A deep supervised cross-attention strategy for ischemic stroke segmentation in MRI studies.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Manual delineation of brain lesions from MRI is time-consuming and subjective.
• Accurate lesion localization is crucial for stroke diagnosis and patient management.

Purpose of the Study:

• To develop an automated method for brain lesion delineation using a novel autoencoder architecture.
• To address challenges in lesion detection, including class imbalance, complex shapes, and variable textures.

Main Methods:

• Implementation of a cross-attention deep autoencoder with hierarchical deep supervision.
• Utilizing convolutional saliency maps and skip connections to preserve lesion morphology.
• Employing a weighted loss function to mitigate class imbalance issues.

Main Results:

• The proposed autoencoder achieved state-of-the-art results on the ISLES2017 dataset.
• Achieved a Dice score of 0.36 and a precision of 0.42 in lesion delineation.
• Integration of ADC, TTP, and Tmax MRI sequences yielded the best performance.

Conclusions:

• Deeply supervised cross-attention autoencoders enhance the accuracy of ischemic lesion estimation in MRI.
• The method improves discrimination between healthy and lesion tissues, supporting patient follow-up.
• This automated approach offers a more objective and efficient tool for stroke diagnosis.