This is a retrospective study. All data were collected from June 2019 to September 2020 at the Affiliated Eye Hospital of Wenzhou Medical University, China. A total of 115 subjects (8–78 years old) were included in the study; Only one eye of each subject was randomly selected and included. Among them, 64 subjects were diagnosed with Demodex blepharitis, and the other 51 subjects without Demodex infestation and blepharitis were randomly selected and included as the control group.

The exclusion criteria were as follows: 1) a history of ocular trauma or surgery; 2) systemic drugs or eye drops affecting MG function or tear film used in the last 2 weeks; 3) contact lenses worn in the last 2 weeks; and 4) other ocular or systemic diseases affecting MG function or tear film.

Patients with MG density below 0.05 were removed to reduce the errors caused by the extreme atrophy of MGs. When the MG density was very low, and gland atrophy was severe, each index of MG was close to 0, and the evaluation efficiency of MG morphology was lost.

This study was performed in accordance with the Declaration of Helsinki and approved by the Research Ethics Committee of the Eye Hospital, Wenzhou Medical University (approval number: 2020-209-K-191). Written informed consent was obtained from all study patients. This study was registered at http://clinicaltrials. gov (NCT04451122).

Demodex blepharitis was diagnosed based on the diagnostic criteria (at least one symptom, such as redness, eye itching, foreign body sensation, abnormal eyelashes with cylindrical dandruff, and positive results on light microscopic examination of the eyelashes) (Amescua et al., 2019). Demodex mites were detected and counted by a professional technician. Three eyelashes were pulled from each eyelid using forceps under a slit-lamp microscope. The eyelashes from each eyelid were placed separately parallel to a glass slide, covered with a cover slide, and the Demodex mites were counted under a light microscope. Each glass slide was dropped with 0.25% fluorescein, with added 20 μl of cedar oil and 100% ethanol when there were too many scales in the root of the extracted eyelashes to improve detection and counting of mites. Patients were diagnosed with Demodex positive if 1) Demodex mites in each stage (mites whether dead or alive, larvae, and eggs) were observed; 2) more than three mites were found on any eyelid; and 3) less than the above standards may have been diagnosed as a suspicious infestation, which needed to be combined with clinical manifestations (Asian Dry Eye Association China Branch, 2018).

Meibography images were captured with a K5M (Markoulli et al., 2017). Then, the MGs from these images were extracted using a novel MG extraction method based on CNN with a residual neural network, which was introduced in our previous study (Zhang et al., 2022).

We replaced the 50-layer ResNet (ResNet50) with the max-pooling layers of the U-net model; however, the up-sampling layer remained the same, and this model was called ResNet50_U-net. For model training, the meibography images were firstly optimized, converted to grayscale, and then standardized and normalized (Dai et al., 2021). Forty annotation meibography images were included as the basis for the training set. Meanwhile, the data enhancement model, imgaug (https://github.com/aleju/imgaug#citation, 6 February 2020), was used to generate new training images from randomly selected original annotation meibography images. In this ResNet50_U-net model, the batch size of each iteration was four, so imgaug generated four random new images each time to form a mini batch. A complete training consisted of 120 epochs, while each epoch consisted of 512 iterations. Thus, it contained a total of 61,440 iterations, using 245,670 training images generated by imgaug. Another 20 annotated original meibography images apart from the training set were used as the validation set.

Intersection of unions (IoU) was used to evaluate the accuracy of the MG segmentation model. It can be simply understood as the ratio of the intersection of the ground truth (manual annotation) and AI result (AI segmentation) to their union. Finally, the ResNet50_U-net model achieved 92.00% in IoU, and 100% in repeatability.

The flowchart of image processing and segmentation is shown in Figure 1.

After extracting the MGs, the following indexes were automatically calculated with the aid of a computer (Zhang et al., 2022).

All participants underwent the ocular surface disease index (OSDI) questionnaire (Markoulli et al., 2017), slit-lamp examination and ocular anterior segment photography. The ophthalmic examinations included the following: 1) tear meniscus height (TMH): the height was measured 5 s after blinking, and the central TMH of the lower eyelid was measured. 2) Tear film break-up time (TBUT): The time before the first defect appeared on the stained tear film was measured as the TBUT after fluorescein instillation (repeated three times, and the mean value was recorded). 3) Corneal fluorescein staining (CFS): CFS was performed after the installation of fluorescein, and the degree was graded using the Baylor grading scheme from 0 to 4. 4) Lid margin score including anterior or posterior displacement of the mucocutaneous junction, vascular engorgement, plugged meibomian gland oriflces, and irregularity of the lid margin. It was scored from 0 to 4 according to the four aforementioned parameters. 5) Meibum expressibility score: The function of the 15 glands on each lower eyelid was assessed. The secretion for each gland was graded as follows: 0 = no secretion, 1 = inspissated (toothpaste consistency), 2 = cloudy, and 3 = clear-to-normal secretion. The scores ranged from 0 to 45 (Cochener et al., 2017).

