A framework for medical image retrieval using machine learning and statistical similarity matching techniques with
Md Mahmudur Rahman1, Prabir Bhattacharya, Bipin C Desai
1Department of Computer Science Software Engineering, Concordia University, Montreal, QC H3G 1M8, Canada. mah_rahm@cse.concordia.ca
Summary
This study introduces a content-based image retrieval framework using machine learning for diverse medical images. The system effectively categorizes and retrieves images, improving search efficiency and accuracy.
Area of Science:
- Medical Imaging Informatics
- Computer Vision
- Machine Learning
Background:
- Content-based image retrieval (CBIR) is crucial for managing large medical image databases.
- Existing CBIR systems often struggle with the diversity of medical imaging modalities, anatomical regions, and orientations.
- Bridging the semantic gap between low-level image features and high-level semantic categories remains a challenge.
Purpose of the Study:
- To propose a novel CBIR framework for diverse medical image collections.
- To enhance retrieval efficiency and accuracy through category-specific searching and machine learning techniques.
- To reduce the semantic gap by associating low-level image features with high-level semantic categories.
Main Methods:
- Implementation of a CBIR framework incorporating machine learning for prefiltering, statistical distance measures for similarity matching, and a relevance feedback (RF) scheme.
- Utilizing supervised and unsupervised learning techniques, including probabilistic multiclass support vector machine (SVM) and fuzzy c-mean (FCM) clustering, for image categorization and search space reduction.
- Developing category-specific statistical similarity matching and an RF mechanism for dynamic query parameter updates.
Main Results:
- Experimental validation on a ground-truth database of 5000 diverse medical images across 20 categories.
- Demonstrated improvement in retrieval effectiveness and efficiency through cross-validation (CV) accuracy and precision-recall analysis.
- The proposed framework effectively narrows the semantic gap and enhances image retrieval performance.
Conclusions:
- The developed CBIR framework offers an effective solution for retrieving diverse medical images.
- The integration of machine learning, category-specific matching, and relevance feedback significantly improves retrieval accuracy and efficiency.
- This approach provides a robust and adaptable solution for medical image database management and retrieval.