Pneumonia detection from chest X-rays using stratified cross-validation and DenseNet121 Fine-Tuning

Abstract

Pneumonia is a very common disease, especially in high Andean areas, its diagnosis is often slow due to the lack of technology or personnel, therefore it is essential to develop an Artificial Intelligence model that can diagnose pneumonia using X-ray plates, in this work, an optimized model was trained for the diagnosis of pneumonia from chest X-rays, using convolutional neural networks (CNN) together with stratified cross-validation. The database used was Chest X-ray Images (Pneumonia), publicly accessible on Kaggle, which was preprocessed and balanced to reduce class biases. The DenseNet-121 architecture was selected and adjusted using fine-tuning, dropout regularization, and batch normalization. It was subsequently evaluated using stratified cross-validation with five partitions. The model showed excellent performance, with an average AUC-ROC of 0,9967 and high F1-scores, especially for the pneumonia class, demonstrating strong ability to discriminate between normal and pathological cases. These results outperform other results from similar approaches without stratified validation or fine-tuning. It is concluded that the combination of pre-trained architectures, modern regularization techniques, and rigorous statistical evaluation significantly improves the automatic diagnosis of pneumonia, particularly in regions with limited access to specialists or advanced technology.