Keywords: 3DCNN, MRI, saliency maps, GradCAM, augmentation, gradcam
TL;DR: Applying saliency methods for 3DCNN interpretation to avoid overfitting of gender prediction models on healthy subjects.
Abstract: MRI-based prediction models are one of the most exploited AI solutions in neurology. Numerous computer-vision models showed their predictive ability for diverse psychoneurological conditions. Although most of these models are based on weak or no annotation, only a few reported studies interpret the predictions and perform the model saliency regions' analysis. We utilize 3DCNN interpretation with GradCAM to explore learned patterns for basic demographic characteristics prediction on the healthy cohort. We compare the saliency maps of the gender prediction models with the different types of MRI data preprocessing and augmentation. We assess the quality of learned patterns and examine the ways of models overfitting. We propose a data augmentation strategy based on optimal transport to avoid model overfitting on the brain volumes.
Paper Type: methodological development
Primary Subject Area: Interpretability and Explainable AI
Secondary Subject Area: Detection and Diagnosis
Paper Status: original work, not submitted yet
Source Code Url: https://github.com/PolinaDruzhinina/CrystalNeuroimaging
Data Set Url: https://www.humanconnectome.org/study/hcp-young-adult
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.