Model-data-driven adversarial active learning for brain tumor segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Active learning (AL) aims to reduce labeling effort in machine learning by selecting informative samples.
• Existing AL methods for segmentation often adapt classification techniques and overlook medical imaging specifics like data scarcity and class imbalance.
• Medical image segmentation faces challenges including high class imbalance, domain differences, and limited annotated data.

Purpose of the Study:

• To develop a novel active learning (AL) framework specifically designed for medical image segmentation.
• To address limitations of current AL methods in handling medical image characteristics.
• To improve the efficiency and effectiveness of medical image segmentation by reducing the need for extensive manual annotation.

Main Methods:

• Introduced a pseudo-label-based filter to handle excessive blank patches in medical abnormality segmentation (e.g., lesions, tumors).
• Proposed a novel query strategy combining model impact and data stability using adversarial attacks for sample selection.
• Utilized adversarial samples generated during the query process to enhance model robustness.

Main Results:

• The proposed AL framework demonstrated effectiveness compared to state-of-the-art methods in medical image segmentation.
• Achieved competitive results with less than 14% annotated patches for 3D brain MRI multiple sclerosis (MS) segmentation.
• Achieved competitive results with 20% annotated patches for Low-Grade Glioma (LGG) tumor segmentation, comparable to full supervision.

Conclusions:

• The novel AL framework significantly reduces the required annotation effort for medical image segmentation tasks.
• The method enhances model performance and robustness while alleviating the time burden on expert annotators.
• This approach facilitates advancements in medical image segmentation by improving efficiency and accessibility.