Table 4 The table presents a review of several investigations using deep learning-based approaches for the detection and classification of strabismus.

Jaehan Joo et al.²⁹

This work introduces a generative data augmentation method via StyleGAN2-ADA for improving strabismus classification performance in situations of limited available data. The authors investigate the performances of two different classifiers and compare the newly generated data with classic augmentation techniques.

StyleGAN2-ADA for data generation and ResNet50 ResNext101 for classification tasks.

The accuracy:

Without augmentation and generated data addition:

ResNet-50 = 85.71%

ResNext101 = 87.75%

Ce Zheng et al.³⁰

The study collected 479 strabismus surgery videos that were segmented manually into 3345 clips representing eight surgical steps.

Two hybrid deep learning (DL) algorithms were considered: a Recurrent Neural Networks (RNN)-based model and a Transformer-based model.

Used a Convolutional Neural Network (CNN) followed by RNN layers (specifically Gated Recurrent Units) and a pre-trained DenseNet for feature extraction, utilizing a Transformer architecture for classification.

Accuracy:

Transformer-based model: 0.96

RNN-based model: 0.83

Precision:

Transformer-based model: 0.90–1.00

RNN-based model: 0.75–0.94

F1-score:

Transformer-based model: 0.93–1.00

RNN-based model: 0.78–0.92.

Dawen Wu et al.³¹

The study aimed at building an artificial intelligence platform embodied as a mobile app for the screening and management of strabismus. The model used primary gaze photos, including 6,194 images with corneal light reflection: 2,938 exotropia, 1,415 esotropia, 739 vertical deviation, and 1,562 orthotropy, sourced from two independent datasets covering a wide age range. The Visual Transformer (ViT_16_224) architecture was implemented to process the visual information. The AI model was evaluated using 5-fold cross-validation and subsequently tested on a separate independent dataset.

Visual Transformer (VIT_16_224)

The AI model obtained the following results in the internal validation set: Precision 0.941, Accuracy 0.980, Specificity 0.979, F1-Score 0.951, and the independent test set:

Accuracy :0.967, Precision :0.980, Specificity :0.970, F1-Score :0.975,

Ayesha Jabbar et al.³²

This work presents a new method for the diagnosis of strabismus by combining a Federated Convolutional Neural Network (FedCNN) and eXtreme Gradient Boosting (XGBoost). The new method utilizes eye-tracking data, namely Gaze deviation (GaDe) images, to improve the diagnosis accuracy of strabismus. It consists of recording ocular movements, generating GaDe images, and applying CNN and XGBoost for feature extraction and classification.

Convolutional Neural Network (CNN) and Federated Convolutional Neural Network (FedCNN).

The model’s accuracy rate of 95.2%.

Rashid Amin et al.³³

The study made use of deep learning algorithms in the detection of strabismus and other eye-related ailments. The methodology included image acquisition, region of interest extraction, extraction of features, and classification

ResNet50 and VGG16.

The study reported an accuracy of 92% for the ResNet50 model and 79% for the VGG16 model.

Yena Christina Kang et al.³⁴

An automated algorithm measures ocular deviation in strabismus patients using nine cardinal gaze positions. It uses deep learning and image processing techniques, including data collection, preprocessing, and training a U-Net convolutional neural network.

The deep learning model used the U-Net architecture to segment the limbus and sclera.

the Dice Similarity Coefficient (DSC) for segmenting the limbus was 95.71%, while the DSC for segmenting the sclera was 96.88%.

Jiewei Lu et al.²⁸

An end-to-end framework for automatic strabismus identification called RF-CNN was created by the researchers. There were two primary processes in the process: first, each image’s eye regions were segmented; second, deep neural networks were used to classify the segmented eye regions.

RF-CNN

The study reported 93.89% as accuracy, 93.30 as sensitivity and 96.17% as specificity