Visual saliency-based active learning for prostate magnetic resonance imaging segmentation.

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Prostate segmentation from MRI is crucial for diagnosis and treatment planning.
• Traditional methods often require extensive manual annotation, which is time-consuming and labor-intensive.
• Developing efficient and accurate automated segmentation techniques is a significant challenge.

Purpose of the Study:

• To propose an active learning (AL) approach for automated prostate segmentation in magnetic resonance images (MRIs).
• To enhance the efficiency of AL by incorporating visual saliency principles for label query selection.
• To improve segmentation accuracy and reduce computational cost compared to fully supervised methods.

Main Methods:

• An active learning strategy inspired by visual saliency principles was developed.
• A graph representation encoding classification maps and low-level features was used.
• Random walks identified informative nodes (label query samples) within a defined volume of interest (VOI).
• Semi-supervised random forest classifiers generated probability maps, and graph cuts optimized a Markov random field for segmentation.

Main Results:

• The proposed active learning method demonstrated superior performance in prostate segmentation.
• The approach achieved better results compared to conventional fully supervised learning methods.
• The use of a VOI and saliency-based query strategy reduced computational time.

Conclusions:

• The active learning approach integrating visual saliency is effective for prostate MRI segmentation.
• This method offers a more efficient and accurate alternative to fully supervised learning.
• The strategy holds promise for improving automated medical image segmentation workflows.