Real-Time Deep Recognition of Standardized Liver Ultrasound Scan Locations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Liver ultrasound (US) is crucial for diagnosing liver diseases.
• Accurate identification of liver segments in US images is challenging due to patient variability and image complexity.

Purpose of the Study:

• To develop an automatic, real-time system for recognizing standardized liver US scans.
• To guide clinicians by coordinating US scans with reference liver segments.

Main Methods:

• A novel deep hierarchical architecture was proposed for classifying liver US images into 11 standardized scans.
• Hierarchical classification with distinct features per level and feature space proximity analysis for ambiguous images were employed.
• Experiments utilized hospital-acquired US image datasets, with distinct patient groups for training and testing to assess performance under variability.

Main Results:

• The proposed method achieved an F1-score exceeding 93%, demonstrating high sufficiency for a guiding tool.
• The hierarchical architecture outperformed non-hierarchical approaches in accuracy.
• The system effectively handles patient variability, a key challenge in liver US interpretation.

Conclusions:

• The developed deep hierarchical architecture provides a robust solution for automatic liver US scan classification.
• This AI tool can significantly assist clinicians in accurate liver segment identification and diagnosis.
• The approach addresses the long-standing challenge of variability and complexity in liver US imaging.