home

Anomaly Detection in X-Ray Images

The aim of this project is to classify x-ray images of hands into normal or not normal hands. Because of the high cost of labelling the data the task should be done in an unsupervised way. That means that the labels should not be included while training.

The data is a subset of the MURA dataset (paper) and includes x-ray images of hands. The data seems not to be very clean (see First look at dataset). To handle this problem a data cleaning pipeline was implemented (see Data cleaning & preprocessing and Results of hand center localisation).

The summary of the research about different unsupervised methods to find anomalies can be found here. Different metrics are presented to evaluate different models.

General

• see Evaluation strategy and performance metrics
• see How to ...

Data Cleaning

• see Data cleaning & preprocessing
• Hand detection
  • for how to train a hand detection model see How to train an object detection model
  • results: Results of hand center localisation

Research Different Methods

• see Evaluation strategy & metrics
• see Unsupervised Deep Learning for Medical Image Analysis
• DBScan see Anomaly Detection Study

Models

Evaluation of Models

• Failure Cases
• Augmentation Effects