Monitoring the transition from corneal ulceration to healed scar using a Scheimpflug tomography-based densitometry.

PURPOSE: To compare corneal haze between active ulcer and healed scarring using a Scheimpflug densitometry. MATERIALS AND METHODS: A prospective longitudinal study enrolled 30 patients (30 eyes) with ulcerative keratitis (UK). Each subject's corneal optical density (COD) was measured with a Scheimpflug corneal densitometry, Pentacam® AXL (Oculus GmbH, Wetzlar, Germany), at the active ulcerative and complete scarring stage. The COD data were analyzed through distinct methods (inbuilt, sorted annular partitions, and ulcer-matching densitometric maps). We compared different CODs to select the better index for clinically monitoring the transition from corneal ulceration to healed scar. RESULTS: The CODs of the periphery (P = 0.0024) and outside of the active ulcer (P = 0.0002) significantly decreased after scarring. Partitioning the cornea into different depths and annular zones, the anterior layer, center layer, and the 2-6 mm annular zone had a more remarkable COD decrease after scar formation. The 3rd-sorted COD in the anterior layer revealed the highest area under the receiver-operating characteristic curves (0.709), in which 90% of subjects had COD reduction during the ulcer-to-scar transition. CONCLUSIONS: Aside from subjective judgment based on clinical signs, the Scheimpflug tomography-based densitometry could provide objective and efficient monitoring of the corneal opacity evolution in UK patients. Because the 3rd-sorted annular COD is a better index than the inbuilt or mapping CODs in differentiating active ulcers from healed scars, this COD could be a clinically promising parameter to monitor the progression of UK patients.

Monitoring the Progression of Clinically Suspected Microbial Keratitis Using Convolutional Neural Networks.

Purpose: For this study, we aimed to determine whether a convolutional neural network (CNN)-based method (based on a feature extractor and an identifier) can be applied to monitor the progression of keratitis while managing suspected microbial keratitis (MK). Methods: This multicenter longitudinal cohort study included patients with suspected MK undergoing serial external eye photography at the 5 branches of Chang Gung Memorial Hospital from August 20, 2000, to August 19, 2020. Data were primarily analyzed from January 1 to March 25, 2022. The CNN-based model was evaluated via F1 score and accuracy. The area under the receiver operating characteristic curve (AUROC) was used to measure the precision-recall trade-off. Results: The model was trained using 1456 image pairs from 468 patients. In comparing models via only training the identifier, statistically significant higher accuracy (P < 0.05) in models via training both the identifier and feature extractor (full training) was verified, with 408 image pairs from 117 patients. The full training EfficientNet b3-based model showed 90.2% (getting better) and 82.1% (becoming worse) F1 scores, 87.3% accuracy, and 94.2% AUROC for 505 getting better and 272 becoming worse test image pairs from 452 patients. Conclusions: A CNN-based approach via deep learning applied in suspected MK can monitor the progress/regress during treatment by comparing external eye image pairs. Translational Relevance: The study bridges the gap between the investigation of the state-of-the-art CNN-based deep learning algorithm applied in ocular image analysis and the clinical care of suspected patients with MK.

Comparing Tear Film Viscosity between Sjögren and Non-Sjögren Dry Eye Disease.

This study aimed to compare tear film viscosity (TFV) in Sjögren and non-Sjögren dry eye diseases (DEDs). This was a cross-sectional observational study. A total of 68 DED patients were enrolled, including 32 patients with Sjögren syndrome (SS) and 36 without SS. TFV was assessed by a tear film analyzer and determined by the momentary moving speed (MMS; MMS (t) = α × t-ß, t = time (s)) with its power-law fitting-derived parameters (α and ß). Among the four indices of TFV (MMS (0.1 s), MMS (2.0 s), α, and ß), the SS-DED patients had significantly lower MMS (0.1 s) (p = 2.01 × 10-5), α (p = 0.0375), and ß (p = 0.0076). The SS-DED group also had significantly higher OSDI, lower central and nasal tear meniscus height (TMH), and higher OSS. MMS (0.1 s) was significantly correlated with nasal TMH and OSS (ρ = 0.2520, p = 0.0381 in nasal TMH; ρ = -0.3487, p = 0.0036 in OSS). Index ß was not correlated with any non-TFV tests. In conclusion, MMS (0.1 s), α, and ß are promising TFV indices in distinguishing SS-DED from non-SS-DED patients early. Among these TFV indices, lower MMS is the best alternative clue for detecting SS-DED.

