Ensembling shallow siamese architectures to assess functional asymmetry in Alzheimer’s disease progression

Abstract

The development of methods based on artificial intelligence for the classification of medical imaging is widespread. Given the high dimensionality of this type of images, it is imperative to use the information contained in relevant regions for further classification. This information can be derived from the morphology of the region of interest, in terms of measurements such as area, perimeter, etc. However, the performance of the classification system strongly depends on the correct selection of the type of information employed. We propose in this work an alternative for evaluating differences between brain regions that relies on the basis of Siamese neural networks. Initially, brain scans are delimited by an anatomical atlas. Next, each pair of regions of interest is then entered into a Siamese network, which is formed by relating the distance between the two individual outputs and the corresponding label. Features are extracted from the embeddings of the final linear layer. Finally, the classification is performed by combining the characteristics of each pair of regions into an ensemble architecture. Performance was assessed by determining how asymmetry between the right and left hemispheres changes during progressive brain degeneration, from mild cognitive impairment to severe atrophy associated with Alzheimer’s disease (AD). Our method discriminates with an accuracy of 98.95% between controls and AD patients, and most important, it predicts the cognitive decline in patients suffering from mild cognitive impairment that will develop AD before it occurs with an accuracy of 78.41%. These results demonstrate the applicability of our proposal in the study of a wide range of pathologies.