Leverage Weakly Annotation to Pixel-wise Annotation via Zero-shot Segment Anything Model for Molecular-empowered...

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Accurate cell identification in whole slide imaging (WSI) is crucial for clinical applications.
• Current AI model training requires laborious pixel-level annotations by experts, which are prone to errors and scalability issues.
• Molecular-empowered learning using immunofluorescence (IF) images shows promise but still necessitates manual delineation.

Purpose of the Study:

• To develop a more efficient and scalable annotation method for pathological image segmentation.
• To explore the use of Segment Anything Model (SAM) with weak box annotations for zero-shot learning.
• To assess the performance of SAM-assisted molecular-empowered learning (SAM-L) compared to traditional methods.

Main Methods:

• Utilized SAM to generate pixel-level annotations from weak box annotations.
• Trained a segmentation model using SAM-generated labels.
• Employed a molecular-empowered learning approach with lay annotators and IF images.
• Evaluated annotation accuracy and segmentation model performance.

Main Results:

• SAM-L significantly reduces annotation effort for lay annotators by using only box annotations.
• Annotation accuracy and deep learning-based segmentation performance are maintained.
• The proposed method enables non-expert annotators to contribute effectively to pathological image segmentation.

Conclusions:

• SAM-assisted molecular-empowered learning (SAM-L) offers a scalable and accurate solution for pathological image annotation.
• This approach democratizes the training of AI models for pathological image segmentation.
• It minimizes reliance on expert annotators and intensive manual delineation.