Deep Learning-Based Acute Ischemic Stroke Lesion Segmentation Method on Multimodal MR Images Using a Few Fully...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Neurology
• Neuroimaging

Background:

• Acute ischemic stroke (AIS) requires prompt diagnosis and quantitative lesion evaluation for effective treatment.
• Current deep learning methods for AIS lesion segmentation on MRI often demand extensive fully labeled datasets, which are time-consuming to create.
• The need for efficient segmentation techniques that minimize reliance on fully annotated data is critical.

Purpose of the Study:

• To develop an automated method for segmenting acute ischemic stroke lesions on MRI.
• To address the challenge of limited fully labeled data by proposing a hybrid approach.
• To improve the efficiency and reduce the annotation burden in AIS lesion segmentation.

Main Methods:

• Proposed a multifeature map fusion network (MFMF-Network) with a dual-branch architecture.
• Utilized a large dataset of weakly labeled subjects to train the classification branch.
• Employed a small set of fully labeled subjects to fine-tune the segmentation branch.

Main Results:

• Achieved a mean Dice coefficient of 0.699 ± 0.128 on a test set of 179 subjects.
• Demonstrated strong performance with a lesion-wise F1 score of 0.886.
• Attained a subject-wise detection rate of 1, indicating high accuracy in identifying affected subjects.

Conclusions:

• The proposed MFMF-Network effectively segments AIS lesions using a combination of weakly and fully labeled data.
• This approach significantly reduces the dependency on fully annotated datasets, making AIS lesion segmentation more practical.
• The method shows promising results for accurate and efficient AIS diagnosis in clinical settings.