Abstract
This paper proposes a novel framework called Bag of Adversarial Features (BAF) for identifying mild traumatic brain injury (MTBI) from diffusion magnetic resonance images (MRIs). MTBI is a significant public health issue, yet accurate diagnostic measures are currently lacking as standard scans often appear normal. Unlike traditional methods that rely on hand-crafted statistical features, our approach employs unsupervised feature learning via adversarial auto-encoders to extract discriminative patch-level representations. These features are aggregated using a bag-of-words (BoW) approach to create a fixed-length signature for classification. Experimental results on a dataset of 227 subjects demonstrate that the BAF framework achieves 83.8% accuracy, significantly outperforming conventional methods based on regional mean values.
Introduction
Mild Traumatic Brain Injury (MTBI), or concussion, affects over 1.7 million people annually in the USA. Diagnosis remains challenging because it typically relies on clinical history and self-reported symptoms; standard CT and MRI scans often fail to show visible abnormalities. While advanced Diffusion MRI (dMRI) can detect microscopic tissue changes, automated classification is hindered by small medical datasets and the limitations of "hand-crafted" features, such as regional averages, which often obscure subtle, localized injury patterns.
Methodology
To address these challenges, we developed a multi-stage pipeline that leverages unsupervised learning to maximize the utility of available data.
The BAF Pipeline
The process begins with the extraction of thousands of overlapping 16x16 patches from brain regions highly implicated in MTBI, specifically the thalamus and the splenium of the corpus callosum (sCC).
Unsupervised Feature Learning and Aggregation
We utilize an Adversarial Auto-encoder (AAE) with a convolutional structure to learn patch-level features without requiring manual labels. The AAE compresses patches into a latent representation while an adversarial loss forces this representation to follow a prior distribution (e.g., Gaussian), ensuring more robust and discriminative features.
These learned features are then aggregated using a Bag of Visual Words (BoW) approach. This method treats the brain as a collection of local patterns, generating a histogram that represents the frequency of specific "visual words" found within the tissue. This signature, combined with demographic data, is processed through feature selection and classified using a Support Vector Machine (SVM).
Experimental Results
The study utilized a dataset of 227 subjects (109 MTBI patients and 118 healthy controls). We derived 11 different metrics from multi-shell diffusion MRI, including DTI, DKI, and WMTI metrics.
- Accuracy: The BAF framework achieved 83.8% accuracy, compared to only 76.6% for traditional methods using a subset of statistical features.
- Sensitivity: The model demonstrated a high sensitivity of 86.1%, which is vital for minimizing false negatives in clinical diagnostics.
- Key Indicators: Feature selection identified kurtosis measures in the thalamus and axonal diffusion (DA) in the splenium as the most discriminative indicators of injury.
Conclusion
The Bag of Adversarial Features approach successfully bypasses the "small data" problem in medical AI by using unsupervised pre-training. By focusing on local patch-level patterns rather than global averages, the model can detect the subtle signatures of MTBI that are often invisible to standard diagnostic techniques. Future work will explore expanding the model to include more brain regions and validating its ability to predict long-term patient outcomes.
Summary updated