Detection and Classification of Brain Tumours from MR Images Using Texture Features and Machine Learning

Abstract

Brain tumour is among the more lethal forms of cancer, and its position close to the neuronal centre of the body means 
that even a small lesion can have serious consequences. Early detection substantially raises the chance of successful 
treatment, but expert evaluation of every patient is costly and impractical at scale, which is what motivates computer
aided detection. This paper describes a system for detecting and classifying brain tumours from magnetic resonance 
images. The pipeline comprises four stages. Pre-processing with a Gaussian low-pass filter removes low-frequency 
background noise and normalizes intensity. Segmentation using Otsu thresholding partitions the image by selecting 
the grey-level threshold that minimizes intra-class variance, separating the region of interest from the background; 
because it operates directly on the grey-level histogram it is fast and reasonably noise-tolerant. Post-processing 
applies morphological erosion and dilation to remove the imperfections that simple thresholding leaves in the binary 
regions. Feature extraction computes fourteen Haralick texture descriptors from the grey-level co-occurrence matrix, 
including contrast, correlation, energy, entropy and homogeneity, using a formulation that is asymptotically invariant 
to the quantization step. Classification into benign and malignant is performed using K-nearest neighbours and 
support vector machine classifiers on a database of 180 brain MR images, of which 144 are benign and the remainder 
malignant.