Co-optimization Learning Network for MRI Segmentation of Ischemic Penumbra Tissues

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neuroscience imaging

Background:

• Supervised Convolutional Neural Networks (CNNs) show promise in medical image auxiliary diagnosis.
• A significant limitation is the scarcity of labeled medical image data, which is expensive and time-consuming to acquire.
• This data shortage hinders the performance of supervised CNNs in medical image segmentation tasks.

Purpose of the Study:

• To address the challenge of limited labeled data in medical image segmentation.
• To propose a novel co-optimization learning network (COL-Net) for Magnetic Resonance Imaging (MRI) segmentation of ischemic penumbra tissues.
• To enhance the accuracy and generalization of medical image segmentation models with minimal labeled samples.

Main Methods:

• Developed a co-optimization learning network (COL-Net) comprising an unsupervised reconstruction network (R), a supervised segmentation network (S), and a transfer block (T).
• The reconstruction network extracts robust features from pseudo-unlabeled samples to aid the segmentation network.
• A mix loss function was employed to co-optimize feature maps between the reconstruction and segmentation networks' bottlenecks.

Main Results:

• COL-Net achieved a Dice coefficient of 0.79 on the public ischemic penumbra segmentation challenge (SPES) using limited labeled samples.
• Demonstrated high predictive accuracy and generalization capabilities.
• Outperformed most supervised segmentation methods in extended experiments.

Conclusions:

• COL-Net effectively alleviates the problem of limited labeled samples in medical image segmentation.
• Presents a meaningful advancement for auxiliary diagnosis in medical imaging.
• Highlights the potential of co-optimization strategies in deep learning for healthcare applications.