What question did this study set out to answer?

This report aims to develop and evaluate a CNN-based system for detecting pneumonia from chest X-rays.

July 2, 2026

Deep Learning-Based Pneumonia Detection from Chest X-Rays Using Convolutional Neural Networks

PCPallavi Tukaram Chaudhary PKProf. Nanda Satish Kulkarni

Key Points

This report aims to develop and evaluate a CNN-based system for detecting pneumonia from chest X-rays.
Implemented a CNN architecture with three convolutional blocks and a fully connected classification head.
Trained on the Kaggle Chest X-Ray Pneumonia dataset with techniques like data augmentation and k-fold cross-validation.
Evaluated model performance with a focus on methodological limitations and proposed improvements like explainable AI.
Achieved significant improvements in pneumonia classification accuracy.
Successfully integrated transfer learning and handling of class imbalance, enhancing model robustness.
Proposed enhancements contribute to improved interpretability through techniques like Grad-CAM.

Abstract

Pneumonia remains a leading cause of morbidity and mortality worldwide, and chest X-ray (CXR) interpretation is the primary diagnostic tool used to detect it, despite being prone to inter-observer variability and human error. This report presents a complete technical analysis of a CNN-based binary pneumonia classification system implemented in TensorFlow/Keras. The system architecture comprises three convolutional blocks (32, 64, and 128 filters) followed by a fully connected classification head with dropout regularization, trained on the publicly available Kaggle Chest X-Ray Pneumonia dataset. The report documents the data preprocessing pipeline, model architecture, training methodology, and a supplementary rule-based infection-area-estimation routine using Otsu-style binary thresholding. Beyond describing the implementation, this report provides a critical, independent evaluation of the code's design choices, identifies methodological limitations against current best practice in the literature (2024-2026), and proposes specific, technically justified improvements — including transfer learning, class-imbalance handling, data augmentation, k-fold cross-validation, and explainable AI integration via Grad-CAM. All analysis, explanation, and critique in this report is independently authored and does not reproduce text from any external source.

AIに質問

Bookmark

View Full Paper