A Stacked Generalization U-shape network based on zoom strategy and its application in biomedical image segmentation.

Area of Science:

• Biomedical imaging informatics
• Deep learning for medical image analysis

Background:

• Deep neural networks offer flexibility but are sensitive to initial conditions and limited data.
• Biomedical imaging often faces challenges with small labeled datasets, impacting model generalization.

Purpose of the Study:

• To develop and test a stacked generalization U-shape network (SG-UNet) for biomedical image segmentation.
• To address the limitations of deep learning models with small and noisy datasets in medical imaging.

Main Methods:

• Proposed a novel SG-UNet architecture utilizing a zoom strategy and multi-supervision.
• Employed hybrid features from multi-resolution images for segmentation and disease detection.
• Introduced a zoom loss function to focus training on challenging samples without pre-training.

Main Results:

• Achieved improved Dice coefficients (3.116%, 2.676%, 2.356%) and F2-scores (3.044%, 2.420%, 1.928%) across CT, colonoscopy, and histopathology datasets.
• Demonstrated slower diminishing marginal efficiency with limited rectal cancer CT data.
• Showcased comparable performance to state-of-the-art methods in gland segmentation.

Conclusions:

• The SG-UNet enables efficient training on small datasets without fine-tuning.
• The proposed algorithm achieves higher accuracy with reduced computational complexity compared to other stacked ensemble networks.
• Offers a robust solution for biomedical image segmentation challenges.