Brain Tumor Classification Using Hyperspectral Image Analysis

Abstract:

The capability of Hyperspectral Imaging (HSI) in rapidly acquiring abundant reflectance data in a non-invasive manner, makes it an ideal tool for obtaining diagnostic information about tissue pathology. Identifying features that provide the most discriminatory clues for specific pathologies will greatly assist in understanding their underlying biochemical characteristics. In HSI, reflectance data is acquired over several wavelengths and identifying the most relevant wavelengths appropriate for a specific application is critical. This not only reduces the dimensionality of the data, but also provides insights onto the biochemical composition of tissues based on their reflectance responses within those relevant wavelengths. In this thesis, brain cancer HSI data was analyzed to extract simple and computationally efficient morphological features from the pixel spectra. In addition, an efficient and computationally inexpensive method of determining the most relevant spectral bands for brain tumor classification was proposed. Empirical mode decomposition was used in combination with extrema analysis to extract the relevant wavelengths based on the morphological characteristics of the spectra.
 
Using a database of 26 brain cancer HSI images and simple morphological features with a SVM classifier, a maximum pixel classification accuracy of 87.9%  and 78.6% was achieved for binary (tumor vs normal) and multigroup pixel classifications (tumor, normal, hypervascular, and background) respectively. The proposed band selection approach identified 12 discriminatory bands (which is 3 times less than the benchmark work) and whose spectral reflectance values achieved a maximum pixel classification accuracy of 99.6% and 94.7% for binary and multigroup pixel classifications respectively using a SVM classifier.

(Anyone can attend this zoom session.  Please send an email to Dr. K. Umapathy requesting the zoom URL and passcode)