The normality of data distributions was analyzed using the Kolmogorov–Smirnov test, and abnormal data distributions were analyzed using non-parametric statistical analyses. Values are expressed as the mean ± standard deviation (SD) or median [interquartile range (IQR)], and the range of data. The independent samples t-test or Mann–Whitney U-test was used to compare differences between the two groups. A generalized estimating equation was used to adjust for age differences. The χ2 test was used to compare sex ratios between the two groups. Receiver operating characteristic (ROC) curve analysis was used to determine the predictive value of total variation in the diagnosis of MGD with Demodex infestation. Multivariate regression analyses were used to evaluate the relationship between the total variation and MG parameters. A two-sided p < 0.05 was considered statistically significant. All statistical analyses were performed using SPSS Statistics (version 23.0; IBM, Armonk, NY, United States).

In this retrospective study, the records of 115 subjects were analyzed at the dry eye department, Eye Hospital Wenzhou Medical University, Zhejiang, China. Of these, 64 subjects diagnosed with Demodex blepharitis were in the Demodex group, and 51 subjects without Demodex blepharitis were in the control group. The basic characteristics and clinical parameters are shown in Table 1.

Participants in the Demodex group were significantly older [39.0 (28.0–57.0) years]) than those in the control group (28.98 ± 11.46 years), so the analysis of covariance was used to adjust for age for the comparisons between the two groups. Significant differences were noted in the OSDI, TBUT, lid margin score, meiboscore, and meibum expressibility scores between the two groups (all p < 0.05; Table 1). No difference was noted in the TMH (p = 0.086), and CFS (p = 0.587).

Table 2 shows the analysis results of the morphological changes in MGs in the Demodex and control groups. Compared to the control group, patients in the Demodex group had significantly lower densities and shorter heights of all MGs in the upper eyelid (all p < 0.05). However, there were no statistically significant differences in the average tortuosity (p = 0.426), width (p = 0.557), and gland number (p = 0.885) between the two groups.

The total variation was significantly higher in the Demodex-positive group than in the Demodex-negative group (p = 0.004).

As shown in Figure 2, the unevenness of gland height in the meibography images of Demodex group (A) was significantly higher than that in the control group (B). However, there was no significant difference in MG density between them. The total variation, represented by the amplitude of the red line (A④ and B④), can quantitatively measure and distinguish the unevenness of the gland height.

The total variation is calculated based on the height of each MG; therefore, it may be affected by various MG parameters such as height, density, and quantity, among others. Multivariate regression analysis was used to analyze the factors that impacted the total variation, and the results are shown in Table 3.

Multivariate linear regression analysis revealed that these three factors, N (gland number), H (total MG height), and D (MG density), had an impact on total variation. But the coefficient of H (total MG height) is too small to draw a correlation between H and total variation. Therefore, we corrected the total variation using these two factors (gland number and MG density) and analyzed the statistical results of the corrected total variations as Uneven indexes between the two groups, as shown in Table 4.

When considering the influence of the gland number and MG density on the uneven degree of MG, we found that there were also statistical differences between the two groups on the two types of Uneven indexes. The average values of total variation (p = 0.004), Uneven index 1 (p = 0.003), and Uneven index 2 (p < 0.001), quantitative measures for the unevenness of MG atrophy, were all significantly higher in the Demodex group than in the control group.

The potential value of total variation and Uneven index was used to discriminate Demodex blepharitis individuals from the control group. Based on the multivariate regression analysis results, ROC curves were generated to show the prediction of Demodex blepharitis using the variates (Figure 3). There was one type of Uneven index (Uneven index 2) that showed the ability to discriminate Demodex blepharitis subjects from the control group.

The area under the curve (AUC) was 0.667 for total variation. The sensitivity and specificity were 67.2% and 66.7%, respectively, at a cut-off value of 778.5.

The AUC was 0.694 for Uneven Index 1 (total variation/N). The sensitivity and specificity were 48.4% and 82.4%, respectively, at a cut-off value of 59.625.

The AUC was 0.822 for Uneven Index 2 (total variation/D). The sensitivity and specificity were 59.4% and 92.2%, respectively, at a cut-off value of 3971.667.