CCI.net fungal keratitis cases dataset
Data files
Jan 23, 2024 version files 258.77 MB
Abstract
Patients diagnosed with FK were enrolled from three cities in China between April 2015 and June 2022. The diagnosis was confirmed either by corneal cultures or corneal biopsy. Patients who were over 65 years old, had systemic diseases such as diabetes, autoimmune diseases, thyroid-related diseases, or were pregnant, were excluded. Patients with bacterial, amebic, or other infections confirmed by culture were also excluded. A total of 355 patients were included, comprising 212 from Zhongshan Ophthalmic Center in Guangdong Province, 83 from Xi'an First Hospital in Shanxi Province, and 60 from Guangxi Zhuang Autonomous Region People's Hospital in Guangxi Zhuang Autonomous Region.
There were 5 licensed ophthalmologists using standard processing to grade the image quality. A total of 8697 images were graded as good quality and were further divided into three datasets: a classification dataset (N=7086, including 4255 in the training set, 1413 in the validation set, and 1418 in the test set), a segmentation dataset (N=56, including 46 in the training set and 10 in the test set), and an external validation set (N=1555, including 1455 to validate the classification model and 100 to validate the segmentation model).
README
There were 5 licensed ophthalmologists using standard processing to grade the image quality. A total of 8697 images were graded as good quality and were further divided into three datasets: a classification dataset (N=7086, including 4255 in the training set, 1413 in the validation set, and 1418 in the test set), a segmentation dataset (N=56, including 46 in the training set and 10 in the test set), and an external validation set (N=1555, including 1455 to validate the classification model and 100 to validate the segmentation model) (Figure 1).
Clinically, three experienced ophthalmologists from Zhongshan Ophthalmic Center were responsible for ophthalmological research and clinical work. According to the results of fungal culture, each image was divided into five categories. Two ophthalmologists separately classified images of each type of fungal keratitis, and the third ophthalmologist reclassified the images with inconsistent classification results from the first two ophthalmologists. The labeling results were 147 for Aspergillus flavus, 3150 for Fusarium, 182 for Mucor, 197 for Penicillium, and 3410 for images without fungal hypha according to the corneal cultures or corneal biopsy results.
The image dataset was divided into three categories with a ratio of 6:2:2 for the training set, validation set, and test set. Before the experiment, all images were resized to 224224 pixels. The inverse frequency algorithm was used for training images to assign different weights to each type of image, and random horizontal flip, horizontal and vertical offset were applied to enhance the robustness of the network.
Establishment and annotation of segmentation datasets. For labeling fungal hyphae, drawing software was used, where the brush color was set to red and the brush width was set to 1 pixel. Two ophthalmologists independently labeled fungal hyphae in 56 images to visualize them. Out of these images, 46 were used for the training set and 10 were used for the testing set.
Establishment and annotation of the external validation set. The dataset comprises a total of 1555 images, out of which 1455 are annotated and used to validate the classification model. The labeling results show that there are 683 normal pictures and 772 pictures with fungal hyphae. Additionally, 100 images are annotated and used to validate the segmentation model.
Methods
Patients diagnosed with FK were enrolled from three cities in China between April 2015 and June 2022. The diagnosis was confirmed either by corneal cultures or corneal biopsy. Patients who were over 65 years old, had systemic diseases such as diabetes, autoimmune diseases, thyroid-related diseases, or were pregnant, were excluded. Patients with bacterial, amebic, or other infections confirmed by culture were also excluded. A total of 355 patients were included, comprising 212 from Zhongshan Ophthalmic Center in Guangdong Province, 83 from Xi'an First Hospital in Shanxi Province, and 60 from Guangxi Zhuang Autonomous Region People's Hospital in Guangxi Zhuang Autonomous Region.
There were 5 licensed ophthalmologists using standard processing to grade the image quality. A total of 8697 images were graded as good quality and were further divided into three datasets: a classification dataset (N=7086, including 4255 in the training set, 1413 in the validation set, and 1418 in the test set), a segmentation dataset (N=56, including 46 in the training set and 10 in the test set), and an external validation set (N=1555, including 1455 to validate the classification model and 100 to validate the segmentation model).