Understanding Stain Separation Improves Cross-Scanner Adenocarcinoma Segmentation with Joint Multi-Task Learning.

Area of Science:

• Digital pathology
• Computational pathology
• Medical image analysis

Background:

• Image variability, or domain shift, in digital pathology due to tissue preparation and scanner differences causes color inconsistencies.
• These inconsistencies negatively impact the performance of automated tumor segmentation algorithms.

Purpose of the Study:

• To develop a robust method for tumor segmentation in digital pathology that overcomes image variability.
• To improve the generalization of segmentation models across different scanners and tissue preparation methods.

Main Methods:

• A joint multitask U-net architecture was developed for simultaneous segmentation and stain separation.
• The model isolates stain matrix and stain density to normalize color variations.
• Training and validation were performed on histopathological images from multiple scanners.

Main Results:

• The proposed model achieved a Dice score of 0.898 and an Intersection over Union (IoU) score of 0.816 on a dataset from three scanners, outperforming conventional methods.
• On external datasets from six different scanners, the model maintained strong performance with an average Dice score of 0.792 and IoU of 0.792.
• The stain separation technique effectively handled color variations, enhancing segmentation accuracy.

Conclusions:

• The developed joint multitask U-net with stain separation significantly improves tumor segmentation accuracy and generalization in digital pathology.
• This approach addresses the challenge of domain shift caused by scanner heterogeneity.
• The findings support the development of more reliable and consistent diagnostic tools for diseases like breast adenocarcinoma.