities introduced by the imaging systems. Segmentation results are compared with those obtained from medical experts for stereotactic radiosurgery purposes. Finally, automatic planning software is tested on 40 lesions that were treated earlier using conventional treatment planning methods.