Multimodal learning for clinically consistent RGP fitting in keratoconus

Abstract Keratoconus is a progressive corneal disorder characterized by highly heterogeneous corneal morphology, which makes rigid gas permeable (RGP) lens fitting strongly dependent on clinician experience and iterative trial processes. This procedure is often time-consuming and may lead to patient discomfort. Existing artificial intelligence approaches typically rely on either corneal imaging or clinical parameters alone, limiting their ability to capture the complex coupling relationships underlying lens fitting. In this study, we propose a unified multimodal deep learning framework, termed CME-Net, to jointly predict key RGP fitting parameters, including base curve (BC), lens diameter (Dia), and fitting strategy. The framework integrates boundary artifact suppression preprocessing with a dual-branch feature extraction network and employs structured multimodal feature fusion to explicitly model interactions between corneal topography and clinical indicators. A unified output head enables collaborative multi-task inference. By integrating complementary information from multiple data sources within a unified framework, the proposed method enables a more comprehensive characterization of corneal morphology and its clinical implications. The proposed method was developed and evaluated using a labeled analytic cohort of 368 keratoconus eyes with complete usable reference labels, derived from a retrospective cohort of 515 eyes collected at the Eye Center of the Second Affiliated Hospital, Zhejiang University School of Medicine. Under the held-out evaluation protocol, CME-Net achieved mean absolute errors of 0.2178 mm for BC and 0.1250 mm for Dia. Compared with the strongest unimodal baseline, the proposed framework reduced the Total MAE by approximately 15%. These results suggest that multimodal integration can enhance the accuracy and consistency of RGP lens fitting prediction. The proposed framework may provide supportive initial fitting guidance for personalized RGP fitting in keratoconus, while clinician oversight remains necessary, particularly for boundary or high-error cases.