Abstract
Structural and functional MRI (s/fMRI) are extremely useful methods to measure brain structure and function non-invasively. Unfortunately, review studies have suggested that many s/fMRI studies have poor reproducibility and poor ability to infer causal mechanisms. This limitation is contributed by the high complexity and high Signal-to-noise ratio of MRI data.
In recent years, the advancement in Deep Convolution Neural Networks (DCNN) have seen an exponential growth in research on large datasets resulting in increasing number of real-world applications. The advantages of DCNN are the ability to recognize and learn relevant features in complex datasets. This is opposed to traditional machine learning methods which involve manual decision making, particularly in feature selection and analysis. However, due to the architectural inability to explain the decision-making processes, DCNN is often refers to as "black box".
One way to overcomes this limitation is to implement a Class Activation Map (CAM). Unlike traditional CNN, CAM network keeps the spatial information, which in turn allows extraction of activation maps (a map of Region-of-Interest, ROI, that the network deems critical in the classification).
In this study, CAM networks were developed for s/fMRI scan where it was trained to classify clinical/neurological grouping. Activation maps were extracted from a structural T1 scans and quantitative MRI of 21 developmental prosopagnosics (DP) and 17 healthy controls. The accuracy at classifying DP vs. healthy based on leave-one-participant-out was 66%. Similarly, we also validated this approach on structural T1 and resting state fMRI comparing 40 individuals with history of Traumatic Brain Injury (TBI) to 40 matched controls without TBI.
This analytical method will allow automated as well as meaningful and robust data mining of large datasets, which is ideal for s/fMRI application.