Pitfalls Related To Computer-Aided Diagnosis System Learned From Multiple Databases

• Published in: Proceedings (International Symposium on Biomedical Imaging) pp. 806 - 809
• Main Authors: Touvron, Hugo, Faisan, Sylvain, Tilquin, Florian, Noblet, Vincent
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.04.2019
• Subjects: classification and regression; Computer-aided diagnosis; Magnetic resonance imaging; nuisance variables; Shape; Sociology; Solid modeling; Statistics; Testing
• ISSN: 1945-8452
• DOI: 10.1109/ISBI.2019.8759550

The growing availability of large neuroimaging databases offers exceptional opportunities to train more and more efficient machine learning algorithms. Nevertheless, these databases may be prone to several sources of variability (age, gender, acquisition parameters,...). These nuisance variables can hamper the performance of a classification method and can even lead to misinterpret its behavior. We focus in this paper on how to account for data coming from different databases. First, we present experiments on simulated data that illustrate how interactions with other confounds such as age can be problematic for the adjustment of data from multiple databases. Then, we compare three standard strategies to adjust data and evaluate them in the scenario of a Computer-Aided Diagnosis system that discriminates healthy from Alzheimer's Disease subjects based on volumetric characteristics derived from MRI. We highlight that classifiers with apparently similar performance do not all rely on relevant information depending on the correction strategy.