Accelerating Volumetric Medical Image Annotation via Short-Long Memory SAM 2

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual annotation of volumetric medical images (MRI, CT) is time-consuming.
• Foundation models like Segment Anything Model 2 (SAM 2) offer potential for faster annotation via slice propagation.
• Current SAM 2 performance is limited by error propagation, especially at slice boundaries.

Purpose of the Study:

• To develop an improved segmentation model for medical images.
• To address the limitations of SAM 2 in propagating masks across volumetric data.
• To enhance the accuracy and efficiency of automated medical image annotation.

Main Methods:

• Proposed Short-Long Memory SAM 2 (SLM-SAM 2) architecture with distinct short-term and long-term memory banks and separate attention modules.
• Evaluated SLM-SAM 2 on four public datasets (organs, bones, muscles) across MRI, CT, and ultrasound videos.
• Compared SLM-SAM 2 performance against the default SAM 2.

Main Results:

• SLM-SAM 2 significantly outperformed default SAM 2, with average Dice Similarity Coefficient improvements of 0.14 (5 volumes) and 0.10 (1 volume).
• The proposed method demonstrated greater resistance to over-propagation errors.
• Reduced manual correction time by 60.575% per volume compared to SAM 2.

Conclusions:

• SLM-SAM 2 offers a substantial improvement over SAM 2 for medical image segmentation.
• The dual memory bank architecture effectively mitigates error propagation in volumetric annotation.
• SLM-SAM 2 represents a significant advancement towards accurate and efficient automated medical image annotation for segmentation model development.