Automated Final Lesion Segmentation in Posterior Circulation Acute Ischemic Stroke Using Deep Learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Stroke Research

Background:

• Final lesion volume (FLV) is a crucial outcome measure in anterior circulation stroke (ACS).
• Automatic FLV quantification in posterior circulation stroke (PCS) is underdeveloped due to limited methods and lower incidence.
• Deep learning applicability in PCS is hindered by data scarcity compared to ACS.

Purpose of the Study:

• To develop and evaluate strategies for a convolutional neural network (CNN) for PCS lesion segmentation.
• To compare the performance of CNNs trained on ACS data, PCS data, combined data, and via transfer learning.
• To determine the optimal approach for accurate FLV quantification and lesion detection in PCS.

Main Methods:

• Utilized follow-up non-contrast CT scans from 1018 ACS and 107 PCS patients.
• Trained CNNs on ACS data (ACS-CNN), PCS data, combined ACS/PCS data, and fine-tuned ACS-CNN using PCS data (transfer learning).
• Evaluated strategies based on volume agreement (intra-class correlation) and lesion detection rate.

Main Results:

• Transfer learning demonstrated superior performance with an intra-class correlation of 0.88 (95% CI: 0.83-0.92) for volume agreement.
• The transfer learning strategy achieved a lesion detection rate of 87%, significantly outperforming other methods (41-77%).
• Other strategies included ACS-CNN generalization, PCS-only training, and combined dataset training.

Conclusions:

• Transfer learning significantly enhances FLV quantification and detection rates for PCS lesions.
• Adapting ACS-trained CNNs via fine-tuning is a more effective strategy for PCS lesion segmentation than other evaluated methods.
• This approach addresses the understudy of FLV in PCS, potentially improving outcome assessment.