Fully Automatic Segmentation of Acute Ischemic Lesions on Diffusion-Weighted Imaging Using Convolutional Neural...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Accurate segmentation of acute ischemic lesions on diffusion-weighted imaging (DWI) is crucial for timely stroke treatment.
• Conventional segmentation algorithms often struggle with lesion variability and accuracy.

Purpose of the Study:

• To develop and evaluate convolutional neural network (CNN) algorithms for automatic segmentation of acute ischemic lesions on DWI.
• To compare the performance of CNN algorithms against conventional segmentation methods.

Main Methods:

• Retrospective analysis of 429 patients with acute cerebral ischemia.
• Development of CNN models (U-Net, DenseNet) with squeeze-and-excitation blocks for lesion segmentation.
• Comparison with thresholding-based algorithms using Dice index and 5-fold cross-validation.

Main Results:

• CNN algorithms demonstrated significantly superior performance (p < 0.001) compared to conventional methods.
• Achieved Dice indices of 0.85 (U-Net, DenseNet) and 0.86 (ensemble), versus 0.52-0.58 for conventional algorithms.
• CNNs showed high accuracy for both small (0.81-0.82) and large (0.88-0.89) lesions.

Conclusions:

• CNN-based algorithms provide highly accurate and superior automatic segmentation of acute ischemic lesions on DWI.
• These findings support the clinical utility of CNNs for improving stroke assessment and management.