Enhancing Strabismus Diagnosis from Detection to Classification with Deep Learning

Abstract

Accurate and early diagnosis of strabismus, a disorder characterized by ocular misalignment, is essential to prevent long-term visual impairment. Traditional diagnostic approaches heavily rely on clinical expertise, often introducing subjectivity and inconsistency. In this study, we propose a deep learning-based framework for automated strabismus classification using four state-of-the-art convolutional neural networks — EfficientNet-B7, ResNet-50, AlexNet, and VGGNet-19. The models are trained on a diverse open-source dataset to classify strabismus into five categories: Esotropia, Exotropia, Hypertropia, Hypotropia, and Normal eye alignment. Our extensive experiments demonstrate that EfficientNet-B7 outperforms other models, achieving superior accuracy and generalization across evaluation metrics, including precision, recall, and F1-score.

Cite this paper

@inproceedings{gupta2025strabismus,
  title     = {Enhancing Strabismus Diagnosis from Detection to Classification with Deep Learning},
  author    = {Gupta, Bhumit and Arora, Madhav and Garg, Shubh and Ghosh, Debabrata},
  booktitle = {Third International Conference on Artificial Intelligence and Machine Learning Applications (AIMLA)},
  year      = {2025},
  address   = {Tiruchengode, Tamil Nadu, India},
  note      = {Paper ID 1140}
}