An Edge-Based Selection Method for Improving Regions-of-Interest Localizations Obtained Using Multiple Deep Learning...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Computer-aided diagnosis (CAD) systems enhance breast cancer diagnosis from breast ultrasound (BUS) images.
• Accurate localization of the tumor region-of-interest (ROI) is crucial for CAD system performance.
• Existing deep learning object-detection models generate ROIs, but their quality varies.

Purpose of the Study:

• To propose a novel edge-based selection method for selecting the optimal ROI from multiple deep learning models.
• To improve the localization accuracy of tumor regions in BUS images for better cancer diagnosis.
• To evaluate the effectiveness of the proposed method against existing object-detection and combining techniques.

Main Methods:

• Utilized the Dense Extreme Inception Network (DexiNed) for computing edge maps of BUS images.
• Developed an edge-based selection method to analyze and select the best ROI from various deep learning object-detection models.
• Evaluated the method on a private dataset (380 images) for cross-validation and a public dataset (630 images) for generalization.

Main Results:

• The edge-based selection method achieved an overall ROI detection rate of 98%, with mean precision, recall, and F1-score of 0.91, 0.90, and 0.90, respectively.
• The proposed method significantly outperformed four individual deep learning object-detection models.
• Outperformed three baseline methods designed for combining ROIs from multiple models.

Conclusions:

• The novel edge-based selection method effectively improves tumor region localization in BUS images.
• This method enhances the performance of computer-aided diagnosis systems for breast cancer.
• Demonstrated the potential of deep learning edge detection for selecting optimal ROIs in medical imaging analysis.