AI is a branch of computer science, in which “artificial” indicates that the systems are man-made and “intelligence” denotes features such as consciousness and thinking (Thrall et al., 2018). The major purpose of AI is to simulate human thinking processes by learning from existing experiences to solve problems that cannot be solved through traditional computer programming (Bischoff et al., 2019). ML is a subset of AI that has become the mainstream of AI technology (Shin et al., 2021). ML extracts and analyzes the features of input samples to classify new homogeneous samples (Totschnig, 2020). ML automatically improves and optimizes computer algorithms and programs by analyzing the data rather than relying on explicit program instructions (Nichols et al., 2019; Cho et al., 2021). Among the various ML models that have emerged, neural networks simulate the synaptic structure of human neurons and improve the computational ability of ML by adjusting the parameters of network models (Starke et al., 2021). Convolutional neural networks (CNNs), which have an encoding structure similar to that of visual cortical neurons (Hou et al., 2019), have become one of the most popular neural network models (Mintz and Brodie, 2019).

In human vision, each neuron in the visual cortex responds to stimulation by activating specific regions in the visual space that form the entire visual field (Figure 1) (Brachmann et al., 2017). Similarly, CNNs extract features from the input image and output a feature map using convolution and pooling operations (Le et al., 2020). A convolution layer consists of a set of two-dimensional numerical matrices that are also known as filters. The CNN obtains the pixel value of the output images by multiplying the value in the filter by the value of the corresponding pixel in the image and summing the product, that is, via convolution operations (Brachmann et al., 2017). To avoid similar sizes of the output pixels after the convolution operation, the CNN changes the size of the output pixels by reducing the input values through the pooling operation. By repeating the convolution and pooling operations, the CNN continuously self-corrects so that the output values become closer to the human ratings (Larentzakis and Lygeros, 2021). New neural network models, such as UNet and ResNet, have been developed to overcome the difficulty of training CNNs with deep layers. These neural network models improve the framework of a CNN by expanding its depth, convolutional layer, or pooling layer. For example, while traditional CNN models can only classify images and output the labeling of an entire image, UNet can achieve pixel-level classification and output the class of each pixel, which makes it well-suited for image segmentation tasks (Yin et al., 2022). ResNet solves the gradient vanishing and gradient exploding problems of traditional CNNs by adding a residual block (He et al., 2020).

To process large amounts of data, multilayer neural networks have been cascaded to form DL algorithms (Kaluarachchi et al., 2021). Compared with traditional ML algorithms, DL has a greater ability to analyze large-scale matrix data (Jalali et al., 2021). The relationship between AI, ML, and DL is shown in Figure 2. Currently, DL-based technologies are widely used in the diagnosis of certain ophthalmic diseases, such as cataracts (Wu et al., 2019) and glaucoma (Sudhan et al., 2022), and the segmentation of medical images, including those of retinal vessels (van der Heijden et al., 2018).

Automatic recognition and labeling of anatomical structures of the eye orbit can be achieved through the segmentation of medical images based on AI technology (Hou et al., 2019). Furthermore, AI can segment bony structures from orbital CT/MRI images. Hamwood et al. (2021) developed a DL system for the segmentation of bony regions from orbital CT/MRI images that exhibited excellent efficiency, particularly in terms of computational time. Li et al. (2022a) extracted bony orbit features and analyzed Asian aging characteristics through the popular deep CNN (DCNN) model. Some commercial software can also automatically segment orbital regions from CT images (Hamwood et al., 2021). In addition, using AI-based technology, irregular soft tissues, such as fat and abscesses, have been reliably segmented from orbital CT/MRI images. Brown et al. (2020) used a UNet-like CNN to segment orbital septal fat from orbital MRI images, and the results showed that AI segmentation was consistent with manual segmentation. Fu et al. (2021) trained and evaluated a context-aware CNN (CA-CNN) to segment orbital abscess regions from CT images of patients with orbital cellulitis, with the AI results being similar to those obtained by medical experts.

In addition, AI can automatically quantify certain anatomical structures based on image segmentation. Umapathy et al. (2020) established an MRes-UNet model to segment and quantify the volume of the eyeball based on orbital CT images. Pan et al. (2022) achieved automatic calculation of the size and height of the bony orbit regions using a U-Net++ based on pre-3D images reconstructed from orbital CT images.

