Segmentation of central nervous system tumor images with neural networks
Abstract
Tumor image segmentation represents a crucial step not only in the diagnosis of the disease but also in its evaluation and monitoring of treatment. In the present work, neural networks are used to recognize the presence of tumoral tissue in central nervous system on magnetic resonance images generated by in vivo spectroscopy and relaxometry. Relaxation data was validated and categorized by means of spectroscopic data, used as a sort of virtual biopsy. Neural networks were trained with the relaxation data in a supervised mode, assuming three categories for the tissue: tumoral, normal or unaffected and liquid or necrosis. Segmentation performed in this way correlates closely to other methodologies previously developed, shortening drastically the processing time what makes it very useful in its clinical application.
