A Heuristic Neural Network Structure Relying on Fuzzy Logic for Images Scoring

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence
• Fuzzy Logic Systems

Background:

• Traditional deep learning struggles with ambiguous features in medical image classification, leading to human error in semantic scoring.
• The Breast Imaging Reporting and Data System (BI-RADS) scoring involves nuanced categories with unclear boundaries, complicating accurate classification.
• Existing methods focus on kernel structures, neglecting the semantic ambiguity inherent in scoring tasks.

Purpose of the Study:

• To introduce a dominant fuzzy fully connected layer (FFCL) to enhance BI-RADS scoring accuracy.
• To develop a model that complements semantic scoring paradigms by incorporating human thought patterns into fuzzy rules.
• To reduce the impact of semantic conglutination and improve classification of ambiguous features.

Main Methods:

• Proposed a novel fuzzy fully connected layer (FFCL) integrated into a deep learning architecture.
• Developed a semantic-sensitive defuzzier layer to project features into semantic space.
• Implemented a fuzzy decoder to adjust output layer probabilities based on global trends and category relationships.
• Utilized Euclidean Distance for score refinement and two-sample t-tests for validation.

Main Results:

• The FFCL structure demonstrated superior performance in both triple and multiclass BI-RADS classification tasks.
• The proposed model effectively reduced the influence of semantic conglutination in scoring.
• Ambiguous semantic spaces between relative categories were observed to shrink during learning phases.
• The FFCL architecture outperformed state-of-the-art methods on the CBIS-DDSM dataset.

Conclusions:

• The FFCL is a universally applicable structure for improving classification of ambiguous features in medical imaging.
• This approach offers a significant advancement in accurate BI-RADS scoring, reducing potential human error.
• The FFCL represents a promising direction for deep learning models dealing with semantically ambiguous data.