Transformer Dil-DenseUnet: An Advanced Architecture for Stroke Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate segmentation of stroke lesions in MRI is crucial for patient diagnosis and treatment.
• Manual segmentation is time-consuming and prone to errors, necessitating automated solutions.
• Existing automated methods struggle with capturing complex spatial relationships and long-range dependencies.

Purpose of the Study:

• To introduce Transformer Dil-DenseUNet, a novel deep learning architecture for precise stroke lesion segmentation in MRI.
• To evaluate the performance of Transformer Dil-DenseUNet on benchmark datasets for stroke lesion segmentation.
• To demonstrate the superiority of the proposed architecture over existing state-of-the-art methods.

Main Methods:

• The Transformer Dil-DenseUNet architecture integrates DenseNet for feature extraction, dilated convolutions for expanded receptive fields, and Transformer blocks for modeling long-range dependencies.
• DenseNet utilizes dense connections to enhance feature reuse and capture fine-grained details.
• Transformer blocks employ multi-head self-attention mechanisms to address limitations of Convolutional Neural Networks (CNNs) in capturing global context.

Main Results:

• The Transformer Dil-DenseUNet model achieved a Dice coefficient of 0.80 ± 0.30 on the SISS 2015 dataset.
• The model obtained a Dice coefficient of 0.81 ± 0.33 on the ISLES 2022 dataset.
• These results surpass current state-of-the-art performance on both evaluated datasets.

Conclusions:

• The Transformer Dil-DenseUNet architecture offers a significant advancement in automated stroke lesion segmentation from MRI.
• The model's ability to integrate dense connections, dilated convolutions, and self-attention mechanisms leads to improved accuracy.
• This approach holds promise for enhancing clinical diagnosis and treatment strategies for stroke patients.