Keratoconus Classification Using Multimodal Imaging Strategy

Data fusion improves the accuracy and robustness of diagnostic models by combining different types of information. This study presents a multimodal framework for keratoconus classification. It uses numeric and textual features from Pentacam reports, extracted with OCR. These are combined with corneal topographic images processed by a dual-branch deep neural network. The method was tested on 2,924 labeled Pentacam scans. Of these, 1,900 were used for training and 1,024 for testing. Scans were labeled as normal, suspicious, or keratoconus. Results show that combining image and text features improves classification. Deep learning accuracy rose from 96.78% to 98.34%. SVM improved from 93.35% to 95.60%. LDA increased from 92.85% to 94.80%, and KNN from 90.50% to 93.94%. These gains, up to 1.56% for deep learning and 3.44% for KNN, show the value of multimodal data for more accurate keratoconus diagnosis.