Detecting Unforeseen Data Properties with Diffusion Autoencoder Embeddings using Spine MRI data

TL;DR

Evaluations on a large-scale dataset consisting of sagittal T2-weighted MR images of three spine regions show that DAE embeddings effectively separate protected variables such as sex and age, and highlight the potential of using advanced embedding techniques like DAEs to detect data quality issues and biases in medical imaging datasets.

Abstract

Deep learning has made significant strides in medical imaging, leveraging the use of large datasets to improve diagnostics and prognostics. However, large datasets often come with inherent errors through subject selection and acquisition. In this paper, we investigate the use of Diffusion Autoencoder (DAE) embeddings for uncovering and understanding data characteristics and biases, including biases for protected variables like sex and data abnormalities indicative of unwanted protocol variations. We use sagittal T2-weighted magnetic resonance (MR) images of the neck, chest, and lumbar region from 11186 German National Cohort (NAKO) participants. We compare DAE embeddings with existing generative models like StyleGAN and Variational Autoencoder. Evaluations on a large-scale dataset consisting of sagittal T2-weighted MR images of three spine regions show that DAE embeddings effectively separate protected variables such as sex and age. Furthermore, we used t-SNE visualization to identify unwanted variations in imaging protocols, revealing differences in head positioning. Our embedding can identify samples where a sex predictor will have issues learning the correct sex. Our findings highlight the potential of using advanced embedding techniques like DAEs to detect data quality issues and biases in medical imaging datasets. Identifying such hidden relations can enhance the reliability and fairness of deep learning models in healthcare applications, ultimately improving patient care and outcomes.

Figures (4)

• Figure 1: t-SNE plots of our DAE embeddings. The embeddings are colored with the patient sex, height, weight, and age. The Region label describes the body region; the acquisition location is the city where the image was recorded.
• Figure 2: Location Biases. Three clusters separate from the head images when we color them by the examination center where the MRI was taken. For each cluster and the rest class, we show two images. The left is a selection of the right edges of the images, and on the right are the images summed together. Both are made from 1000 images and are averaged. We observe that the neck curve differs in the clusters. On the top left, we visualized this by plotting the mean line of all 4 clusters.
• Figure 3: Sex Biases. Left: A subset of subjects are clustered between the male and female blobs, and others are completely pushed to the opposite sex. They must have a set of features that indicate that the spine is the opposite sex. We have no further evidence of how this is reflected in the sex and gender of those persons overall. Right: The training curves of a male/female classifier of ResNet10 and ResNet34. We train only on the male/female label but measure the misclassified group separately and observe that they clearly lag behind during training.
• Figure 4: t-SNE plots of StyleGAN embeddings. The embeddings are colored with the patient sex, height, weight, and age. The Region label describes the body region; the acquisition location is the city where the image was recorded.