Brain Lesion Detection: An Architecture Based On Convolution Neural Network

Abstract

MRI based lesion detection is an essential step for computer aided diagnosis to detect brain
lesion. There are several types of tissues, skull portion in brain MRI images, which lead to false
detection of brain lesion. To get appropriate lesion detection segmentation architecture based on
Convolution Neural Network (CNN) is proposed in the study. Here, 2450 3D RGB images are taken
from The Cancer Genome Atlas (TCGA), where 2400 images are used from 104 patients for training
purpose and 50 images are taken from six patients for testing purpose. From the test images the
architecture achieved accuracy (99.1%), dice similarity (85.4%), jaccard index (74.8%), Mathews
correlation coefficient (82.2%), sensitivity (89.7%), specificity (99.3%), and precision (82.3%) with
95% confidence interval. Besides, the receiver operating characteristic (ROC) curve also plotted with
highest 97.22% area, which proves the constancy and reliability of the architecture. The proposed CNN
based architecture detects the accurate lesion from brain MRI 3D RGB images.