Detection of Viral and Bacterial Pneumonia Through Pattern Recognition and Machine Learning on Chest X-rays (Detecção de pneumonia viral e bacteriana por meio de reconhecimento de padrões e aprendizado de máquina em radiografias de tórax)

Resumo

https://doi.org/10.5281/zenodo.20600900

Abstract

During 2024, there were more than 245 thousand cases of pneumonia in Brazil which led to admissions to hospitals, an increase of 30% compared to the previous year. Therefore, our objective was the development of techniques that provide a quick and reliable identification of this disease. In that regard, we developed an image recognition application, capable of distinguishing unhealthy individuals from a pool of X-ray images, and identifying which type of pneumonia (viral or bacterial) present. This application would be accessible to healthcare professionals such as nurses and physicians via a user-friendly front end application such as a website. In order to identify the images, we trained a machine learning model using a dataset[1] containing labelled X-ray images of both healthy and unhealthy lungs, and later on we switched to a deep learning model. When an external test was run, the results were not impressive so we decided to return to the previous model and improve it using the KFold technique. The result obtained was a model with both accuracy and precision close to 75%, capable of reliably detecting unhealthy lungs and which type of pneumonia was present if any. This model did not use CNN and was composed of XGBoost with KFold and image treatment. After that the model was made available through a proof of concept website, which let the user upload an X-ray image of their chest and receive the model’s prediction, so even lay people could use the application.