FSS-ULivR: a clinically-inspired few-shot segmentation framework for liver imaging using unified representations and...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Computer Vision

Background:

• Accurate liver segmentation is crucial for medical diagnosis and treatment planning.
• Existing segmentation methods face challenges with limited data, poor generalizability, and feature integration.
• There is a need for robust models that can perform precise segmentation with scarce annotated data.

Purpose of the Study:

• To develop a novel Few-Shot Segmentation model with Unified Liver Representation (FSS-ULivR) for precise liver segmentation.
• To enhance feature learning, dynamic feature refinement, and spatial detail recovery in segmentation models.
• To demonstrate the model's effectiveness and generalizability across various medical imaging datasets.

Main Methods:

• Proposed FSS-ULivR model utilizing a ResNet-based encoder with Squeeze-and-Excitation modules.
• Incorporated an enhanced prototype module with a transformer block and channel attention.
• Employed a decoder with attention gates and residual refinement strategies for spatial detail recovery.

Main Results:

• Achieved high performance on the Liver Tumor Segmentation Challenge dataset with Dice coefficient of 98.94% and IoU of 97.44%.
• Demonstrated strong generalizability across multiple datasets (3DIRCADB01, Colorectal Liver Metastases, CT-ORG, MS D Task 3) with Dice scores above 90% in most cases.
• Showcased effectiveness in multi-organ segmentation and brain tumor segmentation tasks, indicating broad applicability.

Conclusions:

• The FSS-ULivR model provides precise and reliable medical image segmentation, particularly for the liver.
• The model's architecture effectively addresses limitations of existing methods, especially in low-data scenarios.
• The demonstrated generalizability and high performance highlight its clinical significance and potential for AI-driven healthcare solutions.