Text-guided few-shot liver and tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Digital Oncology

Background:

• High-precision liver and tumor segmentation is crucial for digital oncology but faces challenges with limited annotations and cross-center domain shifts.
• Existing few-shot learning methods struggle to generalize across diverse clinical settings due to reliance on visual similarity.

Purpose of the Study:

• To develop a text-guided few-shot segmentation framework that leverages clinical semantic information to improve segmentation accuracy and robustness.
• To address data scarcity and domain shift issues in medical image segmentation.

Main Methods:

• Proposed a framework integrating an Automated Semantic Generator, Text-Guided Gating (TGG) mechanism, and Decoupled Prototype Learner.
• Utilized large-scale vision-language models for semantic encoding and adaptive modulation of visual representations.
• Employed per-image averaging and gradient detachment for unbiased class prototype construction.

Main Results:

• The text-guided framework outperformed state-of-the-art supervised, foundation-model-based, and few-shot baselines on LiTS and 3DIRCADb datasets.
• Achieved significant improvements in external liver Dice (8.7 pp) and tumor Dice (26.3 pp) on the 3DIRCADb dataset compared to the strongest few-shot baseline.
• Demonstrated effective mitigation of performance degradation seen in conventional supervised models.

Conclusions:

• Cross-modal semantic guidance significantly enhances robust medical image segmentation, particularly under domain shift.
• The proposed framework offers a data-efficient and robust solution for clinical deployment in digital oncology.