Improving Segmentation of Pilocytic Astrocytoma in MRI Using Genomic Cluster-Shape Feature Analysis

Pilocytic astrocytoma is a type of tumor related lower-grade glioma (LGG). This brain tumor is more difficult to detect, and treat compared to higher grade gliomas. To understand the complexity of LGGs and related disease of epilepsy and cancer, this study proposes genomic cluster-shape feature selection, feature extraction and segmentation methods. The MRI images provides genomic information of DNA sequencing and gene expression by heat map and feature selected by correlation coefficient of p>0.9. The support vector classifier (SVC) and gaussian kernel (GK) used to quantify the linear relation between two variables, and to measure the similarity between two datapoints with high-dimensional space. The twenty-three relevant features are selected by Random Forest classifier and compared with univariate, recursive feature elimination, and principal component analysis method. The semantic segmentation by UNet has encoder captures context information and decoder enables the precise localisation of the object. The ResNext50 incorporates a cordiality parameter, which used to capture the fine-grained features. The UNet with ResNext50 backbone enhance the performance matrix. The calculated metrics of SVC with GK of selected features (90.05%) were higher than without selected features (63.63%). The feature extraction process by random forest classifier with univariate analysis (85.1%) and recursive feature elimination method (85.71%), and with cross-validation achieving an accuracy of 95.2%. The cross-validation (CV) is used to validate the features, with each combination of k folds being multiplied with different batch sizes and numbers of epochs (80.7%, 90.1%, 92.7%, 96.7%, and 95.6%). The segmentation dice score of UNet (72.13%) and UNet with ResNext50 backbone (89.7%) were used to compare the performance of these features. This study used a dataset of LGG patients and found that their improved segmentation accuracy by up to 9.7% compared to earlier analysis of UNet with other residual network and gives the valuable insight of features associated with tumors and reduce the complexity of treatment.