Integrating a learned probabilistic model with energy functional for ultrasound image segmentation

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Image Segmentation

Background:

• Ultrasound (US) image segmentation is crucial for computer-aided diagnosis (CAD) systems.
• US images often suffer from poor quality, including noise, ambiguous boundaries, and heterogeneity, complicating lesion segmentation.
• Lesion regions can lack salience against normal tissues, posing a significant segmentation challenge.

Purpose of the Study:

• To develop and evaluate an advanced US image segmentation algorithm for accurate lesion separation from background.
• To enhance segmentation accuracy by addressing common challenges in US image quality.

Main Methods:

• A learned probabilistic model based on the generalized linear model (GLM) was employed to reduce false positives and enhance the lesion region's likelihood energy term.
• An energy functional incorporating boundary indicators and probability statistics was utilized to define reliable lesion boundaries.
• Probabilistic information was integrated into the energy functional framework to mitigate the impact of poor image quality.

Main Results:

• The proposed algorithm achieved high performance metrics: DICE coefficient of 0.96, Jaccard distance of 0.91, root-mean-square error of 0.059, and mean absolute error of 0.042.
• Analysis confirmed significant performance improvements, including robustness to initialization and noise.
• The integration of probabilistic models effectively improved segmentation accuracy in challenging US images.

Conclusions:

• The proposed algorithm effectively segments lesions in ultrasound images by combining learned probabilistic models with energy functionals.
• This approach significantly overcomes the limitations posed by poor image quality, noise, and ambiguous boundaries.
• The method demonstrates a substantial improvement in segmentation accuracy, offering a valuable tool for computer-aided diagnosis.