Dual-Modality Grading of Keratoconus Severity Based on Corneal Topography and Clinical Indicators

Keratoconus (KC) is a blinding eye disease characterized by corneal dilation, thinning, and conical protrusion. Currently, most studies just focus on KC screening and use only one modality. In this paper, we utilize two modalities of data, corneal topography and clinical indicators, to grade KC severity as normal, mild, moderate, and severe, rather than screening. Considering the data characteristics of each modality, we model their data in a targeted manner. For corneal topography, a Global Feature Extraction (GFE) block based on Self-Attention is designed as a parallel branch to obtain global information. In addition, a Feature Fusion (FF) module is proposed to better fuse local and global information. For clinical indicators, we utilize a tree model to grade KC and remove redundant information through feature selection. Finally, a trainable fusion model is used to make the final decision. Experimental results show that the proposed method outperforms all the competing methods and achieves 94.47% for weighted recall, improving the accuracy of KC grading effectively.