Chest Xray Pneumonia Classification

Author: Benjamin Toler

Overview

This project used nearly 6,000 pediatric xray images from the Guangzhou Women and Children’s Medical Center to develop an image classification model that can accurately identify the presence of pneumonia. The model has been developed for a local hospital that is looking to increase their efficiency and accuracy in diagnosing pneumonia in order to improve patient experience and outcome. The data was processed using ImageDataGenerator from the keras library, and then split into training, validation, and test datasets. The training set was further split into a smaller training set for faster processing on initial modeling iterations. The project uses convolutional neural networks from the keras library. Over several iterations, the complexity of the model architecture is adjusted through the number of layers, the number of convolutional filters, and the number of neurons. Additionally, several regularization techniques are applied to reduce overfitting. The final CNN model showed a 96% accuracy in detecting pneumonia.

Business Problem

A local hospital is looking to improve the efficiency and accuracy in which they diagnose pediatric pneumonia. Using a classification model will decrease patient wait times and reduce errors, making overall patient experience better.

Data

The data for this project comes from the Guangzhou Women and Children’s Medical Center. It consists of 5,956 chest x-ray images (256x256) of patients ages 1-5. The dataset has a class imbalance of 73% pneumonia and 27% normal.

Method

A convolutional neural network(CNN) model was constructed using an iterative approach. The model architecture and hyperparameters were tuned and regularization techniques were deployed until an acceptable training and validation accuracy score was reached.

Results

The final model consisted of 2 convolutional layers each followed by max pooling layers, a flattening layer, and one hidden and one output dense layer. The intial model started with a 67% accuracy, and predicted only pneumonia as seen in the first confusion matrix below. The final model finished with a 96% accuracy and only misclassified 14 of the 646 pneumonia images in the test dataset as seen in the second confusion matrix below.

|

Conclusions

• Use model in evaluating potential pneumonia patients: With a 96% accuracy, the model should be used in conjunction with the standard manual diagnosis process to help reduce errors in pneumonia detection.
• Run model for all chest xrays: The model should be run for all chest xrays, not just when pneumonia is expected. This will prevent missing it in cases where the doctors may not be closely looking for it.
• Use model to train hospital doctors: The model can be used to train doctors where to look for pneumonia and what to look for.

Next Steps

• Test model on adult chest xrays: the current model was built to identify pneumonia in pediatric patients ages 1-5. Future models should be generalized to work for all age groups.
• Further hyperparameter and architecture tuning to see if accuracy can be further improved: due to the processing time needed for these models, only so much variable and architecture tuning was possible. Further tuning may improve the model.
• Create multi-class model that can identify other conditions/diseases in the chest xray: Having a classification model that can identify multiple conditions that can be seen on a chest xray will further improve efficiency in the hospital.

For More Information

See the full analysis in the Jupyter Notebook or review this presentation.

For additional info, contact Benjamin Toler at bentoler22@gmail.com

Repository Structure

├── cnn_model_7
├── final_model
├── images
├── .gitignore
└── Chest_Xray_Pneumonia_Classification.ipynb
├── README.md
├── presentation.pdf