One of AI’s most promising applications: new diagnostic algorithms
The rapid evolution of artificial intelligence (AI) has given rise to several technological innovations, including new applicable solutions for ophthalmic clinical practice. Machine learning is a subset of AI that utilises mathematical algorithms and computer processing power to analyse and find underlying patterns embedded in large volumes of data, thus augmenting human interpretation of clinical information.1 This approach facilitates the classification tasks by combining and integrating multimodal data from different sources, which can improve the overall diagnostic efficiency.2 One of the most promising applications of this technology is the development of new diagnostic algorithms for identifying ectatic corneal diseases.3
Recently, there has been a growing clinical interest in detecting mild forms of corneal ectasia. Indeed, the diffusion of refractive surgery has driven the need to identify subclinical forms at an elevated risk for iatrogenic progressive ectasia after laser vision correction procedures. Furthermore, an early diagnosis of ectatic corneal diseases has become increasingly relevant due to the introduction of new treatments, such as collagen cross-linking, that can slow or halt the progression of the disease.2
The main challenge in diagnosing early keratoconus is that slit-lamp examination is typically unremarkable, and the distinctive signs of advanced ectasia, such as Vogt striae and Fleischer ring, are usually absent. As such, other diagnostic techniques have been proposed over the years to identify the condition before the occurrence of significant visual loss.
Placido-disk-based corneal topography allows for the characterisation of the anterior corneal surface, enabling the detection of mild to moderate forms of keratoconus. Conversely, corneal tomography can evaluate the whole cornea obtaining information from the anterior and posterior surfaces. Since posterior elevation is often the first sign of the disease, tomographic analysis is considered more sensitive than Placido-disk topography in identifying keratoconus.4,5 The Belin/Ambrósio Enhanced Ectasia Display (BAD-D) available on the OCULUS Pentacam® is a comprehensive screening tool combining elevation-based and pachymetric corneal evaluation. The BAD-D is considered one of the Scheimpflug-based indices with the most outstanding predictive accuracy to detect ectatic corneal diseases.6
Nevertheless, the diagnostic accuracy of corneal tomography alone can be unsatisfactory in very mild or sub clinical cases. In clinical studies, eyes with typical front surface topography in patients that have a clinical form of ectasia in the fellow eye are frequently used as models of sub clinical ectasia.
The sensitivity for tomographic indices can be reduced in these cases7 as the corneal shape is frequently not yet altered. However, performing laser refractive surgery in patients with subclinical or forme fruste keratoconus would still likely result in an iatrogenic ectasia.
New technologies, including epithelial mapping and measurement of corneal biomechanics, have been developed to overcome limitations and increase diagnostic sensitivity.
The measurement of corneal biomechanics has recently emerged as a novel promising technology for identifying sub clinical ectasia.8 In recent years, the general interest in corneal biomechanics among clinicians and researchers has rapidly increased. Figure 1 shows all scientific publications on corneal biomechanics indexed on PubMed and grouped by year of publication. Of 2812 total articles published since the 1950s, 1974 articles (70.2%) were published in the past decade, and 943 (33.5%) in the past five years.
The Corvis ST® (OCULUS Optikgeräte) is a noncontact tonometer that employs an ultra high speed Scheimpflug camera to record in vivo the cornea’s response to an air puff-induced deformation (Figure 2). Several studies demonstrated that keratoconic corneas show abnormal deformation amplitudes compared with healthy ones.9-11
While first-generation biomechanical parameters showed a relatively low accuracy for detecting mild forms of ectatic corneal diseases, novel parameters such as the Corvis Biomechanical Index (CBI) were shown to correctly classify eyes with clinical keratoconus in 98.8% of cases, with 98.4% specificity and 100% sensitivity.12
The tomographic/biomechanical index (TBI) is a parameter that integrates corneal morphology from the Pentacam with biomechanical parameters from the Corvis ST®. The index demonstrated the ability to detect subclinical or forme fruste keratoconus accurately. This was proven in eyes with normal topography and asymmetric ectasia in the fellow eye.13 Thus, the TBI is now programmed and included in the commercial OCULUS software (Figure 3).
Figure 3 shows a very relevant case example of a patient with normal tomography who developed ectasia after PRK. The TBI was unavailable at the time of surgery and the BAD-D was within the normal range. However, a retrospective analysis of the preoperative TBI output indicated a high ectasia risk.
Ambrósio et al recently published a cross-sectional, case-control, multicentre retrospective study in which the TBI was optimised using AI technology and a larger data set including 3886 eyes from 3412 patients.14 The optimisation employed a random forest algorithm, a supervised learning method that operates by constructing many decision trees on various subsets of the given data set. The novel algorithm had an area under the curve (AUC) of receiver operating characteristic curves of 0.945 for detecting subclinical ectasia, which was significantly higher than the AUC of the previous version of TBI (0.899; P < .0001). Figure 4 illustrates the clinical relevance of this improvement in a patient who developed ectasia after SMILE. Although preoperative findings of the TBI version 1 were only slightly suspicious for ectasia, the novel version of the TBI was abnormal.
The authors concluded that: “AI optimisation to integrate
Scheimpflug-based corneal tomography and biomechanical assessments augments accuracy for ectasia detection.” These advances may improve clinical decision-making in screening cases at risk of iatrogenic ectasia after laser vision correction. AI may further improve accuracy by integrating data from multimodal imaging technologies, such as epithelial thickness mapping and ocular wavefront data.
Renato Ambrósio Jr, MD, PhD, is president of the International Society of Refractive Surgery (ISRS) and a professor of ophthalmology at Federal University of the State of Rio de Janeiro (UNIRIO).