Data Property Mitigator: Empowering Pneumonia Detection Using Deep Learning Models and Chest X-Rays

We propose the Data Property Mitigator, that incorporates the results of previous research that have studied the impact of Gender and Chest X-Ray imbalance on Pneumonia and COVID-19 Pneumonia aiming to minimize the hidden biases in the adopted datasets. The model is ought to cover the patient’s sex and the Chest X-Ray view properties, and can be expanded to cover more data properties. It is also meant to be generalizable to different detection models, and different medical applications. To evaluate the Data Property Mitigator, segmentation using U-Net and detection using Faster RCNN and YOLOv5 were incorporated. Additionally, MobileNetV2 was used to evaluate it. From the performed experiments, we have spotted the weaknesses of the curated datasets for the deep-learning driven detection of Pneumonia and COVID-19 Pneumonia from Chest X-Rays, and established a benchmark to deploy the Data Property Mitigator in future research contributions.