# Weakly Supervised Pneumonia Localization from Chest X-Rays Using Deep Neural Network and Grad-CAM Explanations

**Authors:** Kiran Shahi, Anup Bagale  
**arXiv:** [2511.00456](https://arxiv.org/abs/2511.00456)  
**Date:** 2025-11-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2511.00456v1/

## Key points

- Weakly supervised deep learning framework for pneumonia classification and localization from chest X-rays
- Utilizes Grad-CAM explanations to generate clinically meaningful heatmaps highlighting regions affected by pneumonia
- Evaluation of seven ImageNet-pretrained architectures under identical training conditions with focal loss and patient-wise splits
- ResNet-18 and EfficientNet-B0 achieved highest overall test accuracy of 98%, ROC-AUC of 0.997, and F1 score of 0.987
- MobileNet-V2 offered a good balance between accuracy and computational cost
- Models focused on clinically relevant lung regions according to Grad-CAM visualizations
- Emphasizes potential of weakly supervised explainable models in enhancing transparency in pneumonia screening and building clinical trust in AI-assisted medical imaging
- Further enhancements include radiologist reader studies and mobile deployment optimizations to strengthen clinical relevance and translational impact

## Summary

, , , , 

This study presents a weakly supervised deep learning framework for pneumonia classification and localization from chest X-rays, utilizing Grad-CAM explanations. By leveraging image-level labels instead of costly pixel-level annotations, the approach generates clinically meaningful heatmaps highlighting regions affected by pneumonia. Seven ImageNet-pretrained architectures were evaluated under identical training conditions with focal loss and patient-wise splits to prevent data leakage. Results on the Kermany CXR dataset show that ResNet-18 and EfficientNet-B0 achieved the highest overall test accuracy of 98%, ROC-AUC of 0.997, and F1 score of 0.987, while MobileNet-V2 offered a good balance between accuracy and computational cost. The Grad-CAM visualizations confirmed that the models focused on clinically relevant lung regions, supporting their interpretability for radiological diagnostics. The study emphasizes the potential of weakly supervised explainable models in enhancing transparency in pneumonia screening and building clinical trust in AI-assisted medical imaging. Further enhancements to the framework include radiologist reader studies and mobile deployment optimizations to strengthen its clinical relevance and translational impact. Overall, this research underscores how explainable and weakly supervised deep learning methods can bridge the gap between black-box image classification and clinically interpretable decision support for pneumonia detection. The author contributions include conceptualization, methodology, software development by K.S., data curation and original draft preparation by both K.S. and A.B., visualization and investigation by both authors, supervision by K.S., validation by both authors, as well as reviewing and editing of the manuscript. The research received no external funding. The dataset used is publicly available (Kermany dataset), with the source code supporting the findings accessible at https://github.com/kiranshahi/pneumonia-analysis. The study acknowledges the use of the Chest X-ray dataset by Kerman. In conclusion, this work demonstrates how explainable AI models can enhance pneumonia detection through transparent heatmaps while maintaining clinical interpretability for improved medical imaging outcomes.