Effect of noise on the robustness of MobileNetV2+U-Net semantic segmentation model for MRI images

Abstract

This study aims to evaluate effect of noise on the robustness of semantic segmentation models for Magnetic Resonance Imaging (MRI) head images with tumor. We implemented the MobileNetV2+U-Net architectural model. We tested the segmentation model with Gaussian and Poisson noises in various levels. The addition of noise was performed five iterations with a variance of 0.01 each iteration. We carried out evaluations by examining the segmentation results, loss function values, accuracy and dice score. Based on the results, increase in noise affects model performance. Evaluation using loss function shows that graph instability is influenced by the noise level. The accuracy results on the highest and lowest validation data were 99.47% and 98.99% for Gaussian noise and 99.64% and 99.5% for Poisson noise. Apart from that, the highest and lowest dice scores were 82.80% and 69.18% for Gaussian noise and 87.83% and 83.10% for Poisson noise. We recommend training the segmentation model using noisy data so that the model can adapt to noisy images.