Deep-learning models can now classify infectious keratitis from slit-lamp images in seconds, with a systematic review published in Cureus reporting that 26 studies published from 2020 onward — 16 on keratitis, 10 on conjunctivitis — met eligibility criteria after screening 800 records down from PubMed, Scopus, and Google Scholar, with the last search conducted in September 2026. The stakes behind that dataset count: corneal opacity ranks fifth among global causes of blindness, and corneal ulceration plus ocular trauma cause an estimated 1.5 to 2 million new cases of monocular blindness each year, more than 90% of them in developing countries — what the review calls a “silent epidemic.”
Why a Culture Plate That Takes Days Loses to a Neural Network
The clinical bottleneck is well defined. In suspected infectious keratitis, the reference standard remains corneal scraping for microscopy and culture after slit-lamp examination. Culture results take several days, require laboratory facilities often unavailable where the burden is highest, and come back negative in a substantial proportion of cases. Clinicians therefore start empirical treatment on clinical appearance alone — and distinguishing bacterial from fungal or Acanthamoeba keratitis by eye is difficult and varies between observers.
The failure mode is concrete. A patient with fungal or Acanthamoeba keratitis may first receive standard antibacterial drops, which are ineffective against those organisms and let the infection progress, sometimes rapidly, to irreversible scarring or perforation. Delays and diagnostic errors raise the risk of urgent corneal transplantation, the review notes.
Convolutional neural networks attack exactly this gap. The architecture works like face recognition software: successive layers detect edges first, then more complex patterns such as corneal infiltrates and ulcer morphology, after training on thousands of labeled slit-lamp or smartphone images. Classification of a new image takes seconds.
The Numbers: 800 Records, 26 Studies, One Imbalance
The review’s methodology, reported in line with PRISMA 2020 guidelines, shows how young this field is. Of 800 identified records (186 PubMed, 274 Scopus, 340 Google Scholar), 176 duplicates were removed, 624 were screened, 542 were excluded at title-and-abstract stage, and 82 full texts were assessed. Sixty-two fell out — 28 with no AI component, 22 without diagnostic performance or outcome data, 12 review articles or editorials. Twenty studies survived the database search; six more were added on update, for 26 total.
That 16-to-10 split between keratitis and conjunctivitis studies is itself a finding. The review states plainly that the evidence for conjunctivitis is considerably smaller, and that imbalance shapes the whole literature. Conjunctivitis is usually diagnosed from history and examination, with lab testing reserved for severe or chronic cases — yet adenovirus causes up to 90% of viral cases, and the different causes are hard to distinguish clinically, which drives unnecessary antibiotic prescribing. The only identified application of AI to clinical records in conjunctivitis was a single study using natural language processing and machine learning to characterize acute conjunctivitis patients from electronic health records.
What the Models Actually Do — and What the Metrics Mean
The review’s terminology table doubles as a buyer’s guide for evaluating these tools. Performance claims hinge on a handful of measures that are easy to conflate:
- AUC — overall ability to distinguish disease from non-disease, from 0.5 (coin flip) to 1.0 (perfect)
- Sensitivity vs. specificity — catching sick eyes versus not alarming healthy ones
- AUPRC — the precision-recall balance, which unlike AUC depends on how common the disease is in the dataset
- Calibration — whether predicted probabilities match observed outcomes, the metric most vendor claims skip
Two engineering techniques recur across the included studies. Transfer learning adapts a model pretrained on a large general image dataset to the medical task, cutting the amount of clinical data needed — critical when labeled corneal ulcers are scarce. Ensemble learning combines several models’ predictions to improve reliability. The review’s caution is overfitting: a model that learns noise specific to its training images performs worse on new data, which is why external validation on data from different centers and populations is the dividing line between a demo and a deployable tool.
Where This Fits in the Wider AI-in-Medicine Fight
Because both conditions present as a red eye and are assessed with the same anterior-segment imaging, the highest-value task is triage: separating usually benign conjunctivitis from sight-threatening keratitis in frontline and tele-ophthalmology settings. The epidemiology explains the urgency. Reported incidence of microbial keratitis ranges from fewer than 10 per 100,000 person-years in some high-income settings to 799 per 100,000 person-years in Nepal; in the United States it is 11 to 27.6 per 100,000, with contact lens-related ulcerative keratitis rising. Smartphone-based imaging models are the logical distribution channel for exactly those low-resource settings where culture labs don’t exist.
The review also extends beyond images. It covers chatbots and large language models — deep-learning systems trained on large text corpora that generate human-like responses — and natural language processing alongside the image classifiers, positioning the field’s newer non-imaging applications against the more mature vision work. Earlier reviews had covered AI in corneal disease broadly, and prior systematic reviews and meta-analyses had evaluated deep-learning diagnostic accuracy for infectious keratitis and AI-based pathogen detection; this one’s stated addition is treating both conditions together, including the non-imaging applications and the ethical and clinical considerations around their use.
The Honest Caveats the Review Itself Flags
The authors acknowledge a search limitation: keratoconjunctivitis, trachoma, and allergic conjunctivitis were not separate search terms, so some studies on those conditions may not have been captured. And because the included studies addressed different tasks — disease detection, pathogen classification, chatbot response evaluation — with different imaging modalities and metrics, the review states that pooling was not appropriate. No head-to-head benchmark table across models exists in this literature; anyone claiming otherwise is overselling.
The field’s trajectory is legible in the review’s own constraint: no date restriction was applied to the search, yet every included study was published from 2020 onward. The science is younger than the smartphone in the average clinician’s pocket — and the next validation step, external testing across populations, is what will decide whether these classifiers reach the clinics that need them most.
