The multi-level classification network (MCN) with modified residual U-Net for ischemic stroke lesions segmentation...

Area of Science:

• Medical Imaging
• Neurology
• Artificial Intelligence

Background:

• Ischemic and hemorrhagic strokes are critical brain injuries requiring accurate diagnosis.
• Manual analysis of 3D brain MRIs for stroke lesions is time-consuming and labor-intensive.
• The Anatomical Tracings of Lesions After Stroke (ATLAS) dataset presents challenges like class imbalance and inter-class similarity.

Purpose of the Study:

• To develop an automated method for segmenting ischemic stroke lesions in brain MRI scans.
• To address the significant class imbalance and inter-class similarity issues within the ATLAS dataset.
• To improve the efficiency and accuracy of stroke lesion detection compared to manual methods.

Main Methods:

• A multi-level classification network (MCN) comprising three cascaded networks was proposed.
• The first network reduces slice-level class imbalance by identifying relevant slices.
• Subsequent networks process overlapping patches for lesion classification and segmentation using a modified U-Net architecture.

Main Results:

• The MCN achieved a mean Dice score of 0.754 on the test dataset.
• This performance surpasses existing state-of-the-art methods on the same dataset.
• The proposed method effectively handles class imbalance and inter-class similarity.

Conclusions:

• The developed multi-level classification network offers a promising automated solution for ischemic stroke lesion segmentation.
• The MCN demonstrates superior performance in addressing key challenges in medical image analysis for stroke detection.
• This approach can significantly assist neurologists in the diagnosis and management of stroke patients.