Delineation and diagnosis of brain tumors from post contrast T1-weighted MR images using rough granular computing and random forest

• Published in: Applied soft computing Vol. 41; pp. 453 - 465
• Main Authors: Koley, Subhranil, Sadhu, Anup K., Mitra, Pabitra, Chakraborty, Basabi, Chakraborty, Chandan
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.04.2016
• Subjects: Brain tumor; Feature extraction; MRI; Rough entropy; Tumor characterization; Brain tumor; Feature extraction; Rough entropy; MRI; Tumor characterization
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2016.01.022

•A new approach of tumor area delineation from T1C brain MR image.•A complete framework of detection and characterization of brain tumors.•The diagnosis can be performed from single slice of T1C MRI. This paper presents a new approach of delineation and characterization of four different types of brain tumors viz. Glioblastoma multiforme (GBM), metastasis (MET), meningioma (MG) and granuloma (GN) from magnetic resonance imaging (MRI) slices of post contrast T1-weighted (T1C) sequence to improve the computer assistive diagnostic accuracy. An integrated framework of identification and extraction of tumor region, quantification of histogram, shape and textural features followed through pattern classification by machine learning algorithm has been proposed. Rough entropy based thresholding in granular computing paradigm has been adopted for delineation of tumor area. After accomplishing quantitative validation and comparison with existing methods, experimental results prove the efficiency and applicability of proposed segmentation approach. In the next stage, the extracted lesions have been quantified with 86 features to develop the training dataset. Random forest (RF), an ensemble learning scheme has been implemented, which learns the training data for accurate prediction of the class label of a given input. The performance of RF has been evaluated by statistical measures from 3 fold cross-validation and compared with five different classifiers. The same experiment has been repeated over the reduced set of features generated by correlation based feature selection strategy. Experimental results show the superiority of RF (Sensitivity achieved in %: GBM-96.7, MET-96.2, MG-98.1 and GN-97.7) with the complete set of features. The comparison of proposed methodology with the existing works signifies its applicability and effectiveness. Additionally a 10 fold cross-validation has been accomplished to justify the statistical significance of the classification accuracy achieved from proposed methodology.