Slit-lamp-based SUN grading aims to provide a standardized assessment of anterior uveitis. Although this method has widely demonstrated its utility in the clinical practice it is subject to biases and shortcomings. SUN grading is operator specific and its reliability and validity rely on observer experience. Furthermore, other technical limitations like slit lamp resolution threshold, beam size calibration, light source, and patient collaboration may jeopardize the quality of its findings.
As consistency, accuracy, and reliability are key criteria for standardized results, a computational, image-based approach to the categorization of anterior uveitis could entail clear benefits. In our study, our image analysis software was able to demonstrate significant correlation between the number of particles detected by AS-OCT imaging and clinical slit-lamp-based SUN grading while providing useful continuous and numerical information (see
Fig. 3) and supporting its potential as a novel application for assessment for intraocular inflammation in the clinical setting. AS-OCT based results are comparable and independent of the operator experience and presents clear advantages when compared with the current gold standard. Thus, image analyses of AS-OCT, have the potential to overcome the biases inherent to SUN grading, improving clinical decision making.
AI and machine learning in diagnostic medical imaging is currently receiving substantial evaluation in different medical fields
15 and together with the latest advances in medical imaging has the potential to revolutionize medical diagnostics. Automated and OCT-based tools for future practice in clinical ophthalmology is currently under development.
10,16–18 Previous groups have been able to demonstrate prosperous achievements in regards to OCT-based uveitis assessment. Two previous groups have been shown correlation between SUN graded slit-lamp assessment and OCT-based uveitis assessment.
19,20 Although these studies show the potential of OCT-based cell counting, they are still subjected to many of the biases as slit-lamp based assessment, namely observer variability. Furthermore, manual assessment would require the experience of a trained ophthalmologist/OCT user limiting availability to clinical settings. Automated OCT-based approaches have also been investigated by previous groups. Two groups were able to demonstrate significant correlation between their automated algorithms and clinical SUN grading assessment.
18,21 Both Li et al. and Sharma et al. developed image-analysis algorithms. Both groups base their algorithms on segmentation of portions of the AC and identifying hyperreflective spots, which is then used as a representative for the whole AC. This method might share some limitation with slit-lamp based assessment as cell concentration in AC might vary geographically and the correct threshold might vary between scans.
Accordingly, we developed a neural network-based AI system which analyses the AC in its entirety and efficiently detects particles based on AS-OCT scans. Previous studies have shown very good correlation between manual and automatic segmentation of AC particles. The presented study and AI particle segmentation model replicate these findings and additionally evaluates OCT-based automatic and manual AC inflammation grading.
Although AI-based clinical diagnosis entails important ethical implications, it has the potential to outperform the clinician
22 in providing unbiased and reproducible clinical results, becoming an important tool for future ophthalmology. Technical developments in optical tomography have provided ophthalmologist with high-definition, real time images of the anterior segment structures.
23 AI models for analysis of these large collections of data, giving an accurate understanding of the underlying principles at hand.
Although quantification of anterior chamber inflammation based on AS-OCT shows promising potential, the current OCT resolution does not allow to distinguish between cells of similar sizes or cell types, like inflammatory cells and pigmentary cells. In the future, OCT based particle size thresholding could help in cell characterization of cell groups, such as erythrocytes, pigmentary cells, and immunologic cells that have different size ranges.
21
In cases of very low inflammation, where only a small number of cells are to be found in the chamber, our current method might share some limitations with SUN grading, stemming from the fact that both methods analyze only a portion of the anterior chamber. We are currently investigating incrementing the number of B-scans which could possibly bring a more accurate count.
AI-based diagnostics is still relatively novel and in a developmental phase. Its results and utilities require comprehensive and extensive validation work in the coming years before it can be utilized independently. Until then, it can be regarded as a novel and experimental part of the diagnostic toolbox in the in ophthalmic practice.
During development, several custom non-AI particle detection algorithms were explored and evaluated in comparison with AI-based algorithms. Particle segmentation models shared similar limitations in regard to patient cornea artifacts detected as particles. We accommodated for this by slightly decreasing the model's segmented AC area and thereby reducing artifacts present in the particle detection model's analyzed area. The models’ performances were evaluated resulting in the choice of an AI-based model. The AI model's performance was evaluated in a validation subset (85 images) using Jaccard (0.99) and Dice (0.98) coefficients. Furthermore, we evaluated the number of particles missed by the AI-based model's particle detection algorithm due to the smaller analyzed area. In 8.7% of the validation subset, in particular, heavily inflamed eyes, one or more particles were found outside the model's segmented AC. We speculate that such missing particles become inconsequential as the software outputs density metrics (particle/area).
