Automated Detection of SDH and EDH due to TBI from CT-scan images using CNN

Traumatic brain injury (TBI) occurs when external forces are suddenly applied to the brain and disrupt the brain's normal functioning. This issue can lead to primary and secondary injuries, such as hematoma, which may cause severe disabilities and even death. Therefore, early and accurate diagn...

Full description

Saved in:
Bibliographic Details
Published in2023 30th National and 8th International Iranian Conference on Biomedical Engineering (ICBME) pp. 164 - 170
Main Authors Mazhari, Ava, Allahgholi, Ali, Shafieian, Mehdi
Format Conference Proceeding
LanguageEnglish
Published IEEE 30.11.2023
Subjects
Online AccessGet full text
DOI10.1109/ICBME61513.2023.10488625

Cover

More Information
Summary:Traumatic brain injury (TBI) occurs when external forces are suddenly applied to the brain and disrupt the brain's normal functioning. This issue can lead to primary and secondary injuries, such as hematoma, which may cause severe disabilities and even death. Therefore, early and accurate diagnosis is essential. Computed tomography (CT) is the appropriate imaging modality to evaluate intracranial hemorrhage. However, manual diagnosis of hematoma is an operator-dependent and time-consuming task, which can lead to a late or incorrect diagnosis. To address this issue, computer-aided diagnosis (CAD) systems have been developed to assist in the accurate and timely management of TBI. These systems are advantageous in early detection; deep learning, in particular, has shown promising results for automatic detection. Regardless, a challenging aspect in this field is diagnosing two types of hemorrhage: subdural and epidural hematoma, which requires a new, more accurate detecting approach. This study uses deep learning methods to detect intracranial hemorrhage using a Head CT scan to identify two types of hematoma subtypes. Using two convolutional neural networks (CNN), AlexNet and ResNet50 classifiers, for the detection process, the performance of these two architectures was compared, and finally, the best model was proposed and evaluated.
DOI:10.1109/ICBME61513.2023.10488625