A Study on the Effect of Noise in Inferential Process of Deep Learning for Medical Image
Mohammad Haekal (a*), Freddy Haryanto (b), Idam Arif (b)
a) Biophysics Laboratory, Department of Physics, Faculty of Mathematics and Natural Sciences, Institut Teknologi Bandung, Jalan Ganesha 10, Bandung 40132, Indonesia
*m.haekal.idn[at]gmail.com
b) Nuclear Physics and Biophysics Research Division, Department of Physics, Faculty of Mathematics and Natural Sciences, Institut Teknologi Bandung, Jalan Ganesha 10, Bandung 40132, Indonesia
Abstract
Noise is an inherent property of medical imaging and could affect the interpretation of medical images either by human or computer. In deep learning study for medical images, the noise is usually introduced during the training phase of the algorithm in order to make the system able to distinguish the image features in the inferential phase despite the existence of noise. Early hypothesis stated that without the introduction of noise in the training phase, the deep learning system could not distinguish the feature well enough which resulted in lower accuracy in the segmentation. However, due to the classification of the feature that the computer “understands” is unknown in deep learning, this study aimed to evaluate whether the inferential phase of image with noise could still be performed without the introduction of noise in the training phase. The noise will be added artificially in the original test data which would be inferred by the system based on the training of normal (without noise) image. The results would then be compared by using Dice’s Similarity Coefficient (DSC) between the test data with and without noise.
Keywords: medical image segmentation, deep learning, automatic segmentation, noise analysis
Topic: Biophysics and Medical Physics