few-shot segmentation

few-shot segmentation performs segmentation of volumetric medical images using few-shot deep neural networks with squeeze & excite mechanisms to enable accurate organ segmentation from limited annotated examples.

Key Features:

• Few-Shot Learning Framework: Learns new classes for volumetric medical images from a minimal number of annotated support examples.
• Novel Architecture (Conditioner and Segmenter Arms): A conditioner arm processes annotated support input to generate task-specific representations that the segmenter arm uses to segment query images.
• Squeeze & Excite Blocks: Integrates channel squeeze & spatial excitation modules to enable interaction between conditioner and segmenter arms and recalibrate channel-wise feature responses.
• Volumetric Segmentation Strategy: Pairs a few slices from the support volume with all slices of the query volume to handle three-dimensional medical scans.
• No Pre-trained Models Required: Operates without reliance on pre-trained networks.

Scientific Applications:

• Organ Segmentation in CT: Applied to organ segmentation in volumetric scans such as whole-body contrast-enhanced CT images.
• Benchmark Evaluation: Validated on the Visceral Dataset with performance reported as superior to multiple baselines.
• Medical Image Analysis: Suited for research and clinical tasks that require precise segmentation from limited annotated data.

Methodology:

Implements deep neural networks with channel squeeze & spatial excitation modules; the conditioner arm processes annotated support slices to produce task-specific representations used by the segmenter arm to segment query volumes; volumetric segmentation pairs a few support slices with all query slices; training does not rely on pre-trained networks.