Developing an Image-Based Deep Learning Framework for Automatic Scoring of the Pentagon Drawing Test.

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• The Pentagon Drawing Test (PDT) is a standard tool for assessing visuospatial function.
• Integrating artificial intelligence (AI) with PDT evaluation can enhance efficiency and reliability, particularly in the context of big data.
• This study focuses on developing a deep learning (DL) framework for automated PDT scoring using image data.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) framework for the automatic scoring of the Pentagon Drawing Test (PDT).
• To compare two distinct DL strategies for PDT image analysis: supervised transfer learning and object detection.

Main Methods:

• Retrospective collection and preprocessing of 823 PDT images into standardized black-and-white, square-shaped formats.
• Implementation of stratified fivefold cross-validation for robust model training and testing.
• Comparison of two convolutional neural network (CNN) strategies: image classification via transfer learning and object detection for shape recognition and positional analysis.

Main Results:

• The image classification (transfer learning) framework achieved moderate performance: 62% accuracy, 62% recall, 65% precision, 63% specificity, and a 0.72 ROC AUC.
• The object detection framework significantly outperformed the classification approach, yielding high averages of 94% accuracy, 95% recall, 93% precision, 93% specificity, and a 0.95 ROC AUC.
• The object detection model demonstrated superior efficacy in automatically scoring the PDT.

Conclusions:

• An image-based DL framework utilizing object detection shows significant clinical potential for efficient and reliable automatic PDT scoring.
• Transfer learning requires cautious application with limited or dissimilar datasets.
• Decomposing complex tasks into simpler, manageable steps aids DL model selection, enhances performance, and clarifies framework logic.