Anomaly Detection in Magnetic Resonance Image Series of the Lungs

Anomaly detection using Artificial Neural Networks (ANN) has shown promising results in disease detection and localisation in Magnetic Resonance (MR) images. New techniques in functional MR image acquisition allow us to image the lungs with high tissue contrast and can be used to continually monitor patients with chronic lung diseases such as cystic fibrosis (CF). Early detection of disease progression and treatment is crucial for improvement of the life quality of the a↵ected patients.
Anomaly segmentation with ANNs in a supervised manner requires manually labeled training data which is difficult to obtain. Additionally, manually labeled data comes
with a bias to the known defects. Autoencoder (AE) networks can be used for anomaly detection in an unsupervised manner, where images from only healthy subjects are used for the training.
In this work, we train various AE networks for anomaly detection in time-resolved 2D MR image series of the lungs. We investigate the potential of 3D, 2D and 1D
AE networks to reconstruct the images of patients with CF, expecting to find high reconstruction errors in defective regions of the lungs. We perform anomaly segmentation by thresholding the resulting error maps and compute the correlation between the results of the segmentation and the subjects’ Lung Clearance Index (LCI). LCI is a global measure of lung ventilation inhomogeneities. We visually compare the error maps and the anomaly segmentation results with previously computed ventilation maps showing the ventilation defects in the lungs.
The results of our study show that the 1D and 2D convolutional AE networks are able to detect anomalies corresponding to the ventilation defects in the lungs and have
moderate correlation when compared to the LCI values.