Image-Based Dysgraphia Detection with Explainability Using CAMs and EDI

View at publisher Summarise with VIRAT
Authors: Varsha Sreekumar, Aadhavan P Anbumunee, Sudha Senthilkumar
Cited by: 0
References: 14

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Dysgraphia is a learning disability that affects people’s handwriting quality and can lead to poor academic outcomes if not diagnosed earlier. Simple diagnosis methods like manual observation can be subjective. You can use specialized tablets for detection based on handwriting, but is not scalable. We propose an explainable image-based detection framework using scanned notebook pages for accessible, low-cost screening in educational and clinical settings. It involves combining Deep Learning with a novel Explainable Dysgraphia Index (EDI) that quantifies handwriting traits weighted by Score- CAM (Class Activation Mapping) activations, thus helping justify classification using clinically meaningful features. The framework uses ResNet-18 trained on the Mendeley Potential Dysgraphia dataset with writer-independent splits for generalization. Score- CAM creates visual heatmaps that help us understand which handwriting regions contributed to decision making. EDI on the other hand combines normalized features with CAM attention for numeric and visual explanations. The CNN (Convolutional Neural Network) and EDI fusion achieved 92% accuracy and 0.98 AUC (Area Under Curve), proving that explainable features improve performance while providing clinicians with interpretable insights for practical adoption.

Fields of Study & Subjects

Handwritten Text Recognition Techniques Writing and Handwriting Education Explainable Artificial Intelligence (XAI)

Publication details

Journal / source
International Conference on Computing, Communication, Control and Cyber-Physical Systems (I5CPS)
Pages
1–6
Publication date
26 February 2026
Language
EN