Images Versus Videos in Contrast-Enhanced Ultrasound for Computer-Aided Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Contrast-enhanced ultrasound (CEUS) is crucial for diagnosing focal liver lesions (FLLs).
• Traditional computer-aided diagnosis (CAD) systems struggle with dynamic lesion characterization due to reliance on static images.
• Integrating spatial and temporal information is key for advanced CEUS analysis.

Purpose of the Study:

• To evaluate the efficacy of Transformer-based architectures in improving CAD performance for liver pathology.
• To compare deep learning models for analyzing CEUS images and videos of FLLs.
• To assess the potential of attention mechanisms in identifying subtle diagnostic differences.

Main Methods:

• A Hybrid Transformer Neural Network (HTNN) combining Vision Transformer (ViT) and convolutional features was used for image-based classification.
• Spatio-temporal Convolutional Neural Network (CNN), CNN with Long Short-Term Memory (LSTM), and Video Vision Transformer (ViViT) were evaluated for video-based tasks.
• Performance was assessed based on classification accuracy with and without manual region of interest (ROI) selection.

Main Results:

• The HTNN achieved 97.77% accuracy for FLL classification but required manual ROI selection.
• Video-based models (CNN-LSTM, ViViT) reached 88% accuracy without needing ROI selection.
• Transformer models demonstrated high accuracy in CEUS-based liver diagnosis.

Conclusions:

• Transformer-based models show significant potential for accurate CEUS-based liver diagnosis.
• Attention mechanisms in Transformers can identify subtle inter-class differences, reducing the need for manual intervention.
• This study paves the way for more automated and accurate AI-driven liver pathology analysis.