TY - GEN
T1 - Automating the Clock Drawing Test with Deep Learning and Saliency Maps
AU - Mayne, Violet
AU - Rogers, Harry
AU - Sami, Saber
AU - de la Iglesia, Beatriz
PY - 2025
Y1 - 2025
N2 - The Clock Drawing Test (CDT) is an important tool in the diagnosis of Cognitive Decline (CD). Using Deep Learning (DL), this test can be automated with a high degree of accuracy, more so where the medium of recording allows the use of temporal information on how the clock was drawn which may not be accessible to clinicians in traditional screening. The high-risk nature of this field makes understanding the reasoning for automated results imperative. A model’s reasoning can often be described using saliency maps, however, there are a number of different methods for generating such maps. Therefore, we propose a methodology to train a DL classifier for use in the CDT which incorporates temporal information and use saliency maps to explain classification predictions. We find that our classifier achieves scores above 98% with F1 for clocks and over 96% F1 on average across a test set of 18 different classes. Our methodology also shows that Integrated Gradients using SmoothGrad produce the best saliency map results visually and statistically.
AB - The Clock Drawing Test (CDT) is an important tool in the diagnosis of Cognitive Decline (CD). Using Deep Learning (DL), this test can be automated with a high degree of accuracy, more so where the medium of recording allows the use of temporal information on how the clock was drawn which may not be accessible to clinicians in traditional screening. The high-risk nature of this field makes understanding the reasoning for automated results imperative. A model’s reasoning can often be described using saliency maps, however, there are a number of different methods for generating such maps. Therefore, we propose a methodology to train a DL classifier for use in the CDT which incorporates temporal information and use saliency maps to explain classification predictions. We find that our classifier achieves scores above 98% with F1 for clocks and over 96% F1 on average across a test set of 18 different classes. Our methodology also shows that Integrated Gradients using SmoothGrad produce the best saliency map results visually and statistically.
KW - Cognitive Decline
KW - Deep Learning
KW - eXplainable Artificial Intelligence
UR - http://www.scopus.com/inward/record.url?scp=85210243628&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-73500-4_8
DO - 10.1007/978-3-031-73500-4_8
M3 - Conference contribution
SN - 978-3-031-73499-1
T3 - Lecture Notes in Computer Science
SP - 86
EP - 97
BT - Progress in Artificial Intelligence
A2 - Santos, Manuel Filipe
A2 - Machado, José
A2 - Novais, Paulo
A2 - Cortez, Paulo
A2 - Moreira, Pedro Miguel
PB - Springer
ER -