Online Hard Patch Mining Using Shape Models and Bandit Algorithm for Multi-Organ Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Hard sample selection is crucial for improving model convergence in machine learning.
• Existing sampling strategies for 3D medical images are computationally expensive and may not adequately exploit hard samples for multi-organ segmentation.

Purpose of the Study:

• To introduce a novel and effective online hard patch mining (OHPM) algorithm for 3D patch-based medical image segmentation.
• To address the limitations of existing sampling strategies in terms of efficiency and exploitation of hard samples.

Main Methods:

• Constructed an average shape model to guide the exploration of hard patches and aggregate feedback.
• Formalized hard mining as a multi-armed bandit problem, solved using bandit algorithms for online and sufficient mining.
• Integrated OHPM with advanced segmentation networks and evaluated on two diverse medical imaging datasets.

Main Results:

• OHPM demonstrated negligible time consumption due to the shape model, intuitively locating difficult anatomical areas.
• Bandit algorithms ensured online and sufficient hard sample mining, leading to improved model convergence.
• Comparative experiments showed OHPM significantly outperformed other sampling strategies, achieving nearly a 2% average Dice score improvement across datasets and networks.

Conclusions:

• OHPM is a superior sampling strategy for 3D multi-organ segmentation, enhancing both segmentation performance and model convergence.
• The proposed method offers an efficient and effective solution for online hard patch mining in medical image analysis.
• OHPM's ability to intuitively locate difficult regions and its computational efficiency make it a valuable tool for deep learning in medical imaging.