Openness-aware multi-prototype learning for open set medical diagnosis.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional image classification assumes closed-set data, failing when novel classes appear during testing.
• Open set recognition (OSR) addresses this by requiring models to identify both known and unknown classes.
• Existing prototype-based OSR methods using single prototypes overlook intra-class variance, increasing open space risk.

Purpose of the Study:

• To propose Openness-Aware Multi-Prototype Learning (OAMPL) to improve OSR performance.
• To address limitations of single-prototype modeling and enhance known-unknown discrimination.
• To introduce novel methods for handling intra-class variance and unknown sample representation in OSR.

Main Methods:

• Developed Adaptive Open Multi-Prototype Formulation (AOMF) for robust class modeling, reducing open space risk.
• Introduced Difficulty Aware Openness Simulator (DAOS) to dynamically synthesize challenging unknown samples.
• Jointly optimized AOMF and DAOS for enhanced known-unknown discrimination and effective learning.

Main Results:

• OAMPL maintained closed-set accuracy while improving OSR performance.
• Achieved approximately 1.5% and 1.2% improvement in AUROC and OSCR, respectively, compared to state-of-the-art models.
• Ablation studies confirmed the effectiveness of AOMF and DAOS components.

Conclusions:

• OAMPL effectively addresses limitations in existing OSR methods.
• The proposed AOMF and DAOS provide significant advancements in recognizing unknown classes.
• OAMPL demonstrates strong potential for OSR applications, particularly in nascent medical fields.