The Influence of Magnetic Resonance Imaging Artifacts on CNN-Based Brain Cancer Detection Algorithms

• Published in: Computational mathematics and modeling 2022-04, Vol.33 (2), p.211-229
• Main Authors: Farias, M. C. Q., de Castro Oliveira, P. H., dos Santos Lopes, G., Miosso, C. J., Lima, J. A.
• Format: Article
• Language: English
• Subjects: Algorithms; Applications of Mathematics; Brain; Brain cancer; Cancer; Computational Mathematics and Numerical Analysis; Decision making; Image degradation; Image quality; Image reconstruction; Machine learning; Magnetic resonance imaging; Mathematical Modeling and Industrial Mathematics; Mathematics; Mathematics and Statistics; Medical imaging; Optimization; Performance evaluation; Random noise; Training; Tumors
• ISSN: 1046-283X; 1573-837X
• DOI: 10.1007/s10598-023-09567-4

Early detection of cancer tumors significantly increases the chances of recovery and usually results in improved quality of life and patient lifespan. In this context, computer systems can help and accelerate health care decision making regarding brain diagnoses, based, for example, on machine learning (ML) algorithms applied to neurological magnetic resonance images (MRIs). In the last decade, several ML approaches have been tested for this purpose, with varied levels of accuracy, precision, and sensitivity. It is already clear that this performance depends on the type of tumor, the quality of the image, the number of training images, and, in the case of supervised training, the quality of previously diagnosed images used for training. An important question remains regarding how the performance of these diagnostic tools is affected by image quality, as it is influenced by several factors common to MRIs, such as artifact like ringing effects, Gaussian noise, reconstruction artifacts associated with lower numbers of measurements, movement artifacts, and so forth. In this paper, we conduct a systematic analysis of this relationship, based on a state-of-art ML diagnosis system, but under different levels of image degradation. Our experiments suggest that severe degradations, leading to reduced mean squared error (MSE) values, can systematically reduce the values of all performance evaluated metrics, but whether such reduction occurs for a specified MSE depends on the type of effect. In addition, for all artifacts tested, we evaluated the range of MSE for which the performance reduction is below a specified threshold. These results can be used in the future to indicate the level of confidence in a certain automatic diagnostic result for different image qualities.