Automatic dementia screening and scoring by applying deep learning on clock-drawing tests.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Dementia is a prevalent neurological syndrome globally.
• Current dementia diagnosis relies on subjective paper-based tests (clock-drawing test), leading to errors and inter-rater variability.
• Need for objective, standardized, and accessible diagnostic tools.

Purpose of the Study:

• To develop and evaluate an automated system for assessing the clock-drawing test using deep neural networks.
• To compare the performance of VGG16, ResNet-152, and DenseNet-121 architectures for this task.
• To provide a standardized, digital estimation of dementia screening results and severity.

Main Methods:

• Utilized deep neural networks (VGG16, ResNet-152, DenseNet-121) for automatic analysis of clock-drawing tests.
• Trained models on a dataset of 1315 individuals, employing optimization strategies to handle data limitations and diverse dementia types.
• Compared deep learning model performance against traditional scoring and human expert judgment.

Main Results:

• Achieved high accuracy rates: 96.65% for dementia screening and up to 98.54% for scoring severity.
• The automated system surpassed the performance of existing state-of-the-art methods and human accuracy.
• Demonstrated the feasibility of using mobile devices for scanning and digital evaluation of the clock-drawing test.

Conclusions:

• Deep neural networks offer a reliable and accurate method for automating clock-drawing test analysis.
• This digital approach enhances diagnostic objectivity and standardization in dementia screening.
• The technology can extend diagnostic capabilities to remote or resource-limited settings, addressing staff shortages and expert unavailability.