Unsupervised Deep Learning for Stroke Lesion Segmentation on Follow-up CT Based on Generative Adversarial Networks.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neurology

Background:

• Supervised deep learning is standard for stroke lesion segmentation on non-contrast computed tomography (NCCT).
• Supervised methods necessitate manual annotations, which are time-consuming and labor-intensive.
• Unsupervised deep learning methods, like generative adversarial networks (GANs), offer an alternative by not requiring manual annotations.

Purpose of the Study:

• To develop and evaluate a GAN for segmenting infarct and hemorrhagic stroke lesions on follow-up NCCT scans.
• To assess the efficacy of unsupervised deep learning in stroke lesion segmentation.
• To provide an automated tool for lesion segmentation, reducing reliance on manual annotation.

Main Methods:

• A GAN was trained on 820 patient NCCT scans from acute ischemic stroke trials.
• The GAN transformed follow-up scans to resemble baseline scans by generating a difference map.
• Lesion segmentations were extracted from the difference map and evaluated using Dice similarity coefficient, Bland-Altman analysis, and intraclass correlation coefficient.

Main Results:

• The GAN achieved a median Dice similarity coefficient of 0.31 for 24-hour and 0.59 for 1-week infarct lesions.
• Segmentation performance for hemorrhagic lesions was significantly lower (median Dice: 0.02-0.08).
• Good volumetric correspondence was observed for infarct lesions (ICC 0.83-0.90).

Conclusions:

• Unsupervised GANs can automate infarct lesion segmentation on NCCT with moderate accuracy.
• The developed GAN shows potential for clinical application in stroke imaging analysis.
• Further research may improve segmentation of hemorrhagic lesions and refine unsupervised deep learning models.