Orbital CT/MRI images are crucial in the preliminary diagnosis of orbital diseases such as orbital wall fractures, orbital tumors, and TAO (Griffin et al., 2018). Orbital blowout fractures are one of the most common injuries caused by orbital trauma. Li et al. (2020) used the Inception V3 DCNN to automatically classify CT images exhibiting orbital burst fractures. Song et al. (2021a) proposed a 3D-ResNet to automatically detect TAO from orbital CT images, and the trained AI algorithm showed excellent performance in a real clinical setting. Lin et al. (2021) used a DCNN to grade TAO based on orbital MRI images, resulting in a labeling of disorder areas that was consistent with that made manually through an occlusion test. Hanai et al. (2022) developed a deep neural network to assess the enlarged extraocular muscles (EEM) of patients with Graves’ ophthalmopathy (GO) from orbital CT images. When applied to the test data, the area under the receiver operating curve (AUC) was 0.946, indicating that the deep neural network could effectively detect EEM in GO patients. Lee et al. (2022) used 288 orbital CT scans from patients with mild and moderate-to-severe GO and healthy controls to train a neural network for diagnosing and assessing the severity of GO. The developed neural network yielded an AUC of 0.979 in diagnosing patients with moderate-to-severe GO. Han et al. (2022) automatically identified the differences in the orbital cavernous venous malformations (OCVM) from orbital CT images by training 13 ML models, including support vector machines (SVMs) and random forests. Nakagawa et al. (2022) implemented a VGG-16 network to determine from CT images whether a nasal or sinus tumor invades the periorbital area. The network model achieved a diagnostic accuracy of 0.920, indicating that CNN-based DL techniques can be a useful supporting tool for assessing the presence of orbital infiltration on CT images.

In addition to diagnosing and grading diseases, AI can extract and determine subtle features from images to differentiate confusing diseases. Orbital cavernous hemangioma and schwannoma differ in terms of surgical strategy but have similar MRI features. Bi et al. (2020) developed a database of orbital MRI images of patients with cavernous hemangioma and schwannoma from 45 hospitals in China and used AI to identify and classify the affected eye, tumor location, and tumor category. The AI system was validated, showing an accuracy greater than 0.900 on a multicenter database. Xie et al. (2022) developed a DL model that combines multimodal radiomics with clinical and imaging features to distinguish ocular adnexal lymphoma (OAL) from idiopathic orbital inflammation (IOI). The diagnosing results yielded an AUC of 0.953, indicating that the DL-based analysis may successfully help distinguish between OAL and IOI. Hou et al. (2021) used an SVM classifier and the bag-of-features (BOF) technique to distinguish OAL from IOI based on orbital MRI images. During an independent verification test, the proposed method with augmentation achieved an AUC of 0.803, indicating that BOF-based radiomics might be a new tool for the differentiation between OAL and IOI. Early detection of hypothyroid optic neuropathy (TON) is crucial in clinical decision-making. Wu et al. (2022) built an AI predictive model to distinguish between TAO and TON by extracting radiomic features from optic-nerve T2-weighted water-fat images from a cohort of patients with TAO and a cohort of patients with TON. Table 1 summarizes the discussed AI-related studies on orbital CT/MRI images.

The accurate measurement of eyelid morphological parameters is crucial in developing an individual eyelid surgery strategy. However, manual measurement of eyelid morphological parameters is difficult to replicate because of subjective errors induced by head movements and changes in facial expressions. AI provides a more objective and convenient tool for quantifying eyelid morphological parameters by parameterizing facial structures and automatically measuring length, area, and volume. Moriyama et al. (2006) achieved eye motion tracking based on the eyelid structure parameters and iris position. Van Brummen et al. (2021) utilized a ResNet-50 model to segment regions, such as the iris and eyebrow, to measure the marginal reflex distance (MRD) in static and dynamic external ocular photographs. Simsek and Sirolu used computer vision algorithms to automatically measure pupillary distance (PD), eye area (EA), and average eyebrow height (AEBH) from external ocular photographs for evaluating the surgical effect of patients who had undergone Muller’s muscle-conjunctival resection (MMCR) surgery (Bahceci Simsek and Sirolu, 2021). The automated measurement of eyelid morphology parameters based on AI technology helps assess eyelid status and improves the accuracy of eyelid surgery.

Compared with other types of imaging data, external ocular photographs can be taken and shared by patients and physicians through smartphones and the internet, which provides sufficient data for AI research on external ocular photographs. Chen et al. (2021) compiled CNN algorithms using the software MAIA to build DL models for the automatic measurement of MRD1, MRD2, and levator muscle strength based on external ocular photographs taken with smartphones. This study was the first smartphone-based DL model for the automatic measurement of eyelid morphological parameters. Compared to those obtained manually, measurements taken with the aid of AI are more objective.

