[Paper Review] Automated Strabismus Detection for Telemedicine Applications

Jiewei Lu, Zhun Fan+5 more|arXiv (Cornell University)|Sep 9, 2018

Retinal Imaging and Analysis18 references18 citations

TL;DR

This paper proposes RF-CNN, an end-to-end deep learning framework for automated strabismus detection in telemedicine. It first segments eye regions using R-FCN and then classifies them as strabismus or normal using a CNN. The method achieves 93.30% sensitivity, 96.17% specificity, and 98.65% AUC on a newly established tele strabismus dataset, demonstrating strong performance for remote diagnosis.

ABSTRACT

Strabismus is one of the most influential ophthalmologic diseases in human's life. Timely detection of strabismus contributes to its prognosis and treatment. Telemedicine, which has great potential to alleviate the growing demand of the diagnosis of ophthalmologic diseases, is an effective method to achieve timely strabismus detection. In this paper, a tele strabismus dataset is established by the ophthalmologists. Then an end-to-end framework named as RF-CNN is proposed to achieve automated strabismus detection on the established tele strabismus dataset. RF-CNN first performs eye region segmentation on each individual image, and further classifies the segmented eye regions with deep neural networks. The experimental results on the established tele strabismus dataset demonstrates that the proposed RF-CNN can have a good performance on automated strabismus detection for telemedicine application. Code is made publicly available at: https://github.com/jieWeiLu/Strabismus-Detection-for-Telemedicine-Application.

Motivation & Objective

• To address the growing need for timely strabismus diagnosis in remote or underserved areas.
• To overcome the limitations of traditional on-site diagnosis requiring specialist presence.
• To develop an automated, scalable solution for strabismus detection using telemedicine.
• To establish a publicly available tele strabismus dataset for research and clinical validation.
• To enable early detection of strabismus to prevent complications like amblyopia and blindness.

Proposed method

• An end-to-end RF-CNN framework is proposed, consisting of two stages: eye region segmentation and strabismus classification.
• R-FCN is used to generate position-sensitive score maps for precise eye region segmentation, reducing background interference.
• The segmented eye regions are fed into a deep convolutional neural network (CNN) for final classification into strabismus or normal.
• The network is trained using stochastic gradient descent with weight decay of 5×10⁻⁴ and dropout rate of 0.5.
• Training is conducted over 5000 iterations with a batch size of 32 and initial learning rate of 0.01 on a GPU-equipped system.
• The framework is implemented using TensorFlow and validated on a newly collected tele strabismus dataset.

Experimental results

Research questions

• RQ1 Can an automated deep learning system achieve high-accuracy strabismus detection in telemedicine settings?
• RQ2 How does the performance of automated strabismus detection vary with the number of training examples?
• RQ3 Can eye region segmentation improve detection accuracy in complex, variable telemedicine images?
• RQ4 Is it feasible to develop a publicly available tele strabismus dataset for research and clinical use?
• RQ5 Can the proposed RF-CNN framework outperform traditional methods in sensitivity and specificity for remote diagnosis?

Key findings

• RF-CNN achieved a sensitivity of 93.30% on the tele strabismus dataset, indicating strong ability to detect strabismus cases.
• Specificity reached 96.17%, showing high accuracy in identifying normal eye images.
• The model achieved an accuracy of 93.89% and an AUC of 0.9865, indicating excellent overall diagnostic performance.
• Performance improved significantly with up to 1500 training examples, after which gains plateaued.
• The sensitivity analysis confirmed that model performance is robust and scalable with increasing training data.
• The study presents the first end-to-end automated strabismus detection framework for telemedicine, validated on a newly established dataset.