Impact of magnetic resonance imaging quality on deep neural network

In the last decades, tasks of classification and segmentation of clinical findings using convolutional neural network shave grown significantly in the sphere of radiology, and more precisely in the modality of magnetic resonance imaging. However, little is known about the behavior of the proposed architectures when faced with factors that degrade the spatial resolution and contrast resolution, since most models are trained with high quality images, which is not consistent with the general daily life. Therefore, it is necessary to analyze the performance of pre-trained neural networks under conditions in which there is deterioration of the input image. In this work, the effects of degradation of both resolutions for three architectures were evaluated – Mobile net, Vgg16 and SEResNeXt50 –, both in classification and segmentation tasks of brain tumors. The results obtained showed that the tasks performed are greatly affected by image quality distortions, especially in cases where the deteriorations become more intense. (author)