The AI particle detection algorithm's specificity and sensitivity metrics were calculated as 0.68 and 0.78, respectively. We believe that false positive predictions are heavily impacted by signal-to-noise ratio, and as OCT technology and image resolution improves, less background noise would be present in scans potentially increasing the software's specificity.
Spearman's rank correlation was used as the measure for correlation between clinical grading and automatic and manual particle segmentation via OCT imaging. Spearman's rank correlation is a nonparametric measure of rank correlation. In our study, we aim to assess the monotonic relationship between a categorical-ordinal variable (SUN grading's 5 increasing gradings) and a numerical-continuous variable (OCT/AI particle count/density). Therefore, it is distinct from agreement estimates (such as Kappa agreement) which requires exclusively ordinal data.
24 As such, we have used correlation measure in lieu of agreement coefficient. Furthermore, Spearman's rank correlation is the most commonly metric used in other groups assessing similar relationship between SUN grading and OCT particle count listed in the systematic review by Liu et al.
10 Furthermore, we chose to fit a nonlinear exponential regression model on association between clinical grading and automatic and manual particle segmentation based on the intrinsic nature of the exponential step-wise increments in SUN gradings. However, correlation between the two continuous variables of automatic and manual particle segmentation was measured using Pearson's (linear) correlation metric. Consequently, we chose a linear regression fit to model the association between automatic and manual segmentation.
Our study and the proposed AI system present themselves with a number of limitations. First, our automatic system is at the current moment not able to analyze multiple slides from a single patient’s OCT scan. We are currently implementing this feature and a prototype of the program is already in the works. Second, we have analyzed OCT scans of a small number of enrolled patient subjects. We believe an increment in the number of enrolled patients is exceptionally necessary to evaluate the effectiveness of the AI model's ability to distinguish between cell number among grades. Consequently, we are currently expanding our study to include a larger patient cohort. In addition, the patient population is skewed with a higher proportion of patients with less severe degrees of inflammation in AC. We believe these data prompt further similar investigations in a more inflammatory diverse patient population. Third, the AI cell counter was not able to demonstrate significant differences between all groups. Our AI software was able to record significant differences between observed particles between all SUN gradings, except grading 0 and 0.5+ (p = 0.0617), and 1+ and 2+ (p = 0.8855). We believe the overlapping range could be due to a number of reasons. The intrinsic nature of SUN grading's nonlinear and noncontinuous scale might assume a role in the overlapping. Furthermore, previous studies of the SUN interobserver agreement show a relatively low agreement among uveitis experts, showing a tendency of discrepancy (especially) in the low spectrum of inflammation.
11 Multiple research groups that have conducted similar OCT to clinical SUN assessments have reported observed hyper-reflective particles in OCT images of eyes clinically classified as SUN 0 by slit-lamp examination by an experienced ophthalmologist
19,25 suggesting the cells might be undetected in cases of very low inflammation during slit-lamp examination. This presented study has observed and reproduced similar events. Furthermore, we were able to record higher correlation and fit of model (Pearson's r = 0.9948, R
2 = 0.9897) between automatic and manual segmentation than both automatic segmentation and SUN grading (Pearson's r = 0.7077, R
2 = 0.8846) and manual segmentation and SUN grading (Pearson's r = 0.8264, R
2 = 0.8698). Similar studies have suggested anterior segment OCT imaging as a promising technique for grading AC cells.
18,20,25 Together with our findings, they might suggest OCT imagining offers a more precise evaluation of anterior inflammation.
As we are determined to formulate a tool for ophthalmologists in a clinical framework, we seek to design the model in accordance with the challenges faced in such settings. Consequently, as diagnosis in low-grade SUN is more clinically relevant, particularly in regard to early treatment, we would like to improve the current practice. That being the case, we believe it is of the highest priority to include a more complete (cellular) validation of the inflammatory conditions of the patients’ eye (e.g. flow cytometric analysis of aqueous humour), of which is the current ongoing effort of our group in the AI validation process.
This study suggests that image analysis of AS-OCT in combination with AI could possibly be used to detect and quantify anterior chamber inflammation in eyes with clinically graded anterior uveitis. We show that the number of particles and particle density correlates with clinical SUN grading (
Fig. 3), whereas the AI observed particles are independent of clinical SUN grading (
Fig. 4). The study highlights the possibilities of the methods in providing a robust, fast, noninvasive, and observer-independent assessment of patients with different grades of anterior uveitis, and its potential to become a key tool in the eye clinic in the near future.