Deep Transfer Learning for Improved Detection of Keratoconus using Corneal Topographic Maps

Abstract Clinical keratoconus (KCN) detection is a challenging and time-consuming task. In the diagnosis process, ophthalmologists must revise demographic clinical ophthalmic examinations. The latter include slit-lamb, corneal topographic maps, Pentacam indices (PI). We propose an Ensemble of Deep Transfer Learning (EDTL) based on maps. consider four pretrained networks, SqueezeNet (SqN), AlexNet (AN), ShuffleNet (SfN), MobileNet-v2 (MN), fine-tune them dataset KCN normal cases, each including also PI classifier. Then, our EDTL method combines output probabilities five classifiers to obtain decision fusion probabilities. Individually, classifier achieved 93.1% accuracy, whereas deep reached classification accuracies over 90% only in isolated cases. Overall, average accuracy networks maps ranged from 86% (SfN) 89.9% (AN). ensemble increased values ranging (92.2% 93.1%) for SqN (93.1% 94.8%) AN. Including ensemble-specific combinations maps’ 98.3%. Moreover, visualization first learner filters Grad-CAMs confirmed that had learned relevant features. This study shows potential creating ensembles fine-tuned with transfer learning strategy as it resulted improved while showing learnable agree knowledge. step further towards deployment computer-assisted system help confirm perform fast accurate treatment.