Liver disease classification from ultrasound using multi-scale CNN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Ultrasound (US) is a primary noninvasive tool for fatty liver disease diagnosis.
• Traditional B-mode US assessments are subjective, necessitating objective diagnostic aids.
• Computer-aided diagnostic tools enhance US specificity and sensitivity for uniform diagnoses.

Purpose of the Study:

• To develop a novel deep learning model for nonalcoholic fatty liver disease classification using ultrasound data.
• To enhance the diagnostic accuracy of ultrasound for fatty liver disease through advanced computational methods.
• To improve the objectivity and consistency of fatty liver disease diagnosis in clinical practice.

Main Methods:

• A multi-feature guided multi-scale residual convolutional neural network (CNN) was designed.
• B-mode US images were combined with local phase filtered and radial symmetry transformed images as multi-feature inputs.
• The model was evaluated on in vivo liver US images from 55 subjects, comparing against traditional CNNs and machine learning methods.

Main Results:

• The proposed multi-feature CNN model achieved an average classification accuracy exceeding 90% across tenfold cross-validation.
• The model demonstrated a 97.8% area under the ROC curve (AUC) in patient-specific leave-one-out cross-validation.
• Significant improvements in classification accuracy were observed compared to mono-feature CNN architectures.

Conclusions:

• Combining multiple features and utilizing multi-scale CNN architectures significantly improves liver classification accuracy.
• The developed deep learning approach shows potential for practical application in assisting radiologists with nonalcoholic fatty liver disease diagnosis.
• This method offers a more objective and accurate approach to diagnosing fatty liver disease compared to traditional subjective assessments.