Pneumothorax-Classifier

Takes in chest radiographs with labels of -1 for no pneumothorax, and values for a mask of the pneumothorax, and it uses different model architectures to determine if out of sample chest radiographs contain pneumothorax and where.

Machine Learning Techniques used:

Preprocessing Techniques:

• Edge detection filtering
• Intensity Threshold to reduce noise
• Cropping radiograph images to focus on ribcage
• Image augmentation through translation, shearing, rotation, etc.

Model Training/Tuning Techniques

• Train dataset split by 50% positive and 50% negative
• Early stopping and saving best model during training
• Architectures include CNN and U-net
• Classification type includes binary and segmentation
• Resampling ensemble
• K-fold cross validation
• Weighted averaging ensemble
• Bagging (Bootstrap Aggregation)

Future improvements

• Add Grid Search for model training
• Add Transfer Learning for model training
• Add Horizontal Ensemble for model training
• Add Snapshot Ensemble for model training
• Add Stacked Ensemble for model training
• Add Testing option for existing model training sessions

This project is being designed to be easy to add in a new 2D radiograph of another body part, and require minimal changes to the code.

Min Requirements (or latest versions):

• Python - 3.5.4
• Keras - 2.2.4
• Tensorflow - 1.8.0
• Pydicom - 1.3.0
• Imageio - 2.2.0
• Pillow - 4.2.1
• CV2 - 4.1.1
• Pandas - 0.22.0
• Scipy - 1.1.0
• Numpy - 1.13.3