I tried breast cancer AI tools: The tech has substance, but one hurdle remains
Community Discussion · Policy

I tried breast cancer AI tools: The tech has substance, but one hurdle remains

Bili GeBili GeAug 182026/08/18 271 views

Last week I met up with a team working on pathology AI, and their pitch really impressed me: they can map centrosome abnormalities onto entire tumor sections, showing defect differences across various regions. Centrosomes manage microtubule organization in cell biology; issues with them are linked to chromosomal instability and tumor heterogeneity. But previously, we could only count them one by one under high-magnification microscopy—no one was