A Unified Framework for Generalized Low-Shot Medical Image Segmentation With Scarce Data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks (DNNs) have advanced medical image segmentation but require extensive data and annotations.
• Acquiring large annotated datasets is challenging and costly, particularly for rare diseases.

Purpose of the Study:

• To propose a unified framework for generalized low-shot (one- and few-shot) medical image segmentation.
• To address the challenge of extreme data and annotation scarcity, crucial for rare disease applications.

Main Methods:

• Utilizing distance metric learning (DML) to learn multimodal mixture representations for each category.
• Employing cosine distances between pixel embeddings and category representations for dense predictions.
• Introducing adaptive mixing coefficients to emphasize relevant modes in multimodal distributions.

Main Results:

• Achieved superior performance in low-shot segmentation on brain MRI and abdominal CT datasets.
• Demonstrated significantly higher mean Dice coefficients (e.g., 81%/69%) compared to 3D U-Net (52%/31%) and ANTs (72%/35%) with only one training sample.
• Effectively overcame overfitting issues with extremely limited data by leveraging inter-subject similarities and intraclass variations.

Conclusions:

• The proposed DML-based framework offers a robust solution for low-shot medical image segmentation, especially in data-scarce scenarios.
• The method shows significant potential for applications involving rare diseases and limited annotated medical data.
• The adaptive multimodal representations provide an effective way to handle limited training samples in medical image analysis.