Sunken upper eyelid correction is clinically and aesthetically important in periorbital reconstruction. However, current evaluation methods for sunken upper eyelid are limited by the expensive and complex medical equipment or subjective judgment. This paper proposes a single image assessment framework based on computer vision (CV) that enables objective evaluation of sunken upper eyelid morphology. The method consists of two stages: The first stage is performed to establish a consistent analysis foundation. Facial landmarks are detected using a 106-point model to crop the periocular region, followed by resolution normalization and eyelid subregion segmentation. In the second stage, three clinically relevant features are extracted: Variance of Gray Value (VGV) for surface depression, Structural Similarity Index (SSIM) for bilateral symmetry, and Degree of Eyelid Wrinkles (DEW) for texture irregularity. These features are integrated into a Support Vector Machine (SVM) model, which outputs the L2 distance to the separating hyperplane () to score overall morphological features. Statistical results on the samples showed significant differences in VGV, SSIM, and DEW between normal group and patient group. Based on the proposed method, most patients had a positive postoperative to preoperative difference of , indicating measurable improvements in surgical outcomes.
Sunken upper eyelid correction surgery is a common reconstructive procedure. Its main purpose is to correct the sunken deformity from the upper eyelid margin to the orbital rim area, restoring the natural flow of the eyelid contour. This surgery not only improves visual function and eyelid closure, but also enhances facial appearance, positively contributing to patient quality of life and psychological well-being. Currently, there are two main approaches for surgical outcome evaluation. The first involves objective measurement using medical imaging equipments such as MRI and CT, as well as clinical measurement tools like the exophthalmometer. The second relies on subjective clinical judgment made by attending physicians. Although the former approach offers objectivity and standardization, it imposes significant economic and time burdens on patients. This is mainly due to high equipment costs, complex operating procedures, and the inability to conduct real-time dynamic evaluations. The latter method heavily depends on the surgeon's experience and is prone to subjective cognitive biases, leading to poor reproducibility and insufffcient objectivity in the evaluation results. Furthermore,none of these methods can balance efficiency and accuracy, affecting communication between doctor and patient.
In recent years, with the popularization of photography technology, image analysis-based methods for postoperative aesthetic evaluation have gradually gained attention. This method has become an essential tool for recording and monitoring patient recovery. It provides an intuitive reference for doctors in diagnosis, surgical planning, and outcome evaluation. Tosan et al. utilized preoperative and postoperative static photographs from frontal and lateral views in the evaluation of open rhinoplasty. They calibrated the images using the interpupillary distance and the distance from the glabella to the pogonion. They measured the upper lip length in the frontal view and the positions of the subnasale and labrale superius in the lateral view. Finally, the surgical outcomes were assessed using a one-sample t-test. Gomez-Rice et al. used preoperative and postoperative full-body standing clinical photographs for the assessment of postoperative satisfaction in adult spinal deformity surgery. It compares the changes in "self-image" and "satisfaction" scores on the SRS-22 questionnaire of the patients before and after the photographs. It confirming that image analysis can improve patients' postoperative subjective satisfaction. Raschke et al. used standardized frontal photos of open and closed eyes taken preoperatively and 3 months postoperatively for postoperative evaluation of blepharoplasty. The photos were captured by a Nikon D80 camera at a fixed distance. The method manually labeled key points and compared seven geometric metrics before and after. They found significant improvements in postoperative lacrimal index and canthus tilt. Ilke et al. evaluated upper blepharoplasty using preoperative and 6 months postoperative frontal full face photos. Their photos were taken by a professional camera under fixed conditions. The images were calibrated by 68-point facial keypoint detection, and geometric parameters such as lid distance and eye opening area were measured. The surgical results were subsequently evaluated by comparing the differences in postoperative parameters. Despite the achievements of these methods, there remain certain limitations. First of all, relying on professional photographic equipment and complex operation process, it is difficult to realize the immediate feedback of postoperative results. It requires specialized skills and is time-consuming, making it difficult to be popularized. Second, feature selection is limited to geometric parameters (e.g., distance and area), ignoring key morphological metrics such as symmetry and texture features. Thus, developing a low - cost, convenient, multi - scenario - applicable assessment method is urgent. This will enhance the assessment of sunken upper eyelid and boost its clinical application value.
Since the human visual system is highly sensitive to the three morphological features of luminance, symmetry, and texture, this paper proposes a computer vision-based method for assessment of sunken upper eyelid. The proposed method employs two stages. In the initial stage, we transformed the original frontal facial photographs into standardized images suitable for analysis through a step-by-step processing framework. First, Face++ 106-point facial recognition model is used to accurately label key eye points such as the corners of the eyes and the edges of the eyelids. This step provides geometric anchor points for subsequent spatial alignment. Next, based on these points, a rectangular region of interest (ROI) is defined. This cropping step isolates the periocular area and effectively filtering out irrelevant background noise. Then, to address the resolution differences caused by different shooting devices, the cropped images are uniformly processed. This processing ensures that all images have consistent clarity and satisfy the recognition requirements of human visual perception. Finally, using the lower edge of the eyebrow as a reference, each eyelid is divided into four small subregions. This division facilitates a more detailed analysis of local feature changes. In the second stage, using the preprocessed data, computer vision techniques were used to extract three morphological features highly relevant to human visual perception. These features are luminance, symmetry, and texture. Given that upper eyelid sunkenness manifests as deepened shadows, poor symmetry, and increased wrinkles, this paper uses these features to construct quantitative indicators. Specifically, the Variance of Gray Value (VGV) is used to measure the uneven distribution of luminance on the eyelid surface. The Structural Similarity Index (SSIM) is employed to quantify the symmetry of bilateral eyelids. Meanwhile, the Degree of Eyelid Wrinkles (DEW) is utilized to describe the fold characteristics of the sunken area. Finally, by training a support vector machine (SVM) model, these features are integrated into a quantitative metric, defined as the L2 distance to the separating hyperplane (). The difference in this distance between postoperative and preoperative states, denoted as , is directly correlated with surgical improvement. Positive values indicate good improvement of the sunken upper eyelid after surgery.