Ptosis is a common eyelid disorder in which a drooping eyelid obscures the pupil, hindering vision in severe cases. Ptosis is generally diagnosed by measuring eyelid morphological parameters, such as the levator muscle strength, lid fissure height, and limbal reflex distance, based on typical clinical symptoms. Surgical therapy is the main treatment for ptosis (Mahroo et al., 2014). Tabuchi et al. (2022) performed an automatic diagnosis of ptosis using a pre-trained MobileNetV2 CNNi applied to photos of patients taken with an iPad Mini. Hung et al. (2022) realized the automatic identification of monocular appearance photos of ptosis patients based on a VGG-16 neural network, and the results showed that AI outperformed GPs in diagnosing ptosis. Combined with devices such as smartphones, the analysis of eye appearance based on AI can be useful in further clinical scenarios. AI provides an objective tool for measuring eyelid morphological parameters and planning surgery strategies instead of relying on the experience of surgeons. Song et al. (2021b) developed a gradient-boosted decision tree (GBDT) for choosing ptosis surgery strategies and trained it with 3D models created by photographing and scanning the eyes of ptosis patients with a structured light camera. The AI model evaluates the external ocular photographs and the 3D model to determine whether surgery is required and establish the surgery strategy to follow. Lou et al. (2021) evaluated the outcome of ptosis surgery by comparing pre- and postoperative values of eyelid morphological parameters, such as MRD1 and MRD2, which were automatically measured by a UNet from ocular appearance photographs of the patients.

Oculoplastic surgery involves the aesthetic restoration and predicting postoperative outcomes through AI can help the surgeon develop a personalized plastic surgery plan. The major purpose of oculoplastic surgery is to realize the expected aesthetic goals. However, it is hard to judge the expected aesthetic results due to a variety of subjective factors (Swanson, 2011). Establishing an objective facial beauty standard is still controversial. Zhai et al. (2019) proposed a new facial detection method based on a transfer learning CNN, which has better classification accuracy than previous geometric assessment methods, laying a foundation for the prediction of oculoplastic surgery effects. Yixin et al. explored the effect of the eyelid on oculoplastic surgery and aesthetic outcomes by comparing the postoperative metrics of oculoplastic patients assessed by the CNN model with those assessed only artificially. The CNN assessment group had better postoperative extent, lower eyelid skin wrinkles, eyelid tear troughs, skin shine, and aesthetic scores than the control group, suggesting that CNN is a beneficial tool for evaluating oculoplastic surgery (Yixin et al., 2022).

Eyelid and periocular skin tumors seriously affect the health and aesthetics of patients (Silverman and Shinder, 2017). Early preliminary screening through external photography helps detect and monitor these tumors. Seeja and Suresh (2019) trained a UNet to automatically segment skin lesions and differentiate melanoma from benign skin lesions, achieving reliable results in the segmentation and diagnosis of melanoma. Li et al. (2022b) used a faster region-based CNN and a DL classification network to build an AI system that automatically detects malignant eyelid tumors from ocular external photographs, obtaining positive performance on both internal and external test sets (AUC ranging from 0.899 to 0.955). CNNs could fully mine image information and distinguish deep features from external photography to detect subtle eyelid and skin tumors that are elusive to the naked eye, thus helping reduce misdiagnosis and missed diagnosis.

Changes in ocular appearance, such as retraction of the upper eyelid, strabismus, and proptosis, are crucial in the diagnosis of TAO (Hodgson and Rajaii, 2020). Huang et al. (2022) used the ResNet-50 model to obtain an automatic diagnosis of TAO based on external ocular photographs. Karlin J et al. (2022) developed a DL model for detecting TAO based on external ocular photographs. A set comprising 1944 photographs from a clinical database was used for training, and a test set of 344 additional images was used to evaluate the trained DL network. The accuracy of the model on the test set was 0.892, and heatmaps showed that the model could identify pixels corresponding to the clinical features of TAO. Orbital decompression surgery can alleviate the symptoms of eye protrusion and repair the appearance of patients with TAO. According to the 2021-EUGOGO guidelines, orbital decompression surgery is the recommended treatment strategy for patients with severe TAO (Smith, 2021). Yoo et al. (2020) trained a conditional generative adversarial network (GAN) using pre- and postoperative external ocular photographs of patients with orbital decompression. The trained GAN could convert the preoperative external ocular photographs into predictive postoperative images, which were similar to the real postoperative condition, suggesting that GAN might be a new tool for the prediction of oculoplastic surgery results. Table 2 summarizes the aforementioned AI-related studies on external ocular images.