This project, conducted as part of the INF8225 course at Polytechnique Montréal, compares several anomaly detection methods in the context of medical imaging. Inspired by the paper Anomaly Detection in Medical Imaging with Deep Perceptual Autoencoders, the project focuses on Deep Isolation Forest (Deep IF), Perceptual Image Anomaly Detection (PIAD), and Deep Perceptual Autoencoder (DPA) methods applied to chest X-ray datasets. The evaluation includes metrics such as ROC-AUC, accuracy, and F1 score, as well as the analysis of confusion matrices. The project compares its results with those reported in the original paper, crediting the code repository where implementations were utilized.
The project is organized as follows:
- notebooks: Contains all the notebooks used to launch the various training sessions and analyzes them from Google Colab.
- anomaly_detection: Package containing the implementation of the various models used (see Credits).
- data: Useful data, including a subset of ChestX-ray14 of over 7,000 images resized to 300x300.
- Comparison_of_Anomaly_Detection_Methods_in_Chest_X-rays.pdf : Our “article” contains an in-depth analysis of our results and comparisons between the different models.
This project is designed to be executed from notebooks, originally intended for Google Colab. To use:
- Clone the repository to your local machine.
- Open and run the provided notebooks in your preferred environment.
- Follow instructions provided in the notebooks to execute the code and reproduce experiments.
The code implementations utilized in this project were sourced from this repository.
See the original paper for more information: Anomaly Detection in Medical Imaging With Deep Perceptual Autoencoders
by Shvetsova, Nina and Bakker, Bart and Fedulova, Irina and Schulz, Heinrich and Dylov, Dmitry V.