Project Description

This project, completed in a single day, demonstrates how to tackle the challenge of Out of Distribution (OOD) data or Open Set Recognition (OSR) during testing phases. The goal is to construct a classifier that is adept at identifying five specific classes of X-Ray images:

1. Teeth
2. Skull
3. Chest
4. Hand
5. Foot
6. OOD (Out of Distribution) or Unknown

It's important to note that a small fraction of the data within each category might be mislabeled, necessitating careful handling to ensure data quality.

Prerequisites

Before you can run the script, ensure you have the following installed:

• Python 3.10
• TensorFlow 2.10
• Pandas
• Numpy
• scikit-learn

You can install the necessary Python packages using conda:

conda env create -f environment.yml

Work Plan

The workflow is organized into three main phases:

Data Preprocessing

The initial phase involves manual screening to eliminate irrelevant or poor-quality images as thoroughly as possible. I created two lists of irrelevant images namely as strict and non-strict cleaned images. Subsequently, the filtered images are converted to grayscale to facilitate the training process. The model is trained on strictly cleaned images but its performance is assessed on both groups.

Model Training

During this phase, we employ a VGG16 model for training on our dataset, utilizing a 75-25 split ratio for training and testing, respectively. Evaluation metrics including accuracy, precision, recall, f1-score, and a confusion matrix are detailed in the test_model.ipynb notebook.

Handling OOD Data

To address OOD data, we explore four straightforward methodologies from oodeel library: Maximum Logit Score (MLS), Maximum Softmax Probability (MSP), Entropy, and Energy. Each method's efficacy is assessed in dedicated notebooks.

Oodeel is a library that performs post-hoc deep OOD detection on already trained neural network image classifiers. The philosophy of the library is to favor quality over quantity and to foster easy adoption.

Important Note: The thresholds for these OOD methods require tuning, a step we did not undertake in this project. Future work could involve optimizing these thresholds for improved performance. You can also run prediction on your image without using any of those mentioned OOD methods by just not passing any input to the -ood parameter.

How to run scripts

To run the train script, navigate to the script's directory in your terminal and use the following command format:

conda activate ood
python train.py --data_dir <path_to_data_directory> --bad_files <bad_files_to_ignore> [--num_classes <number_of_classes>] [--batch_size <batch_size>] [--epochs <number_of_epochs>] [--test_size <validation_split_size>] [--labels_csv <dataset_csv_info_file>]

Arguments

• --data_dir (required): The directory containing the dataset and labels_x_ray.csv file.
• --bad_files (required): The file name of the txt file containing filtered files to ignore them. File must be in data_dir directory.
• --num_classes (optional): The number of classes in the dataset. Default is 5.
• --batch_size (optional): The batch size for training. Default is 32.
• --epochs (optional): The number of epochs for which the model will be trained. Default is 100.
• --test_size (optional): The ratio of the dataset to be used as validation set. Default is 0.25.
• --labels_csv (optional): The file name of the csv file containing labels and file names. File must be in data_dir directory.

To predict the label of an input image, use the following script:

conda activate ood
python predict.py -i <image_file_name> [-ood <ood_method>] [-dd <data_directory>] [-m <model_path>]

Arguments

• -i, --image: The name of the image file you want to predict. This argument is required.
• -ood, --ood-method: Optional. Specifies the Out-of-Distribution detection method to use. Leave blank if not using OOD detection.
• -dd, --data-dir: Optional. The directory where your images are stored. Defaults to data/x_ray.
• -m, --model: Optional. The path to the trained model file. Defaults to model/vgg16-0.96-full_model.h5.
• -s, --strict: Optional. Run prediction based on stricted or non-stricted threshold.
• -rl, --return_label: Optional. Whether to return string label or int index of the predicted class.