On Combining Multiple-Instance Learning and Active Learning for Computer-Aided Detection of Tuberculosis.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Detection

Background:

• Multiple-instance learning (MIL) enables training computer-aided detection (CAD) systems with case-level labels, comparable to supervised methods for tasks like tuberculosis (TB) detection.
• However, inherent MIL uncertainty limits its use for detailed analysis (e.g., pixel-level) or localized lesion detection.

Purpose of the Study:

• To reduce uncertainty in MIL-based CAD systems by embedding a multiple-instance learning classifier within an active learning (AL) framework.
• To develop a novel instance selection mechanism for efficient expert labeling, minimizing labeling effort.

Main Methods:

• Integrated a MIL classifier within an active learning (AL) framework for CAD systems.
• Developed a novel instance selection mechanism using one-class classification, adapted to select meaningful regions for expert labeling.
• Performed a single iteration of active learning, contrasting with typical multi-iteration AL methods.

Main Results:

• The proposed AL framework significantly improved the performance of the MIL-based CAD system.
• Quantitative and qualitative pixel-level evaluations demonstrated the effectiveness of the approach in detecting TB-related textural abnormalities.
• The method successfully reduced uncertainty, enhancing the applicability of MIL for detailed image analysis.

Conclusions:

• Embedding MIL within a single-iteration AL framework with a novel region-selection mechanism effectively reduces uncertainty.
• This approach enhances the performance and applicability of MIL-based CAD systems for precise lesion detection and analysis.
• The method offers a more efficient and effective solution for tasks requiring detailed, pixel-level insights in medical imaging.