Volumetric memory network for interactive medical image segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Automatic medical image segmentation often lacks clinical accuracy, necessitating manual refinement.
• Interactive segmentation methods aim to combine automated efficiency with human expertise for improved results.

Purpose of the Study:

• To develop a novel Volumetric Memory Network (VMN) for efficient and accurate interactive 3D medical image segmentation.
• To enable a robust human-in-the-loop system that refines segmentation based on user guidance and quality assessment.

Main Methods:

• A 2D interaction network generates initial segmentation from user hints on a single slice.
• The VMN propagates segmentation masks bidirectionally across the 3D volume.
• A quality assessment module guides user interaction by identifying slices needing refinement, enabling active learning.

Main Results:

• The VMN effectively encodes and retrieves segmentation information using a memory-augmented design.
• The quality assessment module facilitates preferential user labeling of low-quality segmentation slices for multi-round refinement.
• The VMN demonstrated superior performance across multiple public datasets (MSD, KiTS19, CVC-ClinicDB) compared to state-of-the-art models.

Conclusions:

• The proposed VMN provides a robust interactive segmentation engine for 3D medical images.
• The approach generalizes well to various user annotation types, including scribbles and bounding boxes.
• The VMN significantly enhances segmentation accuracy and efficiency in a clinically relevant manner.