Estimación de la gravedad del Alzheimer según síntomas, usando Machine Learning en imágenes de Resonancia Magnética (MRI).

Abstract

The main objective of this research was to develop an interactive web tool to estimate the severity of Alzheimer's disease, using Machine Learning in MRI corresponding to four groups: mild demented, moderate demented, non demented, very mild demented. The images of the dataset were preprocessed, two methods were used: zero padding to resize the images without modifying them and the correction of bias field. The initial approach was to compare different Machine Learning models, the models we trained were: CNN, VGG 19, Inception V3, MobileNet V3 and EfficientNet. For the training of these models, we used the Python language and the TensorFlow library; as more powerful resources were required, we used Google Colab. For the web page we use HTML5 to design it, JavaScript for some features and CSS to define the styles. Regarding the results, some models had overfitting and others had problems in the behavior of the metrics, the models with the best results were CNN with data augmentation, VGG 19 and Inception V3 with an accuracy of 97.9%, 99.8% and 90% respectively. With the development of this work, it was concluded that using transfer learning doesn’t guarantee that a good result will be obtained, but this depends on the dataset that we are using to train our models. Additionally, it is concluded that when training Machine Learning models overfitting can occur, but it is possible to solve it with techniques that can contribute to obtain a better performance.