Accelerating 2D Abdominal Organ Segmentation with Active Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Abdominal CT imaging assesses body habitus and organ health, requiring semantic segmentation of key structures.
• Deep learning excels at 3D abdominal CT segmentation but is unsuitable for low-dose 2D imaging.
• 2D segmentation faces challenges from lack of 3D context and image level variability, with manual annotation being resource-intensive.

Purpose of the Study:

• To develop an efficient active learning framework for 2D abdominal organ segmentation.
• To reduce the manual annotation effort required for deep learning models in low-dose CT.
• To improve the performance of 2D abdominal organ segmentation by dynamically selecting 'hard cases' for annotation.

Main Methods:

• Designed a gradient-based active learning framework with meta-parameterized exemplar optimization.
• Dynamically selected 'hard cases' to improve annotation efficiency.
• Evaluated performance on the Baltimore Longitudinal Study on Aging (BLSA) cohort, comparing active learning with random selection.

Main Results:

• The active learning framework achieved comparable Dice scores with significantly less data compared to random selection.
• Random selection required 4.4 times more data to reach a Dice score of 0.97 compared to the proposed active learning method.
• Demonstrated maximization of manual annotation efficacy and accelerated learning.

Conclusions:

• The proposed active learning framework significantly reduces annotation effort for 2D abdominal organ segmentation.
• This method enhances the efficiency of deep learning model training for low-dose CT applications.
• The framework offers a practical solution for resource-intensive annotation tasks in medical imaging.