Multi-resolution Tranaform Based Feature Extraction Techniques for Differentiating Glioma Grades Using MRI Images

Abstract

Medical image processing is one of the most attention gaining research areas that utilizes the technology for improving the quality of human life through a more precise and rapid diagnosis systems. This thesis focuses on computer assisted diagnosis of brain neoplasms which is amongst the most fatal cancers. Though, their exact cause is still unknown but early detection anddiagnosisofcorrectneoplasmtypeisveryimportantforpatient’slifeandfurthertreatment planning. Currently, the treatment of brain neoplasm depends on clinically observed symptoms, appearance of radiological tests, and often the microscopic examination of neoplasm’s tissues (histopathology or biopsy report). Magnetic Resonance Imaging (MRI) is the state of art technique to diagnose brain neoplasms and monitor their treatment. It provides a noninvasivewaytoimprovethequalityofthepatient’slifethroughamoreaccurateandfastdiagnosis and with minor side-effects, leading to an effective overall treatment. However, MRI does not provide any information about exact type and grade of neoplasm. The final decision is based on biopsy report of patient which is considered as gold standard, despite all risks associated with surgery to obtain a biopsy. With rapid advancement in technology, the researchers are continuously working on computerized techniques or computer assisted diagnostic tools to provide fast identification, correct diagnosis and effective treatment of brain neoplasm. The aimofthepresentthesisistodesign,implement,andevaluateasoftwareclassificationsystem fordiscriminatingthreegradesofbrainneoplasmonMRI.Limitedbrainneoplasmimagedata isoneofthebiggestissuesinthisresearchareabecausecollectionofthistypeofdatarequires years and years. Normally, we find studies working on images of some specific hospital orwebsite. Inaddition,directcomparisonofthesestudiesisnotpossiblebecauseeachstudyhad worked on different types of neoplasm and various sizes of image data. We have addressed this issue by proposing a new image cropping technique for handling images of different dimension for the same classifier. This new system is capable of handling image datasets from different institutions with various image sizes and resolutions for comparing, regulating and sharing of research. It is also observed, that lesser training and testing images in a particular class of neoplasm badly effect the classification accuracy. By using this generalized system, moreimagesamplesofaneoplasmclasscanbetakenfromotherinstitutionsorwebsitestoimprovetheclassificationaccuracy. ForclassificationofMRIimages,majorityoftheresearchers haveworkedonstatisticalfeaturesofneoplasmregionbutmulti-resolutiontransformsforfeature extraction, are not much explored. Besides this, classification of normal and pathological brain is mostly addressed but very few studies are found on multi-classification of different neoplasm types. The main objective of this thesis is to explore the performance of different multi-resolution transform based feature extraction techniques for multi-classification problem of brain neoplasm type (grade II, grade III and grade IV gliomas). Discrete Wavelet Transform (DWT) is one of the most popular multi resolution transform, extensively used as feature extraction technique for binary (normal vs abnormal brains) brain neoplasm classification systems. In this thesis, a stationary and time invariant Non Subsampled Contourlet Transform (NSCT) with Gray Level Co-occurrence Matrix (GLCM) is used for computation of feature vector in brain neoplasm classification system. This NSCT-GLCM based classification system is also compared with conventional DWT-GLCM based classification system, for the same experimental setup. It is found that NSCT-GLCM based system perform better than DWT-GLCM based system. For further improvement in neoplasm discrimination accuracy, in last algorithm, a multi resolution transform based hybrid feature extraction technique is introduced. This hybrid technique is comprised of conventional DWT, NSCT and GLCM. The quantitative performance analysis showed that hybrid feature extraction technique per formed much better than the previous two techniques (DWT-GLCM and NSCT-GLCM) with the highest accuracy of 88.88%. The developed brain neoplasm classification techniques can better assist the physician’s ability to classify and analyze pathologies leading for a more reliable diagnosis and treatment of disease.