Texture features for classification of ultrasonic liver images

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Engineering

Background:

• Accurate classification of ultrasonic liver images is crucial for diagnosing liver diseases like hepatoma and cirrhosis.
• Traditional texture analysis methods using spatial gray-level dependence matrices, Fourier power spectrum, gray-level difference statistics, and Laws texture energy measures show limitations in speed and accuracy.

Purpose of the Study:

• To evaluate the performance of existing texture features for classifying ultrasonic liver images.
• To propose and validate a novel texture feature set, multiresolution fractal features, for rapid and accurate detection of diffuse liver diseases.

Main Methods:

• Utilized spatial gray-level dependence matrices, Fourier power spectrum, gray-level difference statistics, and Laws texture energy measures for initial classification.
• Developed multiresolution fractal features based on multiple resolution imagery and the fractional Brownian motion model.
• Employed Bayes classifier and Hotelling trace criterion for performance evaluation.
• Implemented a real-time algorithm for classification.

Main Results:

• Initial texture features demonstrated insufficient speed and accuracy for clinical application.
• The proposed multiresolution fractal features effectively capture texture information for liver disease classification.
• A real-time implementation achieved approximately 90% correct classification rates for normal liver, hepatoma, and cirrhosis images.

Conclusions:

• Multiresolution fractal features offer a promising approach for rapid and accurate diagnosis of diffuse liver diseases from ultrasonic images.
• This novel feature set outperforms traditional methods in terms of both speed and classification accuracy.
• The findings support the potential of advanced texture analysis techniques in improving liver disease detection.