The architectures used are ResNet9, ResNet18 and other custom CNN models. The final model selected for use is ResNet18. Hyperparameters have been selected using optuna. Evaluation has been done using k fold cross validation for the final ResNet model. Ensembling used for 5 ResNet models trained on 5 different fold of the training Dataset
- The dataset has been taken from Kaggle Chest X-Ray Images (Pneumonia) dataset which consists of Chest X-ray images (anterior-posterior) selected from retrospective cohorts of pediatric patients of one to five years old from Guangzhou Women and Children’s Medical Center, Guangzhou.
- 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)
- There are 3 notebooks, each containing different models
- Basic Data preprocessing involving:
- Understanding data
- Transforming data
- Applying Random Rotations
- Normalizing RGB values
- Resizing
- Models used:
- ResNet9
- ResNet18
- Other custom CNN models
- Weight decay
- Automatized Hyperparameter search using optuna
- K Fold Cross Validation
- Ensembling, by dividing the training data into 5 folds and training 5 ResNet18 models seperately with the seperate folds of data and using the average of their outputs as the final prediction.
