 |
Ocular
Microbiology and Immunology Group
Back
to OMIG Main Page
< Previous | 2021
Agenda and Abstracts | Next >
2021 OMIG Abstract
Deep Learning for Image-Based Differentiation of Bacterial and Fungal Keratitis
Travis K Redd1, N. Venkatesh Prajna2, Muthiah Srinivasan2, Nisha Acharya3, Thomas M Lietman3,
Jeremy D Keenan3, J Peter Campbell1, Xubo Song1
1Casey Eye Institute, Oregon Health & Science University, Portland, Oregon; 2Aravind Eye Hospital,
Madurai, India; 3Francis I. Proctor Foundation, University of California San Francisco,
San Francisco, California
Purpose: To develop and evaluate convolutional neural networks (CNNs) for automated image-based distinction of bacterial and fungal keratitis.
Methods: We trained 5 state-of-the-art CNNs using transfer learning from ImageNet pre-trained models (ResNet152V2, DenseNet201, MobileNetV2, Xception, and VGG19) to differentiate bacterial from fungal ulcers using standardized handheld photographs collected during previous multicenter clinical trials at Aravind (SCUT, MUTT I, and MUTT II). A total of 446 images were used for training and validation of each CNN. Only images from a single site (Madurai) were included in the training set to avoid label leakage. The models were evaluated on a hold-out test set of 84 images collected from all study sites (Madurai, Coimbatore, Tirunelveli, Pondicherry, Bharaptur, and Lumbini) to assess generalization error. The area under the receiver operating curve (AUC) was compared against 12 expert cornea specialists from South India performing the same task on the test set. Models were developed using the Keras framework in Tensorflow 2.0.
Results: MobileNet and DenseNet achieved the highest AUCs on the hold-out test set (0.84), followed by ResNet (0.83), Xception (0.77), and VGG (0.76). The best-performing CNNs achieved a higher AUC than all 12 human experts (individual human AUCs ranged from 0.59 to 0.79). The ensemble prediction of all 5 CNNs achieved an AUC of 0.85, which was significantly higher than the ensemble prediction of all 12 humans (0.77, P=0.01).
Conclusions: Convolutional neural networks are able to achieve superhuman performance in binary classification of bacterial and fungal corneal ulcers using external photographs. The most successful network architectures in this case were MobileNetV2 and DenseNet201. Further research is ongoing to determine generalization error on new datasets, assess classification accuracy for other forms of keratitis including viral, parasitic, and culture-negative infections, and incorporate elements of the history and clinical exam into multivariate prediction models for the cause of infection. Future iterations of these computer vision models may have utility for guiding empiric antimicrobial therapy in the absence of microbiologic results.
Disclosure: N
Support: NIH (K12EY027720 and P30EY10572) and unrestricted departmental funding from Research to Prevent Blindness
< Previous | 2021
Agenda and Abstracts | Next >
|
|
