LUS-DET: Liver Ultrasound Open-Vocabulary Object Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ultrasound Technology

Background:

• Liver ultrasound lesion assessment relies on standard scanning planes and interpretation of anatomical structures.
• Current AI approaches often separate standard plane recognition and lesion detection, neglecting their clinical interdependence.

Purpose of the Study:

• To develop a unified AI framework, LUS-DET, that bridges liver ultrasound standard plane analysis (LUSP) and liver ultrasound disease diagnosis (LUDD).
• To enable open-vocabulary object detection for lesion analysis by integrating text-guided modeling.

Main Methods:

• Curated a retrospective LUSP dataset with 44,669 region-caption pairs across 12 anatomical categories for in-domain pretraining.
• Developed a region-text alignment mechanism and introduced object prompts for zero-shot lesion detection.
• Evaluated LUS-DET on an open-source LUDD task without lesion-specific annotations.

Main Results:

• LUS-DET achieved competitive zero-shot performance in lesion detection.
• The framework demonstrated superior accuracy and robustness during end-to-end fine-tuning compared to conventional methods.
• Successfully enabled lesion detection without requiring lesion-specific annotations.

Conclusions:

• LUS-DET represents the first clinically coherent model unifying standard plane localization and lesion analysis in liver ultrasound.
• This approach offers a new direction for developing structure-aware and workflow-aligned AI systems in medical imaging.
• Highlights the potential of text-guided modeling to bridge distinct tasks in medical AI.