Global and local multi-valued dissimilarity-based classification: application to computer-aided detection of...

Area of Science:

• Medical Imaging
• Computer-Aided Detection
• Machine Learning

Background:

• Precise localization of lesions is often impossible in computer-aided detection (CAD) systems using weakly labeled data.
• Existing CAD methods struggle with images where abnormalities are known but not precisely delineated.

Purpose of the Study:

• To present a novel CAD approach for effectively handling weakly labeled image data.
• To improve the accuracy of detecting diseases like tuberculosis (TB) in chest radiographs.

Main Methods:

• Utilized multi-valued dissimilarity measures to capture more information from local image features compared to single-valued measures.
• Extended the approach by integrating both local and global image analysis.
• Merged local and global classification results for a comprehensive image assessment.

Main Results:

• The global dissimilarity approach achieved an area under the ROC curve (AUC) of 0.81 for TB detection in chest radiographs.
• Combining local and global classification strategies enhanced the AUC to 0.83.
• This represents the first CAD system applied to a large dataset from a TB screening program.

Conclusions:

• The proposed multi-valued dissimilarity-based CAD framework effectively addresses challenges posed by weakly labeled data.
• Integrating local and global analysis significantly improves detection performance, demonstrating its potential for medical image analysis, particularly for tuberculosis screening.