A novel abnormality annotation database for COVID-19 affected frontal lung X-rays.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Characteristic COVID-19 pneumonia findings on chest X-rays necessitate rapid screening methods.
• Existing machine learning approaches lack annotated datasets for COVID-19 chest X-rays, hindering algorithm development and explainability.

Purpose of the Study:

• To create the COVID Abnormality Annotation for X-Rays (CAAXR) database with annotated abnormalities on over 1700 chest X-rays.
• To establish protocols for semantic segmentation and classification tasks for robust algorithm evaluation.
• To provide benchmark results using popular deep learning models for both classification and segmentation.

Main Methods:

• Annotation of abnormalities in over 1700 frontal chest X-rays from the BIMCV-COVID19+ database.
• Definition of protocols for semantic segmentation and classification.
• Implementation and evaluation of deep learning models including DenseNet, ResNet, MobileNet, VGG, UNet, SegNet, and Mask-RCNN.

Main Results:

• Performance metrics including classwise accuracy, sensitivity, and AUC-ROC for classification models.
• Intersection over Union (IoU) and DICE scores for semantic segmentation models.
• Benchmark results demonstrating the utility of the CAAXR database for evaluating AI models.

Conclusions:

• The CAAXR database provides essential ground-truth annotations for developing and validating AI algorithms for COVID-19 pneumonia detection from chest X-rays.
• Standardized protocols and benchmark results facilitate the advancement of explainable AI in medical imaging for infectious diseases.
• This annotated dataset is crucial for improving the accuracy and reliability of automated screening tools for COVID-19.