Identifying Mild Traumatic Brain Injury via Vision Transformer and Bag of Visual Features

Due to lack of established criteria and reliable biomarkers, timely diagnosis of mild traumatic brain injury (mTBI) has remained a challenging problem. Widefield optical imaging of cortical activity in animals provides a unique opportunity to study injury-induced alterations of brain function. Motivated by the results of medical-imaging studies that employ patch-level-based approaches, this paper proposes to use two patch-based deep learning techniques for classifying brain images of mTBI and healthy Thyl-GCaMP6s transgenic mice. The first approach uses a Bag of Visual Word (BoVW) technique to represent each image as a histogram of local features derived from patches from all training data. The local features are extracted using an unsupervised convolutional autoencoder (CAE). The second approach employs a pre-trained vision transformer (ViT) model. The average accuracy for classifying mTBI and healthy brains for the CAE-BoVW and the ViT are 96.8% and 97.78%, respectively, outperforming results of a convolutional neural network (CNN) model. This work suggests that attention-based models can be utilized for the problem of classifying mTBI and healthy brain images.