Urinary excretion patterns and potential risks of beta-blocker ophthalmic drops in sports.

Beta-blockers have been prohibited by the World Anti-Doping Agency (WADA) in certain sports, but insufficient research data make it difficult to distinguish between therapeutic uses or misuses. This study aimed at investigating the urinary excretion pattern following beta-blocker ophthalmic drops and the potential risk of constituting an adverse analytical finding (AAF) in sports. Prescribed timolol and carteolol ophthalmic drops were used in healthy participants and glaucoma patients. The urine samples were then collected to investigate the urinary excretion pattern following acute and chronic administration of the above beta-blocker ophthalmic drops. The liquid chromatograph-tandem mass spectrometry method was applied for measuring urinary beta-blockers. Our results demonstrated that the levels of both urinary timolol and carteolol exceeded the minimum reporting levels (MRL) following acute and chronic administration. The highest levels of urinary timolol and carteolol observed in the present study were 255.7 and 923.8 ng/ml, respectively. Regarding the acute administration of timolol ophthalmic drop, 26.19 (11/42) of urine samples were detected with timolol higher than the MRL in timed and random sampling. In contrast, the acute administration of carteolol ophthalmic drops made the carteolol levels higher than the MRL among most urine samples. On the other hand, 36.36% (4/11) of urine samples were detected with beta-blockers higher than the MRL during the chronic administration of timolol and carteolol ophthalmic drops. In the context of receiving ophthalmic beta-blocker medications, the present study has highlighted the potential risk of constituting an AAF in specific sports and suggests strengthening athletes' awareness of therapeutic use exemptions.

Deep Learning Approach in Image Diagnosis of Pseudomonas Keratitis.

This investigation aimed to explore deep learning (DL) models' potential for diagnosing Pseudomonas keratitis using external eye images. In the retrospective research, the images of bacterial keratitis (BK, n = 929), classified as Pseudomonas (n = 618) and non-Pseudomonas (n = 311) keratitis, were collected. Eight DL algorithms, including ResNet50, DenseNet121, ResNeXt50, SE-ResNet50, and EfficientNets B0 to B3, were adopted as backbone models to train and obtain the best ensemble 2-, 3-, 4-, and 5-DL models. Five-fold cross-validation was used to determine the ability of single and ensemble models to diagnose Pseudomonas keratitis. The EfficientNet B2 model had the highest accuracy (71.2%) of the eight single-DL models, while the best ensemble 4-DL model showed the highest accuracy (72.1%) among the ensemble models. However, no statistical difference was shown in the area under the receiver operating characteristic curve and diagnostic accuracy among these single-DL models and among the four best ensemble models. As a proof of concept, the DL approach, via external eye photos, could assist in identifying Pseudomonas keratitis from BK patients. All the best ensemble models can enhance the performance of constituent DL models in diagnosing Pseudomonas keratitis, but the enhancement effect appears to be limited.

Tear Proteomics Approach to Distinguishing Primary from Secondary Sjögren's Syndrome for Dry Eye Patients with Long-Term Instillation of Eyedrops.

The diagnosis and monitoring of Sjögren syndrome (SS) is often difficult, requiring a multidisciplinary approach with invasive procedures. Our aim is to elucidate the tear protein alterations of dry eye disease (DED) with primary SS (pSS) and secondary SS (sSS) with the long-term instillation of eyedrops. We collected clinical demographics and tear fluid (TF) samples from DED patients with no autoimmune diseases (non-SS-DED), pSS-DED, and sSS-DED patients, followed by TF screening with tandem mass tagging-labeling gel-free proteomics assay. Bioinformatic analysis via Ingenuity Pathway Analysis was used to identify functional pathways and interacting networks. Validation of candidate proteins with enzyme-linked immunosorbent assay on the tear samples was done. The top functional pathways of the two comparisons (sSS-DED vs. pSS-DED and sSS-DED vs. non-SS-DED) were both associated with inflammation and stress-related signaling. After constructing an interaction network model with the selected candidate proteins, five proteins were identified. A Disintegrin and Metalloproteinase domain-containing protein 10 (ADAM10) was found to be an important candidate biomarker in all groups, followed by epidermal growth factor (EGF) in TF. This study revealed novel DED markers, ADAM10 and EGF, in differentiating between primary and secondary SS patients from tears by in-depth proteomic analysis.

