Traumatic meningeal enhancement detection by deep learning-based biomedical image analysis and handcrafted features extraction

Abstract

Traumatic Meningeal Enhancement (TME) is a critical medical condition characterized by abnormal enhancement of the meninges following trauma, often observed in medical imaging studies. Traumatic meningeal injuries result from external forces hitting the head or skull, damaging the brain’s protective coverings. These injuries can come from falls, car accidents, sports injuries, attacks, or other head trauma. Even in the absence of further trauma-related cerebral abnormalities, TME may be visible on an acute MRI. In addition to highlighting some of the present considerations and unresolved issues of using them, this research aims to address some of the prospective applications of more sophisticated imaging in traumatic meningeal enhancement (TME). A deep convolutional neural network (CNN) model that uses a dataset of 7800 images is used in this study. Testing and training are the two discrete parts of the dataset. We have used the appropriate augmentation method to construct the dataset. Three categories have been used to categorize the data in this study: normal, early (pre), and acute (post). We divided the 6,000 images into three categories for training: normal, early (pre), and acute (post). 30% of the data was used for testing, while the remaining 70% was used for training. The dataset was evaluated against five different transfer learning models and a customized CNN model known as the 13-layered CNN model in the research. We evaluated four transfer learning models, namely VGG19, VGG16, InceptionV3, and MobileNet, using an identical dataset. The accuracy rates obtained were 84%, 86%, 80%, and 89% respectively. Utilizing the same dataset, we proceeded to ensemble these pretrained models and it obtained 88.83% accuracy. Surprisingly, even with the ensemble, our customized CNN model exhibited superior accuracy. Additionally, we conducted SVM and XG Boost hand-crafted feature extraction using techniques like positional orientation (PO), histogram of oriented gradients (HOG), and mean pixel value (MPV). SVM obtained accuray of PO,normal:67% early(pre): 65% and acute(post):67%, for HOG, normal:81% early(pre): 75% and acute(post):77%, for MPV, normal:71% early(pre): 70% and acute(post):70%. XGBoost obtained accuracy of PO,normal:63% early(pre): 60% and acute(post):57%, for HOG, normal:72% early(pre): 69% and acute(post):70%, for MPV, normal:66% early(pre): 63% and acute(post):62%. Subsequently, we applied Support Vector Machine (SVM) and XGBoost algorithms for feature extraction. Despite these efforts, our CNN model consistently outperformed the models built using these feature extraction methods. In contrast, our newly customized CNN model demonstrated a remarkable accuracy of 91%. These results illustrate that when it comes to image processing, our CNN model performs better than any other model in identifying traumatic meningeal brain enhancement.