PneumonIA

Abstract

PneumonIA is a lightweight and open-source tool developed for the automated detection of
pneumonia in chest X-ray images. The system employs a convolutional neural network (CNN) trained on the
Chest X-Ray Images (Pneumonia) dataset from Kaggle. Its architecture comprises three convolutional blocks
followed by dense layers, implemented using Python, TensorFlow, and Keras. The model achieved 86%
accuracy, 99% sensitivity, and an AUC of 0.96, indicating high discriminatory power.The platform uses
grayscale conversion, median filtering, and adaptive thresholding to improve segmentation and highlight
infected regions. Unlike commercial solutions such as Qure.ai or Lunit, PneumonIA is fully operational
without requiring high-speed internet or PACS infrastructure, enabling deployment in first-level care units
and rural clinics. The graphical interface, developed using Gradio, enhances accessibility and educational
value for biomedical and medical students.Validation included 624 test images, demonstrating 82% precision
for pneumonia cases and 97% for typical cases, with a recall of 99% and 65%, respectively. While favoring
sensitivity, the system tends to overdiagnose, which may be acceptable in screening contexts. ROC analysis
yielded an AUC of 0.96, and F1-scores confirmed balanced diagnostic performance.The application is
currently functional on computers via Google Colab, with plans to enable mobile image export and DICOM
integration in the future. All datasets used are publicly available and anonymized, and the tool does not store
user data, ensuring compliance with privacy regulations. The tool has yet to be piloted in clinical workflows;
however, it has been technically validated through reproducible testing.PneumonIA's design addresses equity
in medical AI by focusing on accessibility and minimal infrastructure requirements. It offers excellent
potential as a decision-support tool and educational resource in image processing and AI for health. Future
iterations will incorporate real-world testing with healthcare professionals, utilize broader datasets, and
enhance specificity.