Predicting Pnuemonia Disease

Creating a deep-learning project to predict pneumonia disease in a patient through patterns observed in lung X-rays using transfer learning

IEEE Paper Link - https://ieeexplore.ieee.org/document/9297388

Motivation for selecting the problem statement

Pneumonia can range in seriousness from mild to life-threatening as it causes inflammation and fluid accumulation in the lungs, making it hard to breathe. It is most serious for infants and young children, people older than age 65. It is also a potential complication of COVID-19

To help detect which patients require more medical care, we can use artificial intelligence to analyze lung imaging as part of a clinical research study. The goal would be develop a model that can predict the severity of a pneumonia patient’s lung disease through images of their lung X-ray.

Dataset Details

Dataset link - https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia

The dataset is organized into 3 folders (train, test, val) and contains subfolders for each image category (Pneumonia/Normal). There are 5,863 X-Ray images (JPEG) and 2 categories (Pneumonia/Normal).

Chest X-ray images were selected from retrospective cohorts of patients from five to thirty years old from Guangzhou Women and Children’s Medical Center, Guangzhou. All chest X-ray imaging was performed as part of patients’ routine clinical care.

DL Algorithms/Model details

VGG-16

• VGG-16 Neural Network consists of 16 layers of deep neural network trained on more than a million from the “ImageNet” database.
• After training for 5 epochs, we received an accuracy of 91%

VGG-19

• The concept of the VGG19 is same as the VGG16 except that the model is more complex as the network is 19 layers deep.
• After training for 5 epochs, we received an accuracy of 92%

RESNET-50

• Resnets are a kind of CNNs called Residual Networks. They are very deep compared to VGG-16 and VGG-19.
• After training for 5 epochs, we received an accuracy of 85%

Training and Testing details

Training

Training the model for more than 5216 chest x-ray images. Trained each model for 5 epochs to achieve best accuracy

Testing

Comparing with the training model to achieve an accuracy and loss graph. Testing with 624 images. Altering the parameters, number of epochs to achieve highest accuracy and minimal loss

Results

We receive output in 2 classes (one denoting NORMAL case, other denoting PNEUMONIA case) Upon judging from this output, we can deduce if the person is suffering from pneumonia or not.

Accuracy & Loss graphs

Results obtained by calculations

Comma separated values indicates values for NORMAL and PNUEMONIA class