Establishing a relationship between iron-based blood measures and structural brain changes using neural networks in UK Biobank

• We demonstrate the prediction of haemoglobin concentration levels from brain susceptibility weighted MRI using a convolution neural network model. • Changes in the right thalamus and the white matter tracts between the temporal gyrus, putamen, hippocampus, and thalamus regions were identified as important in prediction for high haemoglobin levels. • The inferred association between two representative measures of iron modality, blood and brain, may have applications in understanding mechanistic changes in neurodegenerative disease. Disturbance of iron homeostasis in both the brain and blood is linked to cognitive impairment and neurodegenerative diseases. Investigation of the associations between changes in blood measurements and the accompanying structural changes in the brain, represented by susceptibility-weighted imaging, would generate evidence of a shared physiopathology between modalities, however linear approaches using derived imaging measures have not detected this relationship. Brain SWI of 4436 participants from UK Biobank at the first imaging visit were used to build a convolutional neural network (CNN) predicting the haemoglobin level. Haemoglobin concentration was grouped in three sets percentiles creating prediction classes. Attention maps were extracted from the CNN using Integrated Gradients to identify which brain regions allowed the CNN to correctly classify haemoglobin. The model predicted haemoglobin level >99% accuracy for three classes and >85% accuracy for ten classes. Using attention maps to understand the model's decision process, we identified regions of interest for the 3-class model. Four out of the nine regions were located in the cerebral cortex. Changes in the right thalamus and the white matter tracts between the temporal gyrus, putamen, hippocampus, and thalamus regions were associated to high haemoglobin class. Our model accurately predicted haemoglobin levels from SWI suggesting an association between the iron measures. Furthermore, it was possible to identify brain areas used by the CNN for decision-making. Using this integrated model of iron modalities will give a more comprehensive view of iron homeostasis and the potential associations with neurodegenerative diseases such as dementia.