Compact and Accurate: DenseNet121 for Automat-ed Classification of Crossed and Normal Eyes

Abstract

The early detection of eye disorders such as crossed eyes (strabismus) is essential for timely treatment and prevention of vision impairment. This paper presents an automated system for the classification of crossed eyes and normal eyes using deep learning techniques. A convolutional neural network (CNN), specifically DenseNet121, is employed to extract meaningful features from eye images and perform accurate classification. The proposed model is trained on a labelled dataset consisting of normal and crossed eye images, achieving high accuracy and reliability. Image preprocessing techniques such as resizing, normalization, and augmentation are applied to improve model performance. The results demonstrate that the system can effectively distinguish between crossed and normal eyes, making it suitable for real-time screening applications. This approach can assist healthcare professionals by providing a fast, cost-effective, and non-invasive diagnostic tool, especially in rural and low-resource settings. Automatic eye condition classification plays a crucial role in early diagnosis of vision disorders such as strabismus (crossed eyes). This paper presents a deep learning-based approach using DenseNet121 for automated classification of crossed and normal eyes. DenseNet121, a densely connected convolutional neural network, enhances feature propagation and reduces the vanishing gradient problem by connecting each layer to every other layer. The proposed system utilizes image preprocessing, data augmentation, and transfer learning to improve model performance on limited datasets. The model is trained on labelled eye images and evaluated using metrics such as accuracy, precision, recall, and F1-score. Experimental results demonstrate that the proposed DenseNet121-based model achieves high classification accuracy while maintaining computational efficiency. The compact architecture with fewer parameters makes it suitable for real-time medical applications. This work highlights the effec-tiveness of DenseNet121 in ophthalmological image classification and provides a reliable solution for automated eye condition screening.

Introduction

The paper presents an automated eye disease detection system that uses the DenseNet121 deep learning model to classify eye images as either Crossed Eye (Strabismus) or Normal Eye. Strabismus is a common vision disorder in which the eyes are misaligned, and if left untreated, it can lead to complications such as amblyopia (lazy eye) and permanent vision loss. Early diagnosis is therefore essential, but traditional methods rely on manual examination by ophthalmologists, which can be time-consuming, subjective, and difficult to access in rural or resource-limited areas.

To address these challenges, the proposed system applies artificial intelligence and deep learning to automate the diagnosis process. DenseNet121 is selected because of its efficient architecture, which promotes feature reuse, reduces the number of parameters, minimizes the vanishing gradient problem, and achieves high accuracy in medical image classification while requiring fewer computational resources than many conventional convolutional neural networks (CNNs).

The system begins by collecting a labeled dataset of eye images from publicly available and reliable sources. The images undergo preprocessing, including resizing, normalization, enhancement, and data augmentation, to improve image quality, reduce overfitting, and enhance model generalization. Using transfer learning, the pretrained DenseNet121 model is fine-tuned to automatically extract meaningful visual features that distinguish crossed eyes from normal eyes.

The proposed application follows a client–server architecture. Users register and log into the web application before uploading an eye image. The Flask backend validates and preprocesses the uploaded image and passes it to the trained DenseNet121 model for classification. The model predicts whether the eye image represents Crossed Eye (Strabismus) or Normal Eye, and the result, along with a confidence score, is displayed through a user-friendly interface. User information, uploaded images, and prediction history are securely stored in an SQLite database.

The system consists of several functional modules, including:

• User registration and login
• Eye image upload
• Image preprocessing
• Eye classification using DenseNet121
• Result display with confidence score
• Dashboard and database management

The application is implemented using Python, Flask, TensorFlow, and Keras for backend development and model implementation, while HTML, CSS, and JavaScript are used to create a responsive frontend interface. Security features such as authentication, input validation, session management, and protected database access ensure safe handling of user data.

Experimental evaluation was conducted using a labeled dataset divided into training, validation, and testing sets. Model performance was assessed using accuracy, precision, recall, and F1-score. Testing confirmed that all major system functions—including user registration, login, image upload, preprocessing, eye classification, result display, and dashboard management—operated successfully. The DenseNet121 model achieved high classification accuracy and demonstrated consistent performance, with transfer learning and image preprocessing contributing significantly to improved prediction quality.

The proposed system has several practical applications, including early strabismus screening, assisting ophthalmologists in diagnosis, supporting telemedicine and remote healthcare services, conducting vision screening in schools and hospitals, facilitating medical research, and enhancing AI-based healthcare systems.

Despite its effectiveness, the system has some limitations. Its performance depends on the quality and size of the training dataset, and prediction accuracy may decrease when images are blurred, poorly illuminated, or incorrectly positioned. Furthermore, the model is designed only to distinguish between crossed and normal eyes and cannot diagnose other eye disorders. Consequently, the system is intended to support clinical decision-making rather than replace professional medical diagnosis.

Conclusion

The proposed system uses the DenseNet121 deep learning model to accurately classify crossed eyes and normal eyes from eye images. The model provides fast and reliable predictions, making it useful for early detection of strabismus. This approach can assist healthcare professionals in improving diagnosis while reducing manual effort. Future work can focus on using larger datasets, improving accuracy, and developing a real-time web or mobile application.

References

[1] H. Singh, C. Indhumathi, R. Moudgil, D. Thakur, et al., “Compact and Accurate: DenseNet121 for Automated Classification of Crossed and Normal Eyes,” in Proceedings of the 2025 International Conference on Decision Aid Sciences and Applications (DASA), 2025, doi: 10.1109/DASA68193.2025.11498887. [2] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger, “Densely Connected Convolutional Networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. [3] Jiewei Lu, Zhun Fan, et al., “Automated Strabismus Detection for Telemedicine Applications,” 2018. [4] X. Huang, S. J. Lee, C. Z. Kim, and S. H. Choi, “An Automatic Screening Method for Strabismus Detection Based on Image Processing,” PLOS ONE, vol. 16, no. 8, 2021. [5] X. Huang, S. J. Lee, C. Z. Kim, and S. H. Choi, “An Improved Strabismus Screening Method with Combination of Meta-Learning and Image Processing Under Data Scarcity,” PLOS ONE, vol. 17, no. 8, 2022. [6] J. Joo, S. Y. Kim, D. Kim, et al., “Enhancing Automated Strabismus Classification with Limited Data: Data Augmentation Using StyleGAN2-ADA,” PLOS ONE, vol. 19, no. 5, 2024. [7] “Automated Strabismus Detection and Classification Using Deep Learning Analysis of Facial Images,” Scientific Reports, 2025. [8] J. C. A. Read, “Stereo Vision and Strabismus,” Eye, vol. 29, pp. 214–224, 2015. [9] Olga Russakovsky, Jia Deng, et al., “ImageNet Large Scale Visual Recognition Challenge,” International Journal of Computer Vision, vol. 115, no. 3, pp. 211–252, 2015. [10] K. W. Chan, L. Deng, and E. M. Weissberg, “Detection of Strabismus by Non–Health Care Professionals in an Ethnically Diverse Set of Images,” JAMA Ophthalmology, vol. 134, no. 1, pp. 30–36, 2016.

Copyright

Copyright © 2026 G. Himabindu , G. Manoj Kumar. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.