TLDR: This research paper investigates how deep learning models for Alzheimer’s disease diagnosis from MRI scans can learn “shortcuts” based on invisible demographic attributes like race and sex. The study demonstrates that these models can accurately classify race and sex from brain MRIs. Furthermore, it shows that imbalanced training data based on these demographics leads to biased diagnostic performance, with models relying on demographic features rather than purely disease-related ones. The authors introduce a method to identify and quantify these “invisible” shortcut features in brain regions, highlighting the need for fairer AI in medical imaging.
Deep learning algorithms are increasingly used in medical imaging, including for diagnosing diseases like Alzheimer’s disease (AD) from MRI scans. While these algorithms hold great promise, a new study highlights a critical concern: they can learn “shortcuts” that lead to biases, particularly when these shortcuts are linked to demographic factors like race and sex.
A research paper titled “Invisible Attributes, Visible Biases: Exploring Demographic Shortcuts in MRI-based Alzheimer’s Disease Classification” by Akshit Achara, Esther Puyol Anton, Alexander Hammers, and Andrew P. King, for the Alzheimer’s Disease Neuroimaging Initiative, delves into this issue. The authors investigate how deep learning models might inadvertently use demographic information, even when it’s not directly visible to humans, to make diagnostic predictions. This can result in unfair performance against underrepresented groups.
Detecting Hidden Demographics
The first key finding of the study is that deep learning algorithms can indeed identify a person’s race or sex from 3D brain MRI scans with high accuracy. This is significant because if a model can distinguish between different demographic groups, there’s a potential for features associated with these attributes to become “shortcuts” in diagnostic tasks. The researchers used two different deep learning models, ResNet and SwinTransformer, and three datasets (ADNI, OASIS-3, and HCP) to demonstrate this capability. For instance, F1-scores for sex classification were consistently high, often above 0.87, and race classification also showed high accuracy, even with data imbalances.
Shortcut Learning and Performance Bias
The core of the research explores whether imbalances in training data, based on race or sex, can cause a drop in model performance for AD diagnosis, indicating shortcut learning and bias. The researchers created “baseline” datasets with balanced demographic representation and “biased” datasets where certain demographic groups were over- or under-represented in relation to the diagnostic labels. They found that when models were trained on these biased datasets, their performance in AD classification dropped significantly, especially for the minority groups in the biased training sets. For example, a steeper drop in F1-scores was observed for the AD class in sex-biased datasets. This suggests that the models learned to rely on demographic features rather than purely disease-related indicators.
Also Read:
- Improving Medical Image Diagnosis with Expert-Guided AI
- Automated Deep Learning Tool Accurately Measures Brain Atrophy from CT Scans
Unveiling the “Invisible” Shortcuts
To understand *what* features the models were using, the team conducted a quantitative and qualitative analysis of feature attributions in different brain regions. They used a technique called GradCAM to visualize which parts of the MRI scans were most influential in the models’ decisions. Unlike some computer vision tasks where spurious correlations (like background in an image) are obvious, identifying these “invisible” shortcuts in medical images is challenging for humans. The researchers developed a novel rank-based analysis to quantify the extent of shortcut learning and pinpoint the brain regions associated with it. They found significant correlations between regions contributing to bias and those associated with protected attributes, particularly for sex-biased experiments.
This work lays a crucial foundation for developing fairer deep learning diagnostic tools in brain MRI. By demonstrating the existence of both race and sex-based shortcut learning and bias, the study emphasizes the need for careful dataset curation and model development to ensure equitable healthcare outcomes. The code for their implementation is publicly available, fostering further research in this area. You can read the full paper here: Invisible Attributes, Visible Biases: Exploring Demographic Shortcuts in MRI-based Alzheimer’s Disease Classification.
