Automatic Diagnosis of Mild Cognitive Impairment Using Siamese Neural Networks

The use of Artificial Intelligence techniques as an aid tool in the medical field is a current and undeniable challenge. In this context, similarity detection methods and Convolutional Neural Networks used in computer vision tasks for feature extraction can greatly contribute to the analysis of medical tests based on freehand drawings. This paper brings together both ideas and proposes the use of Siamese Neural Networks to perform an automatic diagnosis of mild cognitive impairment (MCI) based on the Rey-Osterrieth Complex Figure (ROCF) test. It analyzes the suitability of this type of networks and compares them with an ANN. For this purpose, about 477 drawings collected in a research study in the field of neuropsychology, made by healthy patients or patients with some degree of cognitive impairment, are available. Due to the small number of instances, it is proposed to pre-train the networks with the Transfer Learning technique using a much larger dataset of drawings with similar characteristics.