W e thank Shanthi and associates for their interest in our recently published article on optical coherence tomography (OCT) radiomics and machine learning for the detection of forme fruste keratoconus (FFKC). We appreciate the opportunity to clarify several points regarding the interpretation and potential clinical translation of our findings.
Much of this discussion concerns the distinction between initial model evaluation and readiness for clinical application. Our study was designed as a proof-of-concept investigation to determine whether quantitative radiomic features extracted from routine OCT images contain discriminative information relevant to FFKC detection. It was not intended to establish a stand-alone screening test for routine clinical decision-making. Similarly, our comparison of random forest, C5.0, and XGBoost was not intended to identify a universally superior classifier. Rather, it was designed to assess whether commonly used machine learning algorithms could consistently leverage OCT-derived texture features for FFKC detection. We agree that the absence of statistically significant differences among classifiers should not be interpreted as evidence of formal equivalence. Instead, the consistent diagnostic performance observed across classifiers supports the presence of clinically relevant discriminative information within the radiomic feature set. Larger prospective cohorts with external validation are needed to refine performance estimates and assess generalizability.
The correspondents also note that defining FFKC as the clinically normal fellow eye of a patient with manifest keratoconus represents a high-risk population rather than a general refractive surgery screening population. We agree. This definition is widely used in early keratoconus research and identifies a clinically meaningful group in which subtle abnormalities are particularly challenging to detect. , We further agree that the positive predictive value is strongly influenced by disease prevalence, a point that was discussed in detail in the Discussion section of our original article. A lower positive predictive value under a low-prevalence assumption does not negate the discriminative ability of the model; rather, it highlights that OCT radiomics should be interpreted within the intended clinical context and used as a complementary risk stratification tool alongside tomography, biomechanical assessment, and clinical judgment.
Finally, we agree that radiomic features should not be overinterpreted as direct histologic markers. Nevertheless, emerging OCT-based evidence supports the biological plausibility that OCT image features may contain information related to stromal collagen organization. For example, recent coregistered OCT and second-harmonic generation imaging demonstrated spatial correspondence between OCT scattering patterns and collagen-related microstructural features, supporting the feasibility of extracting collagen organization–related information from OCT images. However, OCT-derived radiomic features remain image-based quantitative descriptors rather than direct molecular or histologic measurements. Their distribution along meridians corresponding to regions commonly affected in early keratoconus supports biological plausibility, but further validation using longitudinal follow-up, biomechanical measurements, high-resolution imaging, or histologic correlation is needed to clarify their structural basis.
In summary, the points raised by the correspondents are important considerations for clinical translation and are consistent with the limitations discussed in our original article. They do not alter the central conclusion that OCT-derived radiomic features provide measurable information that may complement existing approaches for the early detection of FFKC.
CRediT authorship contribution statement
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
