Microbial keratitis (MK) remains a leading cause of blindness. Although the clinical diagnosis of MK can be relatively straightforward, identifying and objectively measuring clinical change can be difficult. Developments in anterior segment optical coherence tomography (AS-OCT) have improved the assessment and measurement of corneal lesions compared with slit-lamp biomicroscopy.¹ Artificial intelligence (AI) plays a significant and increasing role in diagnosing and monitoring conditions in all medical specialties.24 Within ophthalmology, deep learning models have been created that are able to objectively characterise conditions such as cystoid macular oedema,⁵ diabetic retinopathy⁶ and glaucomatous disc cupping.⁷ There are, therefore, many potential benefits from having an automated system that can objectively assess and monitor corneal lesions such as in MK.

One of the limitations of using slit-lamp biomicroscopy to assess disease, such as infiltrate size, is that the assessment is often limited to the anterior component of the lesion. As such, clinicians rely on measuring ulcer and anterior infiltrate size, which may not reflect changes to the posterior extent of the lesions. In contrast, although not all of the lesions may be characterised by AS-OCT, it does enable an increased depth of measurement than apparent on slit-lamp biomicroscopy, so that the extent of the lesion can be better appreciated. This information may help the clinician to better assess response to treatment and whether treatment modification is needed. AS-OCT has become a standard clinical tool; however, it is difficult for the clinician to compute and compare due to the large number of sections in a scan. Manually measuring lesions on AS-OCT is time-consuming, having to go through multiple radial sections and then comparing change over time periods is challenging. This also introduces potential measurement error. Having a computerised algorithm that can rapidly and objectively measure and compare lesions across multiple scans, particularly to quantify and to determine change, would be potentially helpful in a patient’s management. This type of measurement and its compilation and comparison is well suited to AI.

Several deep learning models have been proposed in the past 5 years to assist in the diagnosis of MK. Most of these approaches applied convolutional neural networks (CNNs) to corneal images, often from slit-lamp or external photographs, rather than from OCT scans. For instance, Apushkin et al trained five CNN models on handheld camera images to differentiate fungal keratitis from bacterial keratitis.⁸ Their best-performing model, MobileNet, achieved an area under curve (AUC) of 0.86, surpassing expert ophthalmologists (combined AUC of 0.84 vs experts’ 0.76).⁹ Similarly, Li et al⁷ developed a deep learning system using DenseNet121 to detect keratitis versus other corneal conditions.^{10 11} Their model, tested on smartphone images, achieved an AUC of 0.967, with 91.9% sensitivity and 96.9% specificity for identifying keratitis. In a related study, Wang et al, explored different image crops and sources: using InceptionV3 on slit-lamp images, they achieved an AUC of 0.9588, while performance dropped to 0.8529 on uncropped smartphone photos.¹² This highlighted the significant impact of image quality and context, where even after preprocessing, mobile images performed poorly, likely due to variations in lighting and focus conditions. Notably, none of these models used OCT scans for training or validation. The aim of this study was to develop an AI model and to evaluate the effectiveness of this model in assessing and monitoring change in MK using AS-OCT.

The AS-OCT images were initially captured with Cassia (Tomey GmbH, Nuremberg, Germany) and then on Anterion (Heidelberg Engineering GmbH, Heidelberg, Germany) devices. All lesions on the AS-OCT images were traced by two corneal specialists and used as the ground truth to train and evaluate the AI model. Annotation between corneal specialists showed agreement in >98% of images. When agreement was not made between observers, a third corneal specialist (SK) decided on lesion demarcation. Examples of annotations of each class are presented in online supplemental figure 1.

For the initial training phase, our AI model was trained on AS-OCT scans from healthy eyes and patients with corneal lesions due to corneal dystrophies and MK. This was then repeated on patients presenting with clinically suspected MK at presentation and at 1 week, to assess progression. All patients with clinically suspected MK received a topical fluoroquinolone at presentation and treatment modified according to the microbiological diagnosis and clinical response.

The proposed AI segmentation framework consists of two stages: AI-based segmentation and post-processing, as illustrated in online supplemental figure 2. In the AI segmentation stage, we incorporated two deep learning-based segmenters—a global and a local segmenter—to ensure high accuracy in identifying key anatomical structures and disease markers. These segmenters generated initial segmentations of MK regions, which are then refined in the post-processing step to produce the final segmentation result. The framework is described in detail below. It was not deemed appropriate to involve patients or the public in the design or conduct of our research.

This study evaluated the effectiveness of AI in discerning and measuring corneal lesions in MK and measuring change using AS-OCT. We found that AI is an effective tool in identifying MK using AS-OCT in the majority of cases, with a high sensitivity of 93%. Implementation of our module in hospital clinics could significantly improve the workflow of doctors managing patients with MK. Integration with commonly used OCT machines may allow the software to rapidly characterise the MK lesion and provide the clinician with an approximate estimate of total size of the infiltrate in three dimensions. The goal would then be able to provide the clinician with a comparison tool for infiltrate size over time, which would be more accurate than traditional measurements of epithelial defect and infiltrate size at the slit lamp. One benefit of the application of our algorithm is that it would reduce interobserver variability on the same lesions, which may be particularly beneficial in clinics where patients are being reviewed by different clinicians at each visit.

Cases that were missed by the AI model were typically small peripheral infiltrates near the limbus. The reason that these cases were missed was due to a mechanism incorporated into the AI model where peripheral abnormalities near the iris are excluded. This was created because, in previous iterations of the model, we were detecting high numbers of artefacts in this anatomical area due to hyperreflectivity at the limbus. Future iterations of the model will aim to reduce the size of this exclusion zone to optimise peripheral lesion detection, while maintaining low false positive rates.

