Multiple-Instance Learning for Medical Image and Video Analysis.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Multiple-instance learning (MIL) is a machine learning paradigm well-suited for medical image and video analysis (MIVA).
• MIL algorithms learn from globally labeled data, eliminating the need for manual segmentation required by traditional single-instance learning (SIL) methods.
• The MIVA community shows increasing interest in MIL due to its convenience and potential for improved accuracy.

Purpose of the Study:

• To review existing strategies for modeling MIVA tasks using MIL.
• To recommend general-purpose and MIVA-specific MIL algorithms for various MIVA tasks.
• To discuss recent trends and propose future directions in MIL for MIVA.

Main Methods:

• A meta-analysis of 73 research papers published in the MIVA literature since 1997.
• Compilation and discussion of experimental results from various medical image and video datasets.
• Review of MIL strategies for MIVA problem modeling.

Main Results:

• MIL algorithms are more convenient than SIL solutions for MIVA tasks.
• MIL algorithms demonstrate higher accuracy in many MIVA applications.
• MIL is identified as an ideal solution for numerous MIVA tasks.

Conclusions:

• Multiple-instance learning (MIL) presents a powerful and efficient paradigm for medical image and video analysis (MIVA).
• The absence of manual segmentation requirements makes MIL a practical alternative to SIL.
• MIL's superior accuracy and convenience position it as a key technology for future MIVA research and applications.