Table 5 Comparison of the proposed OL-CDACNN and previous methods
OL-CDACNN | Kuo et al. [24] | Gu et al. [25] | Redd et al. [26] | Ghosh et al. [27] | |
|---|---|---|---|---|---|
What was solved | ● Accurately locate the region of Conj_Cor associated with keratitis lesions from original slit-lamp images ● Provide automatic screening of keratitis, other cornea abnormalities, and normal cornea based on slit-lamp images, with a better performance than conventional CNNs ●Substantial slit-lamp images from multiple-centers verified the effectiveness and generalizability of the model | ● Provide a promising tool for improving first‑line medical care at rural area in early identification of FK | ● The DL algorithm may be useful for computer‑assisted corneal disease diagnosis with excellent performance | ● Develop computer vision models for image-based differentiation of bacterial and fungal corneal ulcers | ● Apply DL algorithms for rapidly discriminating between FK and BK |
Methods | SSD for accurately locating the region of Conj_Cor Deep attention modules and cost-sensitive method were applied to enhance the expression of keratitis-related features and address the imbalanced data problem | DenseNet for discerning FK | Inception-v3 with multi-task multi-label classification layer | MobileNet, DenseNet, ResNet, VGG were compared and applied to distinguish BK and FK | CNN and ensemble learning for distinguishing FK and BK |
Outcomes (95% CI) | |||||
Accuracy (%) | 98.9–100 | 69.4(63.9–74.5) | - | - | 83.0 |
Sensitivity (%) | 96.9–100 | 71.1(62.1–78.6) | - | - | 77.0(0.81–0.83) |
Specificity (%) | 99.2–100 | 68.4(61.1–74.9) | - | - | - |
AUC | 0.997–1 | 0.650 | 0.930(0.904–0.952) | 0.86(0.78–0.93) | 0.904 |
Dataset | 12,407 slit-lamp images collected from Ningbo Eye Hospital, Jiangdong Eye Hospital, Ningbo Ophthalmic Center, and Zhejiang Eye Hospital | 288 slit-lamp images collected from Kaohsiung Chang Gung Memorial Hospital | 5325 ocular surface slit-lamp images collected from Shanghai Eye, Ear, Nose, and Troat Hospital and the Afliated Hospital of Guizhou Medical University | 980 slit-lamp images obtained from one study site within the Aravind Eye Care System in South India | 2167 slit-lamp anterior segment images from 194 patients collected from Ophthalmology Department |