Learning-based ventricle detection from cardiac MR and CT images

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Automatic detection of regions of interest (ROIs) in medical images is challenging.
• Existing methods often struggle with accuracy and require manual intervention.

Purpose of the Study:

• To develop and evaluate a novel two-stage learning algorithm for automatic ROI detection in medical images.
• To improve the accuracy of segmenting anatomical structures like cardiac ventricles.

Main Methods:

• A preliminary threshold is determined using global histogram analysis based on a proposed bell image intensity model.
• A two-stage learning process is employed: threshold value selection and ROI selection from an attention map with dynamic threshold tuning.
• The algorithm was tested on gradient-echo magnetic resonance (MR) images for endocardium boundary detection and cardiac computed tomography (CT) images.

Main Results:

• The method generates an attention map that approximates ROIs but may contain spurious regions.
• The two-stage learning approach refines the initial segmentation to better identify target ROIs.
• Experimental results on MR and CT images demonstrated the potential for locating endocardium boundaries, though further fine-tuning is suggested.

Conclusions:

• The proposed algorithm offers a promising approach for automated ROI detection in medical imaging.
• The two-stage learning strategy enhances the accuracy of segmentation by refining initial thresholding.
• The method provides a foundation for further development and application-specific fine-tuning for precise boundary detection.