EduMirror-XAI: An Explainable and Leakage-Resistant Deep Learning Framework for Detecting Handwritten Letter Reversals in Special Education

Abstract

Handwritten letter reversals may occur during literacy development and may persist among learners with reading and writing difficulties; however, they cannot independently establish dyslexia. This study proposes EduMirror-XAI, a lightweight and duplicate-aware deep-learning framework for classifying correctly oriented and reversed handwritten characters. A balanced sample of 60,000 images was derived from the Dyslexia Handwriting Dataset and partitioned using perceptual-signature grouping to reduce duplicate-related leakage. EduMirror-XAI combined a pretrained MobileNetV3Small backbone, spatial attention, controlled augmentation, and two-stage transfer learning. The validation-selected threshold was fixed before held-out testing. On 9,014 test images, the model achieved 92.69% accuracy, 91.53% sensitivity, 93.85% specificity, a 92.61% F1-score, an MCC of 0.8540, a ROC-AUC of 0.9794, and a PR-AUC of 0.9797. Performance remained stable under blur, rotation, and compression, while reduced contrast produced the largest decline. EduMirror-XAI demonstrates the feasibility of image-level reversal detection for preliminary teacher review but is not intended as a dyslexia diagnostic system.