TY - GEN
T1 - Detecting Cognitive Decline Using a Novel Doodle-Based Neural Network
AU - Pearson, Connor
AU - De La Iglesia, Beatriz
AU - Sami, Saber
N1 - Publisher Copyright:
© 2022 IEEE.
PY - 2022/12/5
Y1 - 2022/12/5
N2 - A key part in the diagnosis of cognitive decline are visuospatial based tests. These visuospatial tests often involve a form of drawing task. In this paper, we build an automated multiclass classifier to assign hand-drawn doodles from Google’s online game Quick, Draw! into 24 unique categories that are simple to draw doodles. Our goal is to create a prototype of an automated online diagnosis tool that resembles the visuospatial portion of established pen and paper cognitive examinations. We built a CNN using the Tensor Flow Keras API, and tested multiple iterations of each model neuron structure. We created a web interface able to capture user inputs from a browser window as they draw the requested doodle for each test stage. The images are relayed back to a server and processed through the same model trained on the Google QuickDraw! dataset to determine a patient’s score. Herein we use these model predictions as a measurement of the users drawing skills. Using a CNN based neural network we achieved a 90.46% model accuracy and around 70% implementation accuracy which is not dissimilar to human pen and paper ratings.
AB - A key part in the diagnosis of cognitive decline are visuospatial based tests. These visuospatial tests often involve a form of drawing task. In this paper, we build an automated multiclass classifier to assign hand-drawn doodles from Google’s online game Quick, Draw! into 24 unique categories that are simple to draw doodles. Our goal is to create a prototype of an automated online diagnosis tool that resembles the visuospatial portion of established pen and paper cognitive examinations. We built a CNN using the Tensor Flow Keras API, and tested multiple iterations of each model neuron structure. We created a web interface able to capture user inputs from a browser window as they draw the requested doodle for each test stage. The images are relayed back to a server and processed through the same model trained on the Google QuickDraw! dataset to determine a patient’s score. Herein we use these model predictions as a measurement of the users drawing skills. Using a CNN based neural network we achieved a 90.46% model accuracy and around 70% implementation accuracy which is not dissimilar to human pen and paper ratings.
UR - http://www.scopus.com/inward/record.url?scp=85144621284&partnerID=8YFLogxK
U2 - 10.1109/MetroXRAINE54828.2022.9967549
DO - 10.1109/MetroXRAINE54828.2022.9967549
M3 - Conference contribution
T3 - 2022 IEEE International Workshop on Metrology for Extended Reality, Artificial Intelligence and Neural Engineering, MetroXRAINE 2022 - Proceedings
SP - 99
EP - 103
BT - 2022 IEEE International Workshop on Metrology for Extended Reality, Artificial Intelligence and Neural Engineering, MetroXRAINE 2022 - Proceedings
PB - The Institute of Electrical and Electronics Engineers (IEEE)
T2 - 2022 IEEE International Workshop on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering
Y2 - 26 October 2022 through 28 October 2022
ER -