Efficient contour-based annotation by iterative deep learning for organ segmentation from volumetric medical images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural network training necessitates extensive human-annotated data, which is time-consuming and inefficient for volumetric medical image organ segmentation.
• Current annotation by iterative deep learning (AID) methods face challenges with long training times and high annotation burdens due to limited domain knowledge integration and inefficient human-interaction tools.

Purpose of the Study:

• To develop an efficient and accurate annotation by iterative deep learning (AID) algorithm for organ segmentation in volumetric medical images.
• To reduce the human annotation burden and accelerate the training process for deep learning models in medical imaging.

Main Methods:

• Developed a contour-based annotation by iterative deep learning (AID) algorithm utilizing boundary representation to integrate organ shape knowledge.
• Proposed a contour segmentation network with a multi-scale feature extraction backbone for enhanced boundary detection accuracy.
• Introduced a contour-based human-intervention method for efficient adjustment of organ boundaries, enabling fast few-shot learning and human proofreading.

Main Results:

• The contour-based AID method significantly reduced annotation time and inter-rater variability compared to traditional contour-interpolation and state-of-the-art (SOTA) voxel-label-based CNN methods.
• The developed contour detection network demonstrated superior performance over SOTA nnU-Net in generating anatomically plausible organ shapes using a small training set.
• Validation involved two human operators annotating abdominal organs in CT images, confirming the method's efficiency and accuracy.

Conclusions:

• The contour-based AID method offers improved efficiency, accuracy, and reduced inter-operator variability for organ segmentation in volumetric medical images compared to existing SOTA AID methods.
• The algorithm's robust shape learning capabilities and flexible boundary adjustment make it well-suited for rapid annotation of regular-shaped organs.
• This approach effectively leverages boundary shape priors and contour representation for streamlined medical image annotation.