Abstract
This paper presents a multimode retrieval system for breast imaging that utilizes feature selection and association rules to bridge the "semantic gap" between low-level image data and clinical interpretation. By digging out the associated relationship between image low-level features and semantic features, the system employs an association classification algorithm to derive visual semantic features. This approach reduces the gap between mathematical features and clinical meaning, enabling more accurate similarity measures. Experimental results show that this method improves the performance of breast imaging case retrieval and provides meaningful decision support for doctors.
Introduction
In medical diagnostics, finding similar historical cases is a vital tool for accurate assessment. However, traditional retrieval methods often fall into two categories: text-based retrieval, which is subjective and labor-intensive, and Content-Based Image Retrieval (CBIR), which relies on mathematical "low-level" features (e.g., pixel intensity, edge histograms). These low-level features often fail to capture the "high-level" semantic concepts clinicians use, such as whether a mass is "lobulated" or "irregular." This disconnect is known as the semantic gap. This paper proposes a multimode retrieval system that automatically translates mathematical features into visual semantic features using data mining techniques.
Methodology
The authors developed a multi-stage pipeline to automate the classification and retrieval process, integrating low-level image processing with high-level semantic mapping.
Feature Selection and Rule Mining
Starting with 32 low-level features, the system uses the StARMiner algorithm to identify the 8 most critical features for distinguishing mass shapes, such as radius curvature and perimeter-area ratios. The core of the system involves finding association rules using the Apriori algorithm that link these specific low-level feature values to semantic categories (e.g., "If Feature X is in Range Y, there is high confidence the mass is Irregular").
Associative Classification and Retrieval
Using the Associative Classification Engine (ACE), the system builds a model to categorize masses into three shapes: Oval, Irregular, and Lobulated. As shown in the system architecture (Figure 1), when a new query image is input, the system determines its semantic shape and then performs a similarity search only within that semantic category, combining mathematical similarity with clinical meaning.
Experimental Results
The study utilized a dataset of 153 breast imaging cases, with 106 used for training and 47 for testing. The associative classification model achieved an overall precision of 85.11% in correctly identifying mass shapes.
Retrieval Performance
The "Multimode Retrieval" approach significantly outperformed standard content-based retrieval. By narrowing the search space to semantically similar cases first, the system maintained higher precision as recall increased.
As illustrated in Figure 2, the multimode retrieval method (especially with dimensionality reduction) provides superior precision compared to traditional CBIR. This ensures that the retrieved cases are not just mathematically similar but also clinically relevant to the user's query.
Conclusion
This research demonstrates that association rules can effectively bridge the semantic gap in medical image retrieval. By focusing on the most relevant features and using a hierarchical classification model, the system provides more accurate and meaningful results for clinical decision support. Future work will aim to expand the semantic categories to include mass margins and optimize the mining algorithms using larger datasets.
Summary updated