University:
Boston University
Course:
GMS BC 630 | Ethical & Policy Issues in Health and Medical Services Emergency Management
Academic year:

2024
Views:

232

Pages:

1
Author:

SpaceCadet101

TRAINING A CONVOLUTIONAL NEURAL NETWORK ON MRI SCANS TO DIFFERENTIATE BETWEEN PTSD & TBI = INTRODUCTION Post Traumatic Stress Disorder (PTSD): A neurological condition triggered by a distressing event characterized by ﬂashbacks, nightmares, and anxiety Traumatic Brain Injury (TBI): Acquired brain damage caused by trauma such as direct impacts, acceleration-deceleration, or shock waves ➢ PTSD & TBI share many symptoms and have a high rate of comorbidity ○ Diﬀicult to distinguish through traditional methods ○ Diﬀerent treatments ■ Transcranial magnetic stimulation is beneﬁcial to TBI patients but can induce seizures for those with PTSD ➢ Compared to civilian controls, veteran populations display higher rates of PTSD, TBI, and comorbid PTSD and TBI ○ Veterans are a primary potential beneﬁciary of this research The goal of this research is to utilize machine learning methods to classify patients with PTSD, TBI, and comorbid PTSD & TBI Fig 1: PTSD & TBI Symptomatology METHODS ➢ Convolutional Neural Network (CNN) ○ A machine learning model commonly used for image classiﬁcation that employs kernels (or ﬁlters) that slide across the input to create a feature map ➢ Training data comes from the University of Southern California ○ ~500 3D MRI scans from Vietnam War veterans ○ 4 classes: PTSD, TBI, comorbid PTSD & TBI, and control ➢ Code written in python using Tensorﬂow, Keras ○ 5,963,291 trainable parameters ○ 3 convolutional layers, 1 dense layer ➢ Training/testing split of 80/20 ➢ Model assessed on metrics of accuracy, area under the curve (AUC) Fig 2: Convolutional layer illustrating the conversion from input to feature map through the kernel Fig 3: Flow chart representing the sequential architecture of the CNN Fig 4 RESULTS Data Files Conv 3D Batch Normalization Preprocessing Activation Function (RELU) x3 Convolutional Layers Flatten Dense Layer Dense Layer Batch Normalization Output Activation Function (RELU) ➢ The model achieved an accuracy of 75% and an area under the curve (AUC) of 0.8125 after training ➢ AUC is the area under the Receiver Operating Characteristic (ROC) curve, the graph of the True Positive Rate (TPR) against the False Positive Rate (FPR). ■ This can be a better measure of performance as it’s more resilient to uneven distributions of data ■ Multiclass AUC can be calculated with a One vs One or One vs Rest approach through averaging ■ Higher is better ➢ In the graph, the decreasing loss shows the minimization of the cost function via gradient descent Fig 4: Example coronal, mid-sagittal, and horizontal MRI scans Fig 5: Graph of loss and val loss over training epochs DISCUSSION/CONCLUSION ➢ Models like this one are expected to assist in the process of diagnosing whether patients have PTSD, TBI, or comorbid PTSD & TBI ➢ Distinction is crucial as TBI treatments can prove ineﬀective or even harmful to those with PTSD or vice versa ➢ Limitations: Some limitations of this research include its nature as a retrospective analysis and the relatively small sample size ➢ Future Directions: In the future, we would like to compare CNNs with other models in hopes of improving performance. Additional preprocessing steps such as adding noise could also be used to increase the amount of training data ➢ Further exploration of this topic will lead to easier and lower-risk personalized patient treatment