Flatiron Data Science Program
Module 4 Project - Neural Networks
December 15th, 2020
X-Ray Image Classification
Objective: Build a model that can classify whether a patient has pneumonia, given a chest x-ray image.
Repository Contents:
- images : images for display through this analysis
- project_4_notebook.ipynb : Google Colab notebook containing full analysis/modeling
- README.md : project summary and contents
- Mod04_presentation.pdf : presentation slides with comments
"Chest X-ray images (anterior-posterior) were provided from pediatric patients of Guangzhou Women and Children’s Medical Center, Guangzhou, China. The diagnoses for the images were then graded by two expert physicians before being cleared for training the AI system."
In the above images, some opacity can be observed in the bottom row of pneumonia lungs. This is due to the X-ray picking up the fluid filled air sacs due to the infection.
Sources: original study source, data download source/Kaggle competition
Full analysis: Colab Notebook
For this project I wanted to optimize the model to arrive at as few fasle predictions as possible. False positives could be costly in resources for the provider and patient, or result in unnecessary treatment. False negatives could allow illness to be missed and treatment delayed. Thus I selected accuracy and the AUC (area under the ROC curve) as target metrics for a balanced model, for both penalize false predictions.
The final model was trained on images with the following preprocessing:
- Images were resized to 124 x 124 pixels, with 3 RGB color channels
- Pixel values were normalized to a 0-1 scale
- To prepare the model to discern noise, four data augmentations were used: rotation, vertical/horizontal shifting, and zoom
- The imbalanced data set (75% pnemonia vs. 25% normal X-rays) was corrected by applying class weights
The final model arcitecture was a convolutional neural network (CNN) with 3 convolution blocks (convolution, drop out, pool).
Resulting in performances scores of:
- Accuracy - 89.77%
- Recall - 94 %
- Precision - 82 %
- AUC - 0.88
- Continue collecting labeled images to progressively train the model.
- Store image data at 128 x 128 to conserve storage memory (this is up to a 10% reduction in original image size).
- Use the model to improve efficiency of Xray review.
This is a supervised learning task and thus performance is based on the quality of the dataset used.
- Collect more labeled images or continue data augmentation to increase the quantity of images in the training set.
- Try transfer learning - use an established x-ray classifier and build model on top of that.
- Progressively resize the model input image size to find the smallest possible input size without sacrificing performance.
Please review the full analysis in the Colab Notebook or view my presentation slideshow.