Cascaded Inner-Outer Clip Retformer for Ultrasound Video Object Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Computer-aided ultrasound (US) imaging aids early clinical diagnosis.
• Current video object segmentation (VOS) models for US images suffer from computational redundancy and accuracy degradation due to blurry tumor areas and uniform attention mechanisms.
• Existing methods struggle with the harsh image quality and indistinct tumor boundaries in ultrasound videos.

Purpose of the Study:

• To develop a lightweight, clip-level VOS framework for enhanced breast lesion segmentation in ultrasound videos.
• To improve segmentation accuracy and maintain processing speed for computer-aided diagnosis.
• To address the limitations of current VOS models in handling challenging ultrasound image quality.

Main Methods:

• A new annotated benchmark dataset for breast lesion segmentation in ultrasound videos was created.
• Proposed the Inner-Outer Clip Retformer for parallel spatial-temporal tumor feature extraction.
• Introduced a Clip Contrastive loss function and Global Retentive Memory for improved feature alignment and reduced computational load.

Main Results:

• The proposed framework achieves higher segmentation accuracy compared to state-of-the-art methods.
• The model demonstrates improved spatial-temporal perception abilities without significant parameter increase.
• The Inner-Outer Clip Retformer effectively extracts both movement and detailed current tumor features for accurate segmentation.

Conclusions:

• The developed lightweight clip-level VOS framework offers a significant advancement in breast lesion segmentation for ultrasound imaging.
• The proposed methods, including the Inner-Outer Clip Retformer and Clip Contrastive loss, enhance segmentation accuracy and efficiency.
• This work provides a promising tool for improving early clinical diagnosis and treatment through more precise tumor identification in ultrasound videos.