Benchmarking of Shallow Learning and Deep Learning Techniques with Transfer Learning for Neurodegenerative Disease Assessment Through Handwriting

Abstract

Neurodegenerative diseases are incurable diseases where a timely diagnosis plays a key role. For this reason, various techniques of computer aided diagnosis (CAD) have been proposed. In particular handwriting is a well-established diagnosis technique. For this reason, an analysis of state-of-the-art technologies, compared to those which historically proved to be effective for diagnosis, remains of primary importance. In this paper a benchmark between shallow learning techniques and deep neural network techniques with transfer learning are provided: their performance is compared to that of classical methods in order to quantitatively estimate the possibility of performing advanced assessment of neurodegenerative disease through both offline and online handwriting. Moreover, a further analysis of their performance on the subset of a new dataset, which makes use of standardized handwriting tasks, is provided to determine the impact of the various benchmarked techniques and draw new research directions.