Diagnosing Fungal Keratitis and Simultaneously Identifying Fusarium and Aspergillus Keratitis with a Dot Hybridization Array.

Fungal keratitis (FK) is one of the most common microbial keratitis, which often leads to poor prognosis as a result of delayed diagnosis. Several studies implied that early differentiation of the two major FK, Fusarium and Aspergillus keratitis, could be helpful in selecting effective anti-fungal regimens. Therefore, a novel dot hybridization array (DHA) was developed to diagnose FK and differentiate Fusarium and Aspergillus keratitis in this study. One hundred forty-six corneal scrapes obtained from one hundred forty-six subjects impressed with clinically suspected FK were used to evaluate the performance of the DHA. Among these patients, 107 (73.3%) patients had actual FK confirmed by culture and DNA sequencing. We found that the DHA had 93.5% sensitivity and 97.4% specificity in diagnosing FK. In addition, this array had 93.2% sensitivity and 93.8% specificity in diagnosing Fusarium keratitis, as well as 83.3% sensitivity and 100% specificity in diagnosing Aspergillus keratitis. Furthermore, it had 83.9% sensitivity and 100% specificity in identifying Fusarium solani keratitis. Thus, this newly developed DHA will be beneficial to earlier diagnosis, more precise treatment, and improve prognosis of FK, by minimizing medical refractory events and surgical needs.

Comparisons of deep learning algorithms for diagnosing bacterial keratitis via external eye photographs.

Bacterial keratitis (BK), a painful and fulminant bacterial infection of the cornea, is the most common type of vision-threatening infectious keratitis (IK). A rapid clinical diagnosis by an ophthalmologist may often help prevent BK patients from progression to corneal melting or even perforation, but many rural areas cannot afford an ophthalmologist. Thanks to the rapid development of deep learning (DL) algorithms, artificial intelligence via image could provide an immediate screening and recommendation for patients with red and painful eyes. Therefore, this study aims to elucidate the potentials of different DL algorithms for diagnosing BK via external eye photos. External eye photos of clinically suspected IK were consecutively collected from five referral centers. The candidate DL frameworks, including ResNet50, ResNeXt50, DenseNet121, SE-ResNet50, EfficientNets B0, B1, B2, and B3, were trained to recognize BK from the photo toward the target with the greatest area under the receiver operating characteristic curve (AUROC). Via five-cross validation, EfficientNet B3 showed the most excellent average AUROC, in which the average percentage of sensitivity, specificity, positive predictive value, and negative predictive value was 74, 64, 77, and 61. There was no statistical difference in diagnostic accuracy and AUROC between any two of these DL frameworks. The diagnostic accuracy of these models (ranged from 69 to 72%) is comparable to that of the ophthalmologist (66% to 74%). Therefore, all these models are promising tools for diagnosing BK in first-line medical care units without ophthalmologists.

Long-term survival of cultivated oral mucosal epithelial cells in human cornea: generating cell sheets using an animal product-free culture protocol.

Previously, we reported a collagenase-based, animal product-free protocol for cultivated oral mucosal epithelial cell sheets for transplantation (COMET). Here, we reported the long-term outcomes of first 2 clinical cases. A 27-year-old man suffered from thermal burn, which resulted in symblepharon of lower fornix OD. COMET was performed, and the cornea remained clear with few peripheral NV and no more symblepharon 34 months postoperatively. Another 42-year-old man suffered from severe alkaline burn OD. He underwent COMET, followed by corneal transplantation half a year later. A biopsy taken two years after COMET showed stratified epithelium positive for keratin 4, 13, and 3 in the suprabasal layer. Staining for p63 and p75NTR was both positive in the basal layer. The graft remained clear up to post-OP 4 years. Our study confirmed the long-term survival of the transplanted OMECs, suggesting that collagenase-based spheroidal suspension culture is a promising technique for COMET.Trial registration ClinicalTrials.gov, ClinicalTrials.gov ID: NCT03943797 Registered 9 May 2019-Retrospectively registered, https://clinicaltrials.gov/ct2/show/NCT03943797 .

Simultaneously Monitoring Whole Corneal Injury with Corneal Optical Density and Thickness in Patients Undergoing Cataract Surgery.

