What are your thoughts on automatic segmentation of metastases using Deep Learning Models?

Question

[clara-28]

Hello everyone,

I am currently exploring the use of deep learning models for automatic segmentation of metastases in histopathological images. While tools like Mesmer, Cellpose, or custom UNet models seem promising, I’ve noticed that many pathologists still rely on manual segmentation.

Given the potential of automation to save time and improve consistency, I’m curious:

• What are your experiences with using deep learning for metastasis segmentation?
• Do you believe these tools can match (or even surpass) the accuracy of manual segmentation?
• Are there specific challenges (e.g., tumor heterogeneity, data quality, or interpretability) that make automation less appealing or effective?
• Do you know of any pre-trained models specifically designed for metastasis segmentation, or have you worked on such tasks yourself? I’d love to hear your thoughts on whether deep learning is ready to replace manual segmentation in this context, or if it’s more of a complementary tool for now.

Thank you for sharing your insights!

[ATpoint]

I am only doing minor work on image analysis, so my comments are general:

What are your experiences with using deep learning for metastasis segmentation?

Any approach that can suggest an analysis is potentially valuable. The user can then still intervene and change things, but like drawing 500 circles around cells rather than using StarDist and maybe changing manually two misclassified ones is definitely a plus.

Do you believe these tools can match (or even surpass) the accuracy of manual segmentation?

They don't need to. They should complement expert-curation.

Are there specific challenges (e.g., tumor heterogeneity, data quality, or interpretability) that make automation less appealing or effective?

I am no pathologist, but in imaging and histology (or immunefluorescence, of science in general) data quality is always an issue. Shitin-shit-out, simple as that. Poor images are hard to analyze, signal and noise will be hard to distinguish.