Democratizing Pathological Image Segmentation with Lay Annotators via Molecular-empowered Learning.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Bioimage analysis

Background:

• Accurate multi-class cell segmentation in whole slide images (WSI) is vital for clinical AI applications.
• Current AI model training relies on time-consuming and error-prone manual annotations by domain experts.
• Differentiating subtle cell types in WSI poses challenges for human annotators.

Purpose of the Study:

• To assess the feasibility of using lay annotators for pathological AI deployment.
• To develop a molecular-empowered learning scheme for cell segmentation using partial labels from non-experts.
• To democratize the development of pathological segmentation deep models.

Main Methods:

• Proposed a molecular-empowered learning scheme for multi-class cell segmentation.
• Integrated Giga-pixel molecular-morphology cross-modality registration and molecular-informed annotation.
• Employed a deep corrective learning method for noisy, partially annotated data.

Main Results:

• Achieved an F1 score of 0.8496 using molecular-informed annotations from lay annotators.
• Outperformed conventional morphology-based annotations from experienced pathologists (F1 = 0.7015).
• Demonstrated superior performance with 3 lay annotators compared to 2 experienced pathologists.

Conclusions:

• The molecular-empowered learning scheme effectively democratizes pathological AI development to lay annotator levels.
• This approach significantly improves cell segmentation accuracy and scalability.
• The method enables AI model training comparable to non-medical computer vision tasks.