Multi-Organ Segmentation Over Partially Labeled Datasets With Multi-Scale Feature Abstraction.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Limited fully annotated datasets hinder deep learning for image segmentation, especially in multi-organ scenarios.
• Partially labeled datasets present a challenge for training robust segmentation models.

Purpose of the Study:

• To propose a unified training strategy for a multi-scale deep neural network using multiple partially labeled datasets for multi-organ segmentation.
• To introduce a novel network architecture for effective multi-scale feature abstraction and fusion.

Main Methods:

• Developed a Pyramid Input Pyramid Output Feature Abstraction Network (PIPO-FAN) with a U-shape pyramid structure.
• Incorporated an equal convolutional depth mechanism to bridge semantic gaps between scales.
• Utilized deep supervision and an adaptive weighting layer for refined, fused multi-scale outputs.

Main Results:

• Achieved very promising performance in multi-organ segmentation across four public datasets (BTCV, LiTS, KiTS, Spleen).
• Demonstrated the effectiveness of the proposed unified training strategy and network architecture.
• Publicly shared source code to encourage reproducibility and further development.

Conclusions:

• The PIPO-FAN model offers a viable solution to the challenge of limited annotated data in multi-organ segmentation.
• The proposed mechanisms effectively integrate multi-scale features for improved segmentation accuracy.
• This work facilitates advancements in deep learning-based medical image segmentation.