CGNet: Few-shot learning for Intracranial Hemorrhage Segmentation

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Pathology

Background:

• Deep learning for medical image segmentation requires extensive annotated data.
• Manual annotation of Intracranial Hemorrhage (ICH) datasets is time-consuming and expensive.
• Few-shot segmentation offers a promising solution for data-scarce medical imaging tasks.

Purpose of the Study:

• To develop a novel few-shot segmentation model (CGNet) for efficient Intracranial Hemorrhage (ICH) segmentation.
• To address the challenge of limited annotated data in medical image analysis.
• To enhance the understanding of lesion details and refine segmentation accuracy.

Main Methods:

• Transformed the ICH segmentation task into a few-shot learning problem.
• Designed CGNet incorporating a Cross Feature Module (CFM) for enhanced feature interaction between query and support sets.
• Implemented a Support Guide Query (SGQ) module to integrate multi-scale features for refined segmentation.

Main Results:

• CGNet demonstrated superior performance on both the BHSD and IHSAH datasets.
• Achieved Dice coefficient score improvements of 3% and 1.8% over state-of-the-art few-shot models.
• Outperformed fully supervised segmentation models when trained on equivalent limited data.

Conclusions:

• CGNet effectively segments Intracranial Hemorrhage (ICH) using limited annotated data.
• The proposed CFM and SGQ modules enhance segmentation detail and accuracy.
• CGNet presents a viable alternative to traditional fully supervised methods in data-limited scenarios.