To pursue the least corneal implication during cataract surgery, this study aimed to monitor corneal wound injury after cataract surgery with a novel method. The prospective cohort study involved thirty-two patients, who were assessed by a Scheimpflug tomography AxL® (Oculus GmbH, Wetzlar, Germany) via the following two kinds of indices: whole corneal optical density (COD) and corneal thickness (CT), two weeks before and one month after cataract surgery. The results of the COD revealed that corneal annuli 0.0-2.0 mm and 2.0-6.0 mm, and the average and maximal values at the incisional site significantly increased postoperatively. Also, the anterior and central stroma of 0.0-2.0 mm, and all three depths of 2.0-6.0 mm, increased remarkably after the operation. For the CT, all ranges of diameters plus incisional sites showed significant increases postoperatively. Furthermore, we analyzed the differences (delta) of COD and CT between pre- and post-operation, and found significant correlations between the delta of COD and the delta of CT, regarding annuli 0.0-2.0 mm, 2.0-6.0 mm, and 6.0-10.0 mm, but no correlation at the incisional site, with either average density or maximal density, was detected. We concluded that whole COD and CT, especially at the central zones of the cornea (annulus < 6 mm), are both valuable parameters in the assessment of corneal damage post-cataract surgery, and are independent indices at the incisional site.

Using Slit-Lamp Images for Deep Learning-Based Identification of Bacterial and Fungal Keratitis: Model Development and Validation with Different Convolutional Neural Networks.

In this study, we aimed to develop a deep learning model for identifying bacterial keratitis (BK) and fungal keratitis (FK) by using slit-lamp images. We retrospectively collected slit-lamp images of patients with culture-proven microbial keratitis between 1 January 2010 and 31 December 2019 from two medical centers in Taiwan. We constructed a deep learning algorithm consisting of a segmentation model for cropping cornea images and a classification model that applies different convolutional neural networks (CNNs) to differentiate between FK and BK. The CNNs included DenseNet121, DenseNet161, DenseNet169, DenseNet201, EfficientNetB3, InceptionV3, ResNet101, and ResNet50. The model performance was evaluated and presented as the area under the curve (AUC) of the receiver operating characteristic curves. A gradient-weighted class activation mapping technique was used to plot the heat map of the model. By using 1330 images from 580 patients, the deep learning algorithm achieved the highest average accuracy of 80.0%. Using different CNNs, the diagnostic accuracy for BK ranged from 79.6% to 95.9%, and that for FK ranged from 26.3% to 65.8%. The CNN of DenseNet161 showed the best model performance, with an AUC of 0.85 for both BK and FK. The heat maps revealed that the model was able to identify the corneal infiltrations. The model showed a better diagnostic accuracy than the previously reported diagnostic performance of both general ophthalmologists and corneal specialists.

Grading reliability of the tear film viscosity examination.

PURPOSE: To assess the reliability of a tear film (TF) viscosity video grading system. METHODS: Thirty-four dynamic TF viscosity videos were obtained by a clinically available TF analyzer and objectively sorted according to the movement speed of three arbitrary reflective light particles. A 4-grade system was constructed on a specially designed window for simultaneous comparison with the three standard videos. Two masked graders were invited to grade these videos under a randomized procedure. Observer reliabilities were determined by Spearman's correlation analysis and Bland-Altman agreement analysis. RESULTS: For this four-grade system, the intra-observer correlation was very strong in the two graders (ρ = 0.96 and 0.82; both P < 0.0001). However, the inter-observer correlation showed moderate strength in normal playback speed (ρ = 0.53, P = 0.002 and ρ = 0.52, P = 0.003 for 1st and 2nd gradings, respectively). In slower playback videos, the inter-observer correlation of the two graders was higher (ρ = 0.70 and P < 0.0001) when reduced to 0.8-times playback speed. Moreover, the 0.8-times mode was also significantly better than normal playback mode (P = 0.0204) in terms of inter-observer agreement. CONCLUSIONS: The dynamic 4-grade system has an excellent intra-observer reliability and a good inter-observer reliability under 0.8-times playback speed. The grading system established in this study provides a promising solution for rapidly determining the level of TF viscosity.

Clinical Characteristics of Presenile Cataract: Change over 10 Years in Southern Taiwan.

