ULS4US: universal lesion segmentation framework for 2D ultrasound images

Area of Science:

• Medical imaging and artificial intelligence
• Deep learning for medical image analysis
• Ultrasound image segmentation

Background:

• Current deep learning (DL) methods for ultrasound (US) image segmentation are organ-specific, lacking a general-purpose framework for computer-aided diagnosis.
• Existing DL models are limited to specific organs like the breast or thyroid, hindering broader clinical application.
• A universal framework is needed to segment lesions across various organs in US images.

Purpose of the Study:

• To develop a universal lesion segmentation framework for ultrasound images (ULS4US) applicable across different organs.
• To address the need for a general-purpose tool in computer-aided diagnosis scenarios.
• To leverage distinct ultrasonic echo intensities and patterns of lesions for segmentation.

Main Methods:

• Proposed a ULS4US framework incorporating a multiple-in multi-out (MIMO) UNet for multiscale feature extraction.
• Implemented a two-stage lesion-aware learning algorithm for recursive lesion localization and segmentation.
• Introduced a lesion-adaptive loss function with weighted and self-supervised components for network outputs.

Main Results:

• ULS4US achieved superior performance compared to six state-of-the-art models on a unified dataset (Accuracy: 0.956, DSC: 0.836, HD: 7.849, mIoU: 0.731).
• The unified dataset comprised over 2200 images from three organ types across public and private sources.
• Comparative experiments indicated ULS4US's scalability with additional data.

Conclusions:

• The study demonstrates the potential of DL-based universal lesion segmentation in clinical ultrasound.
• ULS4US can significantly reduce clinician workload.
• The framework has the potential to enhance diagnostic accuracy in medical imaging.