A Machine Learning Approach for Differential Diagnosis of Vascular and Non-Vascular Intracranial Hemorrhage in Non-Contrast CT Images
Introduction: Accurate diagnosis of acute Intracranial Hemorrhage (ICH) involving vascular and non-vascular bleeding has proven to be challenging due to the visual complexities in non-contrast Computed Tomography images (NCCT). Consequently, there has been a necessity for the adoption of novel techn...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Mashhad University of Medical Sciences
2025-04-01
|
| Series: | Iranian Journal of Medical Physics |
| Subjects: | |
| Online Access: | https://ijmp.mums.ac.ir/article_25937_c48a913a4452f1469e1ab3c88a9b70ce.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Introduction: Accurate diagnosis of acute Intracranial Hemorrhage (ICH) involving vascular and non-vascular bleeding has proven to be challenging due to the visual complexities in non-contrast Computed Tomography images (NCCT). Consequently, there has been a necessity for the adoption of novel techniques to address this issue, recently. This study aims to develop a new framework for automatic and accurate diagnosis of ICH and the ability of machine learning to differentiate vascular and non-vascular causes of Intracranial hemorrhages based on CT scan images without contrast material. Determining whether intracranial hemorrhage is vascular or non-vascular is clinically significant as it influences treatment decisions.
Material and Methods: In this retrospective study, NCCT images were gathered from a group of 370 patients, comprising 67 subjects with vascular bleeding and 303 with non-vascular bleeding. Radiomics features encompassing morphological, texture, and intensity-related characteristics, were extracted for every image slice. Subsequently, the effectiveness of five classification methods—namely, Support Vector Classifier (SVC), Logistic Regression (LR), Random Forest (RF), Decision Tree (DT) and K-Nearest Neighbors (KNN) was evaluated.
Results: Metrics for evaluating classification methods, sensitivity, specificity and accuracy for the Logistic Regression were 55%, 65% and 63%, respectively. The AUC-ROC in this model was 0.66, which is better than other methods with large margin.
Conclusion: In this study, an evaluation of five different classification methods revealed that all of them exhibited sufficient level of specificity. However, when it comes to classification sensitivity and accuracy, the Logistic Regression approach outperformed the others. |
|---|---|
| ISSN: | 2345-3672 |