PURPOSE: The purpose of this study is to investigate the clinical characteristics of presenile cataract and compare that to ten years ago in southern Taiwan. METHODS: The subjects who received cataract surgeries aged 30 to 54 years were recruited in Kaohsiung Chang Gung Memorial Hospital during September 2015 and August 2016. Patients with uveitis or those who received combined cataract surgeries were excluded. Retrospective chart review was performed in this study. RESULTS: A total number of 2439 cataract surgeries were performed, and 302 (12.38%) eyes were having presenile cataract. Mean age was 47.55 ± 5.64 years old, and mean axial length was 26.00 ± 2.89 mm. Among 302 presenile cataract eyes, the leading cause was high myopia (defined as mean axial length ≥ 26 mm, 47.02%), followed by diabetes mellitus (26.82%). In types of lens opacity analysis, 67.55% of the patients were nuclear sclerosis dominant. Compared to the previous study 10 years ago, the leading cause of presenile cataract changed from idiopathy to high myopia, whereas the lens opacity types changed from posterior subcapsular opacity dominant to nuclear sclerosis dominant. CONCLUSIONS: High myopia has become the most important clinical characteristic associated with presenile cataract in a myopia epidemic area, and the most common type of lens opacity was nuclear sclerosis. With the increasing prevalence of high myopia, we should pay more attention to the management of presenile cataracts in high myopes to avoid complications.

Axial length of cataract eyes: a comparison of two cohorts over a span of 10 years apart.

BACKGROUND: To assess the associations of axial length with age-related cataract within a span of 10 years in an Asian population in southern Taiwan. METHODS: A retrospective cohort study examined 960 adults who underwent cataract surgery at the Kaohsiung Chang Gung Memorial Hospital in year 2008 and year 2018. Axial length was assessed with the ultrasound biometry and/or the Zeiss IOLMaster. Eyes with prior blunt eye trauma or had underwent vitrectomy operations were excluded. The significance of the changes in axial length between the two cohorts was determined after performing age-matched analyses. Due to utilization of ultrasound biometry and/or Zeiss IOLMaster, axial length corrections with our mean difference in measurement results, which were similar to previous studies on comparison between the two measurement tools, were carried out. RESULTS: Axial length showed an age-related elongation in 10-year cross-sectional data, from a mean of 23.65 ± 1.80 mm in year 2008 to a mean of 24.30 ± 1.90 in year 2018 (p = 0.003). Patients with high myopia (axial length > 26 mm) increased significantly over the 10-year period from 8.1 to 16 % (p < 0.001). A birth cohort effect on axial length was evident as the axial lengths of year 2008 cohort were shorter than the 2018 cohort when they were in the same operation age group. In particular, persons born after the 1960s demonstrated a predominant increase in axial length in both cohorts. CONCLUSIONS: Our study confirms a trend in increase of axial myopia, especially high myopias, over the 10-year period. A novel finding of this study was discovering a birth cohort effect on axial length, especially in persons born after the 1960s in southern Taiwan.

The severity of visual field defect and the defective angles of localized retinal nerve fiber layer using a novel approach.

BACKGROUND/PURPOSE: A modified novel analytical method for evaluating photographic retinal nerve fiber layer defect (RNFLD) was developed for the purpose of estimating the severity of early glaucomatous eyes. This new method was then compared with the original method, as described below, in relation to the visual field defect severity. METHODS: The clinical records of 47 glaucomatous eyes with qualified photographs were obtained from a prior retrospective study. The reference point of the angle was set at the center of the optic disc center (COD group) and the central vessel trunk (COV group). Angle α was the angular width between the fovea and the proximity of RNFLD, while angle ß (+c) was the sum of the angular width(s) of localized RNFLD. The correlation between the visual field parameters and the angles of RNFLD in the above two groups were analyzed. RESULTS: There were significant differences in angle α (p = 0.001) and angle ß (+c) (p = 0.016) between the two groups. Angle ß (+c) showed significant correlation with MD (p = 0.012), PSD (p = 0.035), VFI (p = 0.042) and MD (p = 0.016), PSD (p = 0.035), VFI (p = 0.031) in the COD and COV group, respectively. No correlation was found between angle α and the presence of central scotoma in both groups. CONCLUSION: Our novel method was more convenient in a clinical setting and noninferior to the original method.

Tear Proteomics Study of Dry Eye Disease: Which Eye Do You Adopt as the Representative Eye for the Study?

