# Comparative study of Deep Learning Models for Binary Classification on Combined Pulmonary Chest X-ray Dataset

**Authors:** Shabbir Ahmed Shuvo, Md Aminul Islam, Md. Mozammel Hoque, Rejwan Bin Sulaiman  
**arXiv:** [2309.10829](https://arxiv.org/abs/2309.10829)  
**Date:** 2023-09-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.10829v2/

## Key points

- Comparative analysis of deep learning models for binary classification on a combined Pulmonary Chest X-ray dataset
- Aimed to determine relative performance of different models on the same dataset
- Dataset combined Shenzhen, China and Montgomery, USA data
- Trained model using eight prominent deep learning models: DenseNet 121, DenseNet 169, DenseNet 201, EfficientNet b0, EfficientNet lite4, GoogleNet, MobileNet and ResNet18
- Maintained controlled comparison environment with same training parameters for all models
- Found distinct differences in performance among the models on pulmonary chest X-ray image dataset
- DenseNet169 achieved precision of 89.38 percent and MobileNet achieved precision of 92.2 percent
- Conducted analysis of fusion datasets by combining two datasets from different geographical regions for improved generalization and robustness
- Comprehensive comparative performance analysis of the eight deep transfer learning models
- Provides valuable insights into effectiveness of these models in tuberculosis diagnosis for clinicians and researchers
- Clear ranking of the models based on their performance
- Examined reasons behind each model's performance and offered guidance for model selection and potential areas of improvement
- Contributes to advancing tuberculosis diagnosis through deep learning models by providing insights into relative performance and recommendations for practical implementation.

## Summary

In this study, the authors conducted a comparative analysis of deep learning models for binary classification on a combined Pulmonary Chest X-ray dataset. The aim was to address the knowledge gap in determining the relative performance of different models when applied to the same dataset. The dataset used in this study combined Shenzhen, China (CH) and Montgomery, USA (MC) data. The researchers trained their model for binary classification using eight prominent deep learning models: DenseNet 121, DenseNet 169, DenseNet 201, EfficientNet b0, EfficientNet lite4, GoogleNet, MobileNet and ResNet18. They ensured that all models followed the same training parameters to maintain a controlled comparison environment. At the end of the study they found distinct differences in performance among the models when applied to the pulmonary chest X-ray image dataset. Specifically DenseNet169 achieved a precision of 89.38 percent and MobileNet achieved a precision of 92.2 percent. To provide more context and expand on their findings, the researchers also conducted an analysis of fusion datasets by combining two datasets from different geographical regions resulting in a larger and more representative dataset for improving generalization and robustness of their analysis. Furthermore they performed a comprehensive comparative performance analysis of the eight deep transfer learning models which provided clinicians and researchers with valuable insights into the effectiveness of these models in tuberculosis diagnosis. The study also presented a clear ranking of the models based on their performance as well as examining reasons behind each model's performance offering guidance for model selection and potential areas of improvement. Overall this research contributes to advancing tuberculosis diagnosis through deep learning models by providing insights into their relative performance and offering recommendations for practical implementation.