Abductive multi-instance multi-label learning for periodontal disease classification with prior domain knowledge.

Area of Science:

• Dentistry
• Machine Learning
• Medical Image Analysis

Background:

• Machine learning aids dental disease diagnosis, but current methods struggle to pinpoint specific affected regions.
• Existing two-stage models for periodontal disease diagnosis lack the precision required for expert-level analysis.

Purpose of the Study:

• To propose a novel ABductive Multi-Instance Multi-Label learning (AB-MIML) approach for enhanced periodontal disease diagnosis.
• To improve the modeling of relationships between dental images, local patches, and diagnostic labels.
• To integrate domain expertise and image structure for more accurate diagnostic reasoning.

Main Methods:

• Developed an AB-MIML framework treating intraoral images as 'bags' and patches as 'instances'.
• Enhanced multi-instance multi-label learning to establish many-to-many correspondences.
• Incorporated a knowledge base of expert knowledge and image structure for abductive reasoning.

Main Results:

• AB-MIML demonstrated superior performance in diagnosing periodontal diseases compared to state-of-the-art methods.
• The approach accurately identified critical diagnostic regions, mirroring human expert assessments.
• Experimental results confirmed the method's effectiveness across various performance metrics.

Conclusions:

• The proposed AB-MIML method significantly advances dental disease diagnosis accuracy.
• Integrating domain knowledge via abductive reasoning enhances machine learning model performance.
• AB-MIML effectively bridges the gap in pinpointing specific affected regions in dental imaging.