Most studies about dry eye disease (DED) chose unilateral eye for investigation and drew conclusions based on monocular results, whereas most studies involving tear proteomics were based on the results of pooling tears from a group of DED patients. Patients with DED were consecutively enrolled for binocular clinical tests, tear biochemical markers of DED, and tear proteome. We found that bilateral eyes of DED patients may have similar but different ocular surface performance and tear proteome. Most ocular surface homeostatic markers and tear biomarkers were not significantly different in the bilateral eyes of DED subjects, and most clinical parameters and tear biomarkers were correlated significantly between bilateral eyes. However, discrepant binocular presentation in the markers of ocular surface homeostasis and the associations with tear proteins suggested that one eye's performance cannot represent that of the other eye or both eyes. Therefore, in studies for elucidating tear film homeostasis of DED, we may lose some important messages hidden in the fellow eye if we collected clinical and proteomic data only from a unilateral eye. For mechanistic studies, it is recommended that researchers collect tear samples from the eye with more severe DED under sensitive criteria for identifying the more severe eye and evaluating the tear biochemical and proteomic markers with binocular concordance drawn in prior binocular studies.

A Genomic Approach to Investigating Ocular Surface Microorganisms: Monitoring Core Microbiota on Eyelid Margin with a Dot hybridization Assay.

A sound ocular surface microbiota has been recognized as a part of ocular surface health following a growing body of evidence from next-generation sequencing technique and metagenomic analysis. However, even from the perspective of contemporary precision medicine, it is difficult to directly apply these new technologies to clinical practice. Therefore, we proposed a model based on dot hybridization assay (DHA) to bridge conventional culture with a metagenomic approach in investigating and monitoring ocular surface microbiota. Endophthalmitis, mostly caused by bacterial infection, is the most severe complication of many intraocular surgeries, such as cataract surgery. Hazardous microorganisms hiding and proliferating in the ocular surface microbiota not only increase the risk of endophthalmitis but also jeopardize the effectiveness of the preoperative aseptic procedure and postoperative topical antibiotics. The DHA model enables the simultaneous assessment of bacterial bioburden, detection of target pathogens and microorganisms, and surveillance of methicillin/oxacillin resistance gene mecA in the ocular surface microbiota. This assay revealed heavier bacterial bioburden in men, compatible with a higher risk of endophthalmitis in male patients who underwent cataract surgery. No occurrence of endophthalmitis for these patients was compatible with non-hazardous microorganisms identified by specific dots for target pathogens. Moreover, the mecA dot detected oxacillin-resistant strains, of which culture failed to isolate. Therefore, the DHA model could provide an alternative genomic approach to investigate and monitor ocular surface microorganisms in clinical practice nowadays.

Analysis of tear film spatial instability for pediatric myopia under treatment.

In Taiwan, the prevalence of myopia in children between 6 and 18 years old is over 80%, and high myopia accounts for over 20%, which turned out to be in the leading place worldwide. Orthokeratology and low-dose atropine are proven treatments to reduce myopia progression, though the potential corneal disturbances remain an issue in young populations. The alteration of the tear film is widely discussed but there is no consensus to date, so we aim to investigate the tear film spatial instability in children with myopia control using atropine or orthokeratology. Thirty-eight treatment-naïve participants and 126 myopic children under treatments were enrolled. The ocular surface homeostasis, spatial distribution of tear break-up, and high-order aberrations (HOAs) of the corneal surface were assessed. We found out that myopic children treated with either atropine or orthokeratology showed ocular surface homeostasis similar to that in treatment-naïve children. Nevertheless, children treated with orthokeratology presented higher HOAs (p < 0.00001) and a tendency of the first tear break-up zone at the inner half of the cornea (p = 0.04). This unique spatial instability of the tear film associated with myopia treatment might provide a more focused way of monitoring the pediatric tear film instability.

A deep learning approach in diagnosing fungal keratitis based on corneal photographs.

Fungal keratitis (FK) is the most devastating and vision-threatening microbial keratitis, but clinical diagnosis a great challenge. This study aimed to develop and verify a deep learning (DL)-based corneal photograph model for diagnosing FK. Corneal photos of laboratory-confirmed microbial keratitis were consecutively collected from a single referral center. A DL framework with DenseNet architecture was used to automatically recognize FK from the photo. The diagnoses of FK via corneal photograph for comparing DL-based models were made in the Expert and NCS-Oph group through a majority decision of three non-corneal specialty ophthalmologist and three corneal specialists, respectively. The average percentage of sensitivity, specificity, positive predictive value, and negative predictive value was approximately 71, 68, 60, and 78. The sensitivity was higher than that of the NCS-Oph (52%, P < .01), whereas the specificity was lower than that of the NCS-Oph (83%, P < .01). The average accuracy of around 70% was comparable with that of the NCS-Oph. Therefore, the sensitive DL-based diagnostic model is a promising tool for improving first-line medical care at rural area in early identification of FK.