Automatic Segmentation of Ischemic Stroke Lesions in CT Perfusion Maps Using Deep Learning Networks
Stroke is the third leading cause of death and the largest cause of acquired disability worldwide. Classification of stroke lesions is vital in recovery, diagnosis, outcome assessment, and treatment planning. The current standard approach for segmenting ischemic stroke lesions is based on thresholdi...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
University of Isfahan
2024-09-01
|
| Series: | هوش محاسباتی در مهندسی برق |
| Subjects: | |
| Online Access: | https://isee.ui.ac.ir/article_29109_4c9dae37a178345e520cd93facbff339.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Stroke is the third leading cause of death and the largest cause of acquired disability worldwide. Classification of stroke lesions is vital in recovery, diagnosis, outcome assessment, and treatment planning. The current standard approach for segmenting ischemic stroke lesions is based on thresholding of computed tomography perfusion (CTP) maps. However, this detection approach is inaccurate (the dice similarity coefficient is around 68%). Accordingly, several machine learning-based techniques have recently been proposed to improve the segmentation accuracy of ischemic stroke lesions. Although these studies have achieved significant results, they still need to be improved before being used in real practice. This research presents a new technique based on deep learning for the segmentation of ischemic stroke lesions in CTP maps. The proposed network architecture includes the 7 Graph Convolutional layer, which can automatically perform feature selection/extraction and classify the resulting feature vector. In this study, the ISLES 2018 database was used to train the proposed network. The indices of the Dice Similarity coefficient and Jaccard Index based on the proposed model are 75.41% and 74/52%, respectively, which is a significant improvement compared to recent studies. In addition, the performance of the proposed model in noisy environments is very promising; so, at SNR=-4, the accuracy of networks is still above 60%. |
|---|---|
| ISSN: | 2821-0689 |