Classification of Acute Intracerebral Hemorrhage Using Radiomics on Brain Computed Tomography Images

• Published in: 2023 15th International Conference on Information Technology and Electrical Engineering (ICITEE) pp. 1 - 6
• Main Authors: Thabarsa, Phattanun, Angkurawaranon, Salita, Madla, Chakri, Vuthiwong, Withawat, Unsrisong, Kittisak, Inkeaw, Papangkorn
• Format: Conference Proceeding
• Language: English
• Published: IEEE 26.10.2023
• Subjects: Acute Intracerebral Hemorrhage; Brain modeling; Computational modeling; Computed tomography; Machine Learning; Manuals; Predictive models; Radiomics; Sensitivity; Support vector machines
• ISSN: 2766-0419
• DOI: 10.1109/ICITEE59582.2023.10317766

Acute intracerebral hemorrhage (ICH) entity accounts for 10 to 15% of all strokes and is associated with a higher mortality rate ischemic stroke or subarachnoid hemorrhage. Causes of ICH are divided into primary, and secondary, including vascular malformation and tumorous. Primary ICH accounts for approximately 80% of all ICH cases. Vascular anomalies rank as the second most common cause of spontaneous ICH overall. Furthermore, hemorrhage resulting from brain tumors can occur in up to 10% of all primary or metastatic tumors. Early recognizing of these three causes of bleeding is critical for clinicians in precise diagnosis, effective treatment management, and helps avoid delayed diagnosis. We proposed a radiomics approach for classifying multiple causes of acute ICH as vascular malformation, tumorous, and primary-related hematoma. Non-contrast brain computed tomography with clinical features was used as input. The regions of both hematoma and perihematomal edema were delineated by using manual segmentation approach. Four feature selection methods were adopted. Also, three classification models were investigated in this study. The results showed that using the features selected by F-value applied with SVM classifier outperformed the other models, achieving weighted average accuracy (± SD) of 0.84 (± 0.07). Additionally, the model demonstrated average sensitivity and positive predictive value of 0.84 (± 0.06) and 0.86 (± 0.05), respectively. We also evaluate the overall performance of discriminating each class from the rest using AUC. The result suggested that our proposed model achieved the weighted average AUC of 0.90. Our proposed method highlights the potential in identifying multiple causes of acute and nontraumatic ICH, which has not been previously explored.