AI-Powered Detection of Stroke in Non-Contrast Brain CT Images
Vikramaditya Tarai1, Nabin Chandra Sahu2, Sonam Pattanaik3 , Gayatree Maharana4
1Department of M Sc Computer Science, Sophitorium Institute Of Technology And Lifeskills, Odisha, India
2Department of Mechanical Engineering, Sophitorium Engineering College, Odisha, India
3 Department of Computer Science, Sophitorium Engineering College, Odisha, India
4 Department of M Sc Computer Science, Sophitorium Institute Of Technology And Lifeskills, Odisha, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract -The identification and categorization of brain tumors represent essential components in the enhancement of human health. A range of medical imaging modalities is capable of detecting brain pathologies associated with hemorrhagic events, including nuclear magnetic resonance, ultrasound, X-ray imaging, radionuclide imaging, laser technology, electron imaging, and photonic methods. Among these modalities, magnetic resonance imaging (MRI) is preferred within the medical community due to its superior image resolution and the absence of ionizing radiation. The interpretation of MRI scans is facilitated by the application of Artificial Intelligence, which alleviates the labor-intensive and time-consuming nature of this process. The efficacy of deep learning and convolutional neural networks in the identification of brain tumors has been well established. In the present study, deep learning transfer methodologies are employed for the classification of tumors. Specifically, ResNet-50 models with pre-training on convolutional neural networks are utilized to automatically predict and categorize brain tumors. The dataset comprises head CT (Computed Tomography) images in JPG format, featuring 2,500 images from the brain window and 2,500 images from the bone window, sourced from 82 patients. The performance of the ResNet-50 models is assessed in comparison to Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and You Only Look Once (YOLO) algorithms, with experimental findings substantiating the effectiveness of the ResNet-50 approach. Consequently, the application of ResNet-50 for tumor classification is affirmed and endorsed.
Key Words:ResNet-50, Support Vector Machines, You Only Look Once, Computed Tomography, Artificial Intelligence.