TripletResNet: A Deep Metric Learning Approach for mTBI Diagnosis from 3D CT

Hanem Ellethy, Shekhar S. Chandra, Viktor Vegh

Abstract: Mild Traumatic Brain Injury (mTBI) is a prevalent neurological condition with significant diagnostic challenges. Accurate and timely diagnosis is critical to guiding effective treatment strategies and improving patient outcomes. Traditional diagnostic techniques often suffer from limited accuracy and sensitivity, leading to potential misdiagnoses. To overcome these limitations, we propose TripletResNet, a novel residual neural network trained using metric learning with triplet loss. Our method effectively distinguishes mTBI cases from healthy individuals by embedding 3D Computed Tomography (CT) scans into a discriminative feature space. Optimizes representations by maximizing distances between dissimilar cases and minimizing them between similar ones, thereby highlighting diagnostic features. Our TripletResNet achieved superior diagnostic performance, demonstrating an average accuracy of 94.3%, sensitivity of 94.1%, and specificity of 95.2% through five-fold cross-validation. Notably, compared to our previously introduced Residual Convolutional Neural Network (RCNN), TripletResNet exhibits significant performance improvements, a 22.5% increase in specificity, 16.2% higher accuracy, and 11.3% enhanced sensitivity. Additionally, TripletResNet requires fewer memory resources, offering a practical and efficient solution. The integration of occlusion sensitivity maps further enhances the interpretability of our model, providing visual explanations that build clinician trust and facilitate transparent informed decision-making.

External IDs:dblp:conf/ausai/EllethyCV25