Image-Based Differentiation of Bacterial and Fungal Keratitis Using Deep Convolutional Neural Networks

Abstract

Purpose: Develop computer vision models for image-based differentiation of bacterial and fungal corneal ulcers and compare their performance against human experts.

Design: Cross-sectional comparison of diagnostic performance.

Participants: Patients with acute, culture-proven bacterial or fungal keratitis from 4 centers in South India.

Methods: Five convolutional neural networks (CNNs) were trained using images from handheld cameras collected from patients with culture-proven corneal ulcers in South India recruited as part of clinical trials conducted between 2006 and 2015. Their performance was evaluated on 2 hold-out test sets (1 single center and 1 multicenter) from South India. Twelve local expert cornea specialists performed remote interpretation of the images in the multicenter test set to enable direct comparison against CNN performance.

Main outcome measures: Area under the receiver operating characteristic curve (AUC) individually and for each group collectively (i.e., CNN ensemble and human ensemble).

Results: The best-performing CNN architecture was MobileNet, which attained an AUC of 0.86 on the single-center test set (other CNNs range, 0.68-0.84) and 0.83 on the multicenter test set (other CNNs range, 0.75-0.83). Expert human AUCs on the multicenter test set ranged from 0.42 to 0.79. The CNN ensemble achieved a statistically significantly higher AUC (0.84) than the human ensemble (0.76; P < 0.01). CNNs showed relatively higher accuracy for fungal (81%) versus bacterial (75%) ulcers, whereas humans showed relatively higher accuracy for bacterial (88%) versus fungal (56%) ulcers. An ensemble of the best-performing CNN and best-performing human achieved the highest AUC of 0.87, although this was not statistically significantly higher than the best CNN (0.83; P = 0.17) or best human (0.79; P = 0.09).

Conclusions: Computer vision models achieved superhuman performance in identifying the underlying infectious cause of corneal ulcers compared with cornea specialists. The best-performing model, MobileNet, attained an AUC of 0.83 to 0.86 without any additional clinical or historical information. These findings suggest the potential for future implementation of these models to enable earlier directed antimicrobial therapy in the management of infectious keratitis, which may improve visual outcomes. Additional studies are ongoing to incorporate clinical history and expert opinion into predictive models.

Keywords: AUC, area under the receiver operating characteristic curve; Artificial intelligence; Bacterial keratitis; CI, confidence interval; CNN, convolutional neural network; Computer vision; Convolutional neural networks; Corneal ulcer; Deep learning; Fungal keratitis; Infectious keratitis; MUTT, Mycotic Ulcer Treatment Trials; SCUT, Steroids for Corneal Ulcers Trial.

Figure 1

Receiver operating characteristic curves of 5 deep convolutional neural network (CNN) models on a single-center external testing set consisting of 100 corneal ulcer images (50% fungal, 50% bacterial) from Coimbatore. The performance of the CNN ensemble is also depicted, representing the average output of all 5 models for each image. MobileNet demonstrated the highest performance among the model architectures tested. Ninety-five percent confidence intervals are depicted next to each area under the receiver operating characteristic curve (AUC) estimate.

Figure 2

Receiver operating characteristic curves of the 5 convolutional neural network (CNN) models and 12 human graders on a multicenter testing set consisting of 80 corneal ulcer images (48 bacterial, 32 fungal). The CNN ensemble demonstrated a statistically significantly higher area under the receiver operating characteristic curve (AUC; 0.84) than the human ensemble (AUC, 0.76; P < 0.01). Ninety-five percent confidence intervals are depicted next to each AUC estimate.

Figure 3

Confusion matrices of the convolutional neural network (CNN) and human ensembles, with prediction categories (bacterial or fungal) assigned according to the threshold defined by Youden’s index. Percent values indicate column proportions. For example, the CNN ensemble showed 81% accuracy for identifying fungal infections and 75% accuracy for bacterial infections.

Figure 4

Receiver operating characteristic curves of the best-performing convolutional neural network (CNN) (MobileNet), best-performing human (grader 1), and the ensemble of the 2 (CNN plus human ensemble). The area under the receiver operating characteristic curve (AUC) of the ensemble was 0.87, compared with an AUC of 0.79 for grader 1 (P = 0.09) and AUC of 0.83 for MobileNet (P = 0.17). Ninety-five percent confidence intervals are depicted next to each AUC estimate.

Figure 5

Representative gradient class activation heatmaps of the images on which the best-performing convolutional neural network model (MobileNet) achieved highest agreement (top 5 fungal and top 5 bacterial) between model prediction and ground truth and lowest agreement (bottom 5 fungal and bottom 5 bacterial). Red coloration indicates regions of the input image that conferred the highest influence on the model’s prediction. Superimposed on each image is the percent agreement between the model’s prediction and the ground truth (for fungal images, P(fungal) × 100%; for bacterial images: 1 – P(fungal) × 100%). Adjacent to each heatmap is the raw image input to the model. The model tended to perform well when focusing attention on the corneal infiltrate, with relatively worse performance when areas of attention strayed from the cornea or when tested on images with quality limitations including overexposure (O), underexposure (U), or eccentric fixation (E).