Automatic Segmentation of Acute Ischemic Stroke From DWI Using 3-D Fully Convolutional DenseNets.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Acute ischemic stroke is a prevalent cerebrovascular disease in aging populations.
• Accurate infarct localization and quantification are crucial for timely diagnosis and treatment, but manual methods are time-consuming.
• Current diagnostic tools require improvement for efficiency and accuracy in stroke lesion identification.

Purpose of the Study:

• To develop and evaluate an automatic method for segmenting acute ischemic stroke lesions from diffusion-weighted images (DWIs).
• To leverage deep 3-D convolutional neural networks (CNNs) for efficient and accurate stroke lesion detection.
• To improve upon existing methods for stroke lesion segmentation, aiming for clinical applicability.

Main Methods:

• A novel deep 3-D CNN architecture with dense connectivity was employed for end-to-end feature learning.
• The model was trained using a Dice objective function to address class imbalance in DWI data.
• A dataset of 242 subjects with acute ischemic stroke was used for training, validation, and testing.

Main Results:

• The proposed method achieved a Dice similarity coefficient of 79.13%, lesionwise precision of 92.67%, and lesionwise F1 score of 89.25%.
• Performance surpassed other state-of-the-art CNN methods, demonstrating superior accuracy in stroke segmentation.
• Validation on the ISLES2015-SSIS dataset confirmed the model's generalization capacity and competitive performance.

Conclusions:

• The developed automatic segmentation method using deep 3-D CNNs is fast, accurate, and effective for acute ischemic stroke.
• The approach shows significant potential for integration into clinical routines, improving diagnostic efficiency.
• This data-driven method offers a promising advancement in neuroimaging analysis for stroke patients.