harishmogili21/pneumonia-detection-inceptionresnetv2

Pneumonia Detection using InceptionResNetV2

Project Description

This project implements a transfer learning approach for detecting pneumonia from chest X-ray images. It utilizes the InceptionResNetV2 model pre-trained on the ImageNet dataset, with the convolutional layers frozen and new layers added for binary classification.

Dataset

The model was trained on the Chest X-Ray Images (Pneumonia) dataset from Kaggle.

• Training set:
  • Normal: 1341 images
  • Pneumonia: 3875 images
• Validation set:
  • Normal: 8 images
  • Pneumonia: 8 images
• Test set:
  • Normal: 234 images
  • Pneumonia: 390 images

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

Model Architecture

The model is based on the InceptionResNetV2 architecture. The convolutional base layers are frozen, and the following layers were added on top:

• Global Average Pooling 2D layer
• Two Dense layers with ReLU activation (512 and 128 units)
• Output Dense layer with Softmax activation for 2 classes (Normal, Pneumonia)

Training

• Optimizer: Adam (learning rate 0.001)
• Loss Function: Categorical Crossentropy
• Epochs: 20

Performance

The model's performance was evaluated on the test set:

• Accuracy on Normal images: 156 out of 234 correct
• Accuracy on Pneumonia images: 375 out of 390 correct

Usage

This model can be loaded and used to predict whether a chest X-ray image shows signs of pneumonia. The input images should be preprocessed (resized to 200x200 and scaled by 1/255.0) before being passed to the model for prediction.

Limitations and Future Work

• The validation set is small, which may lead to erratic validation accuracy plots.
• Image resizing from high-quality originals to 200x200 might impact performance.
• Further hyperparameter tuning and potentially training more layers of the base model are needed to improve overall performance, especially in reducing false positives for normal cases.