Segment anything small for ultrasound: Enhancing segmentation with non-generative augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Ultrasound (US) image segmentation is crucial but challenging for small anatomical structures due to noise and imaging variability.
• Existing deep learning models struggle with accuracy and robustness in segmenting small US structures.

Purpose of the Study:

• To introduce Segment Anything Small (SAS), a novel data augmentation technique to enhance deep learning model performance for segmenting small anatomical structures in ultrasound images.
• To improve the robustness and generalizability of segmentation models across diverse imaging conditions and anatomical variations.

Main Methods:

• SAS employs a dual transformation strategy: simulating scale variations by resizing organ thumbnails and injecting noise to mimic tissue texture variations.
• A promptable foundation model was fine-tuned using SAS on a controlled medical imaging dataset.
• The model's performance was evaluated on internal and external datasets using Dice scores and iterative point prompts.

Main Results:

• SAS significantly improved segmentation performance, with average Dice score gains of 0.16 (95% CI: 0.132, 0.188) and up to 0.35.
• Iterative point prompts achieved performance comparable to bounding box prompts with minimal input.
• The technique enhanced model robustness and generalizability, particularly for small structures, without negatively impacting larger ones.

Conclusions:

• SAS is an effective and computationally efficient data augmentation technique for improving ultrasound image segmentation, especially for small structures.
• The method reduces the need for extensive human labeling, making it valuable for resource-constrained settings.
• SAS offers a practical solution for enhancing medical image analysis through improved deep learning model performance.