Transport-Based Anatomical-Functional Metric Learning for Liver Tumor Recognition Using Dual-View Dynamic CEUS...

Area of Science:

• Medical Imaging
• Oncology
• Artificial Intelligence

Background:

• Primary liver cancer (PLC), including hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC), presents diagnostic challenges due to overlapping ultrasound (US) findings.
• Current qualitative US assessments are subjective and expertise-dependent, necessitating objective diagnostic tools.

Purpose of the Study:

• To develop and validate a novel transport-based anatomical-functional metric learning (T-AFML) method for objective quantification of multi-modal US findings in PLC.
• To improve the differential diagnosis accuracy between HCC and ICC using enhanced US image analysis.

Main Methods:

• Proposed a T-AFML method integrating gray-scale and dynamic contrast-enhanced US data.
• Introduced temporally regularized optimal transport for automatic alignment of hemodynamic enhancement patterns.
• Employed a selector-based metric integration mechanism to adaptively weigh US modalities.

Main Results:

• The T-AFML method achieved a diagnostic accuracy of 88.41% and a sensitivity of 86.16% in a cohort of 174 PLC patients (105 HCC, 69 ICC).
• Demonstrated superior performance in quantifying multi-modal US findings similarity compared to existing methods.

Conclusions:

• The T-AFML method offers an objective and effective approach for the differential diagnosis of PLC.
• This technique holds promise for improving early screening and diagnosis of primary liver cancers using ultrasound.