Weakly-Interactive-Mixed Learning: Less Labelling Cost for Better Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation demands extensive pixel-level annotations, which are costly and time-consuming.
• Existing methods struggle with the high annotation burden for accurate segmentation.

Purpose of the Study:

• To propose a novel Weakly-Interactive-Mixed Learning (WIML) framework to reduce annotation efforts in medical image segmentation.
• To achieve high segmentation accuracy using limited annotations.

Main Methods:

• Developed a Weakly-Interactive Annotation (WIA) component to reduce strong label annotation time via interactive learning.
• Designed a Mixed-Supervised Learning (MSL) component utilizing weak and few strong labels for improved accuracy.
• Introduced a Full-Parameter-Sharing Network (FPSNet) integrating attention modules (scSE) and a Full-Parameter-Sharing (FPS) strategy.

Main Results:

• The WIML framework significantly reduces the time and cost associated with creating high-quality annotated datasets.
• FPSNet with integrated scSE modules improved class activation map (CAM) performance.
• The FPS strategy in FPSNet mitigated overfitting when trained with minimal strong labels.
• WIML-FPSNet demonstrated superior performance compared to state-of-the-art methods on BraTS 2019 and LiTS 2017 datasets.

Conclusions:

• The proposed WIML framework effectively addresses the challenge of limited annotated data in medical image segmentation.
• WIML-FPSNet offers a promising solution for accurate and efficient medical image segmentation with reduced annotation requirements.