The Importance of Model Inspection for Better Understanding Performance Characteristics of Graph Neural Networks

This study highlights the importance of conducting comprehensive model inspection as part of comparative performance analyses. Here, we investigate the effect of modelling choices on the feature learning characteristics of graph neural networks applied to a brain shape classification task. Specifically, we analyse the effect of using parameter-efficient, shared graph convolutional submodels compared to structure-specific, non-shared submodels. Further, we assess the effect of mesh registration as part of the data harmonisation pipeline. We find substantial differences in the feature embeddings at different layers of the models. Our results highlight that test accuracy alone is insufficient to identify important model characteristics such as encoded biases related to data source or potentially non-discriminative features learned in submodels. Our model inspection framework offers a valuable tool for practitioners to better understand performance characteristics of deep learning models in medical imaging.

Analysing race and sex bias in brain age prediction

Brain age prediction from MRI has become a popular imaging biomarker associated with a wide range of neuropathologies. The datasets used for training, however, are often skewed and imbalanced regarding demographics, potentially making brain age prediction models susceptible to bias. We analyse the commonly used ResNet-34 model by conducting a comprehensive subgroup performance analysis and feature inspection. The model is trained on 1,215 T1-weighted MRI scans from Cam-CAN and IXI, and tested on UK Biobank (n=42,786), split into six racial and biological sex subgroups. With the objective of comparing the performance between subgroups, measured by the absolute prediction error, we use a Kruskal-Wallis test followed by two post-hoc Conover-Iman tests to inspect bias across race and biological sex. To examine biases in the generated features, we use PCA for dimensionality reduction and employ two-sample Kolmogorov-Smirnov tests to identify distribution shifts among subgroups. Our results reveal statistically significant differences in predictive performance between Black and White, Black and Asian, and male and female subjects. Seven out of twelve pairwise comparisons show statistically significant differences in the feature distributions. Our findings call for further analysis of brain age prediction models.