Fully automatic acute ischemic lesion segmentation in DWI using convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Stroke is a leading cause of disability and death, often involving acute ischemic lesions.
• Accurate diagnosis and treatment are crucial for managing stroke, but manual lesion segmentation in diffusion-weighted MR imaging (DWI) is challenging.
• Automated segmentation tools are needed to assist clinicians in localizing and quantifying stroke lesions.

Purpose of the Study:

• To develop and validate a novel framework for automatic segmentation of stroke lesions in DWI.
• To improve the efficiency and accuracy of stroke lesion detection and quantification.
• To establish a new benchmark for automated stroke lesion segmentation using deep learning.

Main Methods:

• A novel framework employing two cascaded convolutional neural networks (CNNs): EDD Net and MUSCLE Net.
• EDD Net, an ensemble of DeconvNets, performs initial lesion detection.
• MUSCLE Net refines detections to reduce false positives, enhancing segmentation accuracy.

Main Results:

• The framework achieved a mean Dice coefficient of 0.67 across a large dataset of 741 subjects.
• Mean Dice scores for small and large lesions were 0.61 and 0.83, respectively.
• A high lesion detection rate of 0.94 was achieved, demonstrating the system's effectiveness.

Conclusions:

• The proposed dual-CNN framework offers an effective solution for automated stroke lesion segmentation in DWI.
• This automated approach can aid clinicians in timely diagnosis and treatment planning for acute ischemic stroke.
• The study highlights the potential of deep learning in neuroimaging for improving stroke care.