For the training set, problems arose due to the presence of normal AS-OCT slices in our dataset that did not have an MK infiltrate. Combining normal and abnormal slices into a single segmenter set, however, would lead to an imbalanced training set, as normal slices dominated the other classes. This imbalance can adversely affect the model’s ability to learn and accurately classify the minority classes, such as the MK infiltrate regions. Second, we faced a pixel-level imbalanced class problem within the MK infiltrate classes. The regions corresponding to higher and lower likelihoods of being an MK infiltrate were relatively small compared with the overall image space. These smaller regions, however, contained crucial details to help differentiate between different stages or types of MK infiltrates. If, however, we were to have trained a single segmenter for all the classes together, the model might then have struggled to allocate sufficient attention and resources to accurately segment these small regions. This would have led the model to prioritise the larger and more dominant background class, leading to suboptimal performance in segmenting the MK infiltrate classes. By developing two separate segmenters, however, we addressed these imbalanced class challenges more effectively. The global segmenter focused on excluding general objects/artefacts that resembled the cornea and it handled the sample-level imbalanced class problem by distinguishing between the background and the other classes. The local segmenter specifically targeted the MK infiltrate regions and mitigated the pixel-level imbalanced class problem by prioritising the segmentation of these smaller regions. Overall, the decision to develop two separate segmenters for the total classes was driven by the need to address sample-level and pixel-level imbalanced class problems, ensuring accurate and reliable segmentation results.

During the inference phase, test AS-OCT scans underwent simultaneous processing through both global and local segmenters developed in this study. The goal was to obtain refined segmentation maps by subtracting the segmentation results of the global segmenter from the segmentation results of the local segmenter. This subtraction operation aimed to remove false positive regions that might have been incorrectly classified by the local segmenter. By subtracting the segmentation maps, we focused on the regions that were classified as higher and lower likelihood of being an MK infiltrate by the local segmenter while excluding the general objects/artefacts that resembled the cornea but were not relevant to MK diagnosis. This process helped to refine the segmentation results and enhance the accuracy and reliability of identifying MK infiltrate regions. Other limitations in our study are the absence of MK due to Acanthamoeba and other pathogens. Our framework did not evaluate the algorithms’ ability to work equally well in bacterial and fungal infections; however, this does represent a gap in our knowledge which could be addressed with a future comparative study.

In summary, our methodology involved the development of two AI-based segmenters and ACE for post-processing. The first global segmenter focused on excluding general objects/artefacts that resemble the cornea, while the second local segmenter aimed to classify regions into higher likelihood MK infiltrates, lower likelihood MK infiltrates, or other non-MK areas. These segmenters could serve as valuable tools for diagnosing MK and monitoring its progression.

A limitation of the model is that it does not have external validation. This is particularly relevant as OCT image characteristics vary based on device and settings. We aimed to address this by including images from multiple OCT devices, but to fully assess the validity of the module, it would require validation on external test sets when these become available. The dataset generated through our research would be made available to any other researcher on request.

The specificity of the AI model for the detection of lesions was perfect and no normal scans were flagged as having MK. This is the result of multiple iterations and deep learning cycles aiming to minimise artefact detection. This does not mean that AI can distinguish MK from other causes of corneal lesions or the type of MK. The current model has been trained not to detect any abnormalities in the anterior chamber, eyelids or at the limbus on the AS-OCT. These areas previously were large sources of artefact error. The tradeoff for this is that some smaller peripheral lesions can be missed. This was a sacrifice that we deemed appropriate, as our goal was to develop an AI software that can be used to track the change of MK lesions based on corneal sections, rather than identify small peripheral infiltrates (which was achieved with high accuracy on previous model iterations).

Although the AI model was effective in detecting MK lesions, the accuracy of which it traced these lesions was limited, as is reflected when looking at the progression (change) in lesions in severe cases. This is evident when there is severe disorganisation of the anterior segment and when treatment also includes wearing contact lenses. We used 656 single-sliced manually traced AS-OCT images to train our AI model, which is fewer than used in deep learning models investigating macular oedema (934 OCT images),⁵ diabetic retinopathy (1748 fundus photograph images)⁶ and glaucomatous optic neuropathy (48 116 fundus images).⁷ As we continue to trace and enter more AS-OCT images of severe cases of MK into the model in the future, we anticipate that the accuracy with which the AI measures corneal lesions in severe cases of MK, that is, those with associated abnormalities in the anterior segment, will improve. Previous studies^{5 6} have relied on coding algorithms to enhance the efficiency of image tracing by specialists, which may improve the efficiency of data entry into the deep learning model for clinicians generating similar AI programmes in the future.

The diagnosis of MK is one that is made by the clinician aided by slit-lamp biomicroscopy. The value in AI analysis of MK lesions lies in the ability to detect subtle corneal changes in large lesions that are not easily apparent on slit-lamp examination, which may indicate that the lesion is improving or worsening. This has the potential to help direct treatment or further investigations in such cases. Our study demonstrated that AI can accurately identify change using only corneal sections without anterior chamber or other information in a number (76%) of severe cases compared with specialist OCT interpretation. In this study, we only interpreted corneal changes on the infiltrate, as our AI model was not trained to identify hypopyon size or the presence of anterior chamber cells. Other AI models have shown the ability to quantify the presence of anterior chamber cells¹⁵ and future iterations will aim to also include these factors.