Deep Learning-Based Conjunctival Melanoma Detection Using Ocular Surface Images

Abstract

The human eye could be affected with conjunctival melanoma, which indicates a fatal malignant growth of the eye. Being a very rare disease, there exists a lack of related data in the literature. Also, very few studies performed deep learning-based conjunctival melanoma detection from the ocular surface images. In response to the research gap, we created a more enriched and well-curated dataset validated by medical experts for conjunctival melanoma detection using deep learning. Furthermore, we have made our dataset available on Kaggle. In the present work, we utilized a total of seven pretrained deep learning-based classification models, that are, DenseNet161, DenseNet201, EfficientNet_B7, GoogLeNet, ResNet18, ResNet50, and ResNet152. We outperformed the prior literature in terms of assessment metrics, including different performance parameters like accuracy, precision, sensitivity, F1-score, specificity, confusion matrix, ROCcurve, and AUROC, with better improvement achieved in multi-label classification. The best AUROC and overall accuracy were found to be around 1.00 and 99.51%, respectively, for binary classification whereas these were around 0.99 and 94.42%, respectively, in case of multi-label classification, from the best performing model, EfficientNet_B7. The deep learning applications in medical images require visual interpretation of the features prioritized in decision-making, and a Grad-CAM-based visualization is added in this study to prove the effectiveness of the best-performing model. The research conducted here may make it simpler to identify conjunctival lesions using state-of-the-art deep learning models more meticulously in the future.