Abstract:Abstract To demonstrate the identification of corneal diseases using a novel deep learning algorithm. A novel hierarchical deep learning network, which is composed of a family of multi-task multi-label learning classifiers representing different levels of eye diseases derived from a predefined hierarchical eye disease taxonomy was designed. Next, we proposed a multi-level eye disease-guided loss function to learn the fine-grained variability of eye diseases features. The proposed algorithm was trained end-to-end directly using 5,325 ocular surface images from a retrospective dataset. Finally, the algorithm’s performance was tested against 10 ophthalmologists in a prospective clinic-based dataset with 510 outpatients newly enrolled with diseases of infectious keratitis, non-infectious keratitis, corneal dystrophy or degeneration, and corneal neoplasm. The area under the ROC curve of the algorithm for each corneal disease type was over 0.910 and in general it had sensitivity and specificity similar to or better than the average values of all ophthalmologists. Confusion matrices revealed similarities in misclassification between human experts and the algorithm. In addition, our algorithm outperformed over all four previous reported methods in identified corneal diseases. The proposed algorithm may be useful for computer-assisted corneal disease diagnosis.

What problem does this paper attempt to address?

This paper aims to solve the problem of corneal disease identification. Specifically, the author has developed a new deep - learning algorithm for identifying corneal diseases from ocular surface photos taken by slit - lamp microscopes. This research mainly focuses on the following corneal diseases: infectious keratitis, non - infectious keratitis, corneal dystrophy or degenerative diseases, and corneal tumors. ### Research Background Corneal diseases are one of the leading causes of reversible blindness worldwide, second only to cataracts. It is estimated that about 6.8 million people in India and about 3.2 million people in China are blind in at least one eye due to corneal diseases. Early diagnosis and appropriate treatment can prevent vision loss. However, traditional corneal disease diagnosis relies on doctors' experience, which is time - consuming and prone to inter - observer variation. Automated grading of medical images can reduce doctors' workload, improve the efficiency and repeatability of screening programs, and improve patients' prognosis through early detection and treatment. ### Research Methods 1. **Data Set**: The research used 5,325 slit - lamp photos of the ocular surface from two major ophthalmic centers in China, including 870 photos of normal subjects and 4,455 photos of patients with five tested eye diseases. In addition, a prospective clinical data set containing 510 photos of newly admitted patients was also collected. 2. **Image Grading and Reference Standards**: 32 ophthalmologists participated in the grading of retrospective database images. After training, only 20 ophthalmologists who met the standard of κ value above 0.75 were selected as qualified graders. Each photo was graded by two ophthalmologists after face - to - face communication. 3. **Deep - Learning Algorithm**: The research designed a hierarchical deep - learning network, which consists of multiple multi - task multi - label classifiers, representing different levels of eye disease classification. The algorithm uses the Inception - v3 convolutional neural network as the backbone network and replaces the final classification layer to achieve multi - task multi - label classification. A multi - level eye - disease - guided loss function was used during the training process, including two levels of losses: \[ \text{Loss}=\alpha\times\text{level\_1\_loss}+(1 - \alpha)\times\text{level\_2\_loss} \] where \(\alpha\) is a weighting parameter used to balance the two levels of losses. 4. **Data Augmentation**: Due to data imbalance, the research adopted multiple data augmentation methods (such as flipping, color jittering, etc.) to balance data of each category. ### Research Results 1. **Algorithm Performance**: On the test data set, the area under the ROC curve (AUC) of the algorithm for each corneal disease exceeded 0.910. Overall, the sensitivity and specificity of the algorithm were comparable to or better than the average level of 10 ophthalmologists. 2. **Confusion Matrix**: The confusion matrix shows that the algorithm and human experts show similarities in misclassification, especially having difficulties in distinguishing between infectious and non - infectious keratitis. 3. **Comparison with Existing Methods**: Compared with four other reported methods (ResNet34, DenseNet, Inception - v3, and Ensemble), this algorithm performs better in identifying corneal diseases. 4. **Heat Map Visualization**: The research also used Gradient - Weighted Class Activation Mapping (Grad - CAM) to generate heat maps, helping doctors identify potential corneal lesion areas, thereby increasing the credibility and interpretability of the diagnosis. ### Conclusion The deep - learning algorithm developed in this research performs excellently in identifying corneal diseases, with high accuracy and reliability. By automatically processing ocular surface images, this algorithm is expected to assist in the diagnosis of corneal diseases in clinical practice and improve the efficiency and accuracy of diagnosis.

Deep learning for identifying corneal diseases from ocular surface slit-lamp photographs

[Deep learning based lesion detection from anterior segment optical coherence tomography images and its application in the diagnosis of keratoconus].

Prediction of Different Eye Diseases Based on Fundus Photography via Deep Transfer Learning

Deep learning for multi-type infectious keratitis diagnosis: A nationwide, cross-sectional, multicenter study

A Novel Hierarchical Deep Learning Framework for Diagnosing Multiple Visual Impairment Diseases in the Clinical Environment

Ophthalmic Disease Detection via Deep Learning With a Novel Mixture Loss Function

Deep-Ocular: Improved Transfer Learning Architecture Using Self-Attention and Dense Layers for Recognition of Ocular Diseases

Using Slit-Lamp Images for Deep Learning-Based Identification of Bacterial and Fungal Keratitis: Model Development and Validation with Different Convolutional Neural Networks

Feasibility assessment of infectious keratitis depicted on slit-lamp and smartphone photographs using deep learning

Automatic segmentation of corneal deposits from corneal stromal dystrophy images via deep learning

Determination of probability of causative pathogen in infectious keratitis using deep learning algorithm of slit-lamp images

Automatic Screening and Identifying Myopic Maculopathy on Optical Coherence Tomography Images Using Deep Learning

Differentiation of Active Corneal Infections from Healed Scars Using Deep Learning

Preventing corneal blindness caused by keratitis using artificial intelligence

Deep learning-based classification of infectious keratitis on slit-lamp images

Deep Learning and Glaucoma Specialists: The Relative Importance of Optic Disc Features to Predict Glaucoma Referral in Fundus Photos

Deep learning model for extensive smartphone-based diagnosis and triage of cataracts and multiple corneal diseases

Deep learning in early keratoconus/corneal ectasia

Development and validation of a deep learning system to screen vision-threatening conditions in high myopia using optical coherence tomography images

Multi-label ocular disease classification with a dense correlation deep neural network

Application of a Deep Learning System in Pterygium Grading and Further Prediction of Recurrence with Slit Lamp Photographs