Dynamic pixel-wise weighting-based fully convolutional neural networks for left ventricle segmentation in short-axis...

Area of Science:

• Medical Imaging
• Cardiovascular Diseases
• Artificial Intelligence in Medicine

Background:

• Accurate left ventricle (LV) segmentation in cardiac MRI is crucial for diagnosing cardiovascular diseases.
• Current segmentation methods face challenges, especially with apical and basal slices.

Purpose of the Study:

• To develop a novel, fully automatic LV segmentation method using convolutional neural networks.
• To improve segmentation accuracy by integrating multi-scale features and a dynamic pixel-wise weighting strategy.

Main Methods:

• A novel convolutional neural network architecture integrating hierarchical and multi-scale features.
• Implementation of a dynamic pixel-wise weighting strategy to focus on misclassified pixels.
• Validation using the CAP database for cine MR images.

Main Results:

• The proposed method achieves substantial improvement in LV segmentation compared to existing deep learning approaches.
• Enhanced performance observed particularly in apical and basal slices of cine MR images.
• Demonstrated effectiveness on the CAP database.

Conclusions:

• The novel deep learning approach significantly enhances automatic left ventricle segmentation in cardiac MRI.
• The dynamic pixel-wise weighting strategy is key to improving accuracy on challenging image slices.
• The study also identifies limitations in current CNN-based semantic segmentation for LV analysis.