A hierarchical neural network algorithm for robust and automatic windowing of MR images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate display window adjustment is crucial for interpreting magnetic resonance (MR) images.
• Existing methods may lack adaptability and robustness for diverse MR image datasets.

Purpose of the Study:

• To develop a novel hierarchical neural network algorithm for automatic adjustment of display window width and center in MR images.
• To create an adaptive and extendable algorithm capable of handling a wide range of MR images.

Main Methods:

• A hierarchical neural network architecture combining feature generation (wavelet histogram, spatial statistics), competitive layer clustering, radial basis function (RBF) networks, bi-modal linear estimators, and data fusion.
• Utilizing both RBF and bi-modal linear estimators for subclass-specific parameter estimation.
• Employing a data fusion process to combine estimates for final display parameter computation.

Main Results:

• The algorithm demonstrated satisfactory performance across a wide range of MR images.
• The RBF estimator excelled with familiar image types, while the bi-modal estimator handled diverse images effectively.
• Data fusion ensured accurate estimations for trained images and robustness for unknown images.

Conclusions:

• The proposed hierarchical neural network algorithm provides an effective and adaptive solution for automatic MR image display parameter adjustment.
• The algorithm's adaptive and extendable nature allows it to handle new MR image types through training.
• This approach enhances both the accuracy and robustness of MR image visualization.