A Combination of Image Enhancement and U-Net Architecture for Segmentation in Identifying Brain Tumors on CT-SCAN Images

Abstract

A brain tumor is a condition where the abnormal growth of brain cells around the brain spreads to any part of the body uncontrollably. Brain tumors can be detected by recognizing tumor patterns in brain images by segmentation that separates brain tumor features in brain images. Sometimes the image data still has low quality, so image enhancement is needed to improve image quality. This paper proposed to improve the quality of a 2D brain tumor image and the performance of segmentation. Image enhancement consists of two stages, namely contrast enhancement and noise reduction. The contrast enhancement used is Contrast Limited Adaptive Histogram Equalization (CLAHE) and gamma correction, while the noise reduction used is a Gaussian filter. Parameters in measuring the results of the quality of the image are Mean Square Error (MSE) and Structural Similarity Index Measure (SSIM). U-Net architecture is applied in the brain tumor segmentation process. Parameters in evaluating the performance of the segmentation model are accuracy, sensitivity, specificity, and mean Intersection over Union (IoU). The results of brain tumor image improvement obtained are MSE of 0.018 and SSIM of 0.9748. The results of brain tumor segmentation obtained were 98% accuracy, 77.85% sensitivity, 99.12% specificity, and 82.31 % mean IoU. Sensitivity below 80% is caused by too small a portion of the tumor. Even though the sensitivity was below 80%, the model was able to separate the tumor from the background.