Researchers Compare Deep Learning (DL) Algorithms For Diagnosing Bacterial Keratitis via External Eye Photographs

350
Source: https://www.nature.com/articles/s41598-021-03572-6.pdf

Viral keratitis, bacterial keratitis, fungal keratitis, and parasitic keratitis are all types of infectious keratitis. Bacterial Keratitis (BK) is a kind of Infectious Keratitis that is one of the most frequent and vision-threatening. Contact lens wear is the most prevalent risk factor for Bacterial Keratitis (BK), and it is becoming increasingly popular around the world for a variety of reasons, including exercise, cosmesis, and myopia management.

BK is substantially more fulminant and painful in the clinical course than other Infectious Keratitis(s). A delayed diagnosis of Bacterial Keratitis (BK) can result in large-area corneal ulcerations, melting, and even perforation if not treated.

In the case of Infectious Keratitis, timely detection and treatment of BK are vital goals. However, in many rural areas, the availability of ophthalmologists does not meet the demand for the rapid diagnosis of Bacterial Keratitis (BK).

In subtropical climates, Bacterial Keratitis (BK) is the most prevalent Infectious Keratitis, and it is the primary cause of corneal scarring and vision loss. Some researchers recently revealed that deep learning (DL) based image diagnosis had a high diagnostic rate for Bacterial Keratitis (BK). Their findings demonstrated that different DL algorithms had varying diagnostic performance, with DenseNet and ResNet50 showing the best outcomes. However, these researchers chose different niches to promote their deep learning (DL) models for diagnosing BK, making the actual performance of DL models in these studies incomparable.

A team of researchers intended to explore the faithful performance of several deep learning (DL) models in diagnosing BK using an external eye photo in another recent study influenced by past developments. As a result, this study examined the presentation of deep learning (DL) models at the picture level, using a single external eye photo and no extra preprocessing procedures like image manipulation or segmentation.

From June 1, 2007, to May 31, 2019, the researchers gathered external eye images and evaluated medical data from patients with clinically suspected IK who presented to Chang Gung Memorial Hospital (CGMH) branches. External eye photography was performed by certified ophthalmic technicians utilizing a camera-mounted slit lamp biomicroscope according to individual standard procedures in CGMH branches. In the following studies, one photo was taken using white light illumination for each patient.

The input images were all uniformly enlarged to 224 x 224 pixels, which is the industry standard for deep learning (DL) algorithms. The RGB values of each pixel in a photograph were normalized in a range of 0 to 1. This study developed a deep learning (DL) model for diagnosing BK using an exterior eye snapshot. The training images were used to train a deep learning (DL) model for distinguishing BK from non-BK shots, while the validation images were utilized to evaluate a trained model’s performance. Each diagnostic model was trained following randomization using the appropriate DL method toward the goal with the largest area under the receiver operating characteristic curve (AUROC).

The researchers used the validation data to empirically tweak the hyperparameters of each model, such as learning rate, number of dense blocks, growth rate, and batch size, to develop the best model. For a visual description of these deep learning (DL) models, Grad-CAM++ was used. All of the tests were run on NVIDIA GeForce RTX 1080 GPUs, and the models were written in PyTorch.

Each DL diagnosis model’s sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were determined using five-fold cross-validation. In a nutshell, the images were divided into two categories: BK group and non-BK group. The residual dataset was used to validate the model after four of the five datasets were used to train a DL diagnostic model. For the performance validation of deep learning (DL) models, there were five rounds of experiments.

SE-ResNet50 had the best sensitivity, PPV, NPV, and accuracy of the four deep learning (DL) models, whereas ResNeXt50 had the best specificity. In the four non-EfficientNet deep learning (DL) models, none of the performance indicators for diagnosing BK reached statistical differences between any two models.

EfficientNet B0 exhibited the highest sensitivity of the four efficientNet deep learning (DL) models, whereas EfficientNet B3 had the highest specificity and PPV. EfficientNets B1 and B3, respectively, had the highest NPV and accuracy. In diagnosing BK, however, there was no significant difference in performance between any two of the four EfficientNet DL models.

When the four non-EfficientNet models were compared to the EfficientNet models, they all had much higher sensitivity than the EfficientNet models. Except for EfficientNet B0, which did not attain significance when compared to ResNeXt50 or SE-ResNet50, all EfficientNet models exhibited considerably greater specificities than non-EfficientNet models.

The AUROC and accuracy indices were used to summarise performance indices in the diagnosis of Bacterial Keratitis (BK). The researchers discovered that there was no significant difference in diagnostic accuracy and AUROC between all non-EfficientNet and EfficientNet models.

Conclusion

Before running a deep learning (DL) model for diagnosing Bacterial Keratitis (BK), the team compared the potential deep learning (DL) algorithms using an external eye photo in the same verification setting without adding a fluorescence staining photo, performing image segmentation, or transforming the images. Non-EfficientNet models (ResNet50, ResNeXt50, DenseNet121, SE-ResNet50) were shown to be more sensitive than EfficientNet models (EfficientNets B0, B1, B2, and B3) in this study, while EfficientNet models were found to be more specific. All of the models above have similar accuracy and AUROC.

Adopting a robust deep learning (DL) algorithm is a realistic way to improve the performance of an AI system for imaging diagnosis of ep learning (DL) model for diagnosing Bacterial Keratitis (BK). SE-ResNet, DenseNet, EfficientNets B1 and B3, and DenseNet were shown to have the highest AUROC, with DenseNet being the best DL method for diagnosis. By developing a more effective deep learning (DL) model in detecting Bacterial Keratitis (BK) based merely on an external eye image, the need for an additional fluorescent staining shot, advanced image segmentation or transformation, and a specifically constructed camera will be gradually reduced. In clinical settings lacking ophthalmological medical experts, this strategy may be more practical and effective.

Paper: https://www.nature.com/articles/s41598-021-03572-6.pdf