Explainable brain age prediction: a comparative evaluation of morphometric and deep learning pipelines.

Area of Science:

• Neuroimaging and computational neuroscience
• Artificial intelligence in medicine

Background:

• Brain age is a key biomarker for brain health and early detection of neurodegenerative diseases.
• Current brain age prediction relies on morphometric features or deep learning (DL) on MRI scans.
• Limited systematic comparison exists for these methods' performance and interpretability.

Purpose of the Study:

• To comparatively evaluate morphometric feature-based and 3D convolutional neural network (CNN) pipelines for brain age prediction.
• To assess the interpretability of these pipelines using various eXplainable Artificial Intelligence (XAI) methods.
• To understand the clinical utility and variability of different XAI techniques.

Main Methods:

• Utilized a multisite neuroimaging dataset for model training and validation.
• Employed FreeSurfer for morphometric feature extraction and 3D CNNs for DL-based prediction.
• Applied SHAP, Grad-CAM, and DeepSHAP for interpretability analysis of both pipelines.

Main Results:

• Achieved comparable, state-of-the-art performance for both morphometric and DL pipelines on an independent test set.
• SHAP demonstrated the most consistent and interpretable results for the feature-based pipeline.
• DeepSHAP showed higher variability, and further assessment of Grad-CAM's clinical utility is needed.

Conclusions:

• Morphometric and DL pipelines provide similar performance for brain age prediction.
• XAI methods offer valuable insights into model predictions, with SHAP showing strong interpretability.
• Integrating XAI into clinical practice is vital for understanding and trusting neuroimaging biomarkers.