Leveraging weak supervision for cell localization in digital pathology using multitask learning and consistency loss.

Area of Science:

• Digital pathology
• Computational pathology
• Medical image analysis

Background:

• Cell detection and segmentation are crucial for automated digital pathology systems.
• Encoder-decoder networks are effective but require extensive boundary annotations.
• Limited or weaker annotations, like cell counts, are often sufficient for specific tasks.

Purpose of the Study:

• To propose a novel mixed-supervision strategy for training multitask networks in digital pathology.
• To leverage pathologist 'eyeballing' (visual estimation) cell counts as an auxiliary supervisory signal.
• To develop a multitask network for concurrent cell counting and localization, enhanced by a consistency loss.

Main Methods:

• Implemented a mixed-supervision approach incorporating eyeballing-derived cell counts.
• Designed a multitask network for simultaneous cell counting and localization.
• Introduced a consistency loss to regularize training by minimizing prediction discrepancies between tasks.

Main Results:

• Demonstrated effective utilization of weak annotations (eyeballing counts) on two hematoxylin-eosin stained tissue image datasets.
• Achieved improved performance in cell counting and localization tasks, especially when strong annotations were scarce.
• Validated the feasibility of integrating eyeballing-derived ground truths into network training.

Conclusions:

• The proposed mixed-supervision strategy significantly reduces the reliance on labor-intensive, precise cell boundary annotations.
• Integrating eyeballing cell counts offers a practical and efficient way to train digital pathology models.
• This approach holds promise for advancing automated analysis in digital pathology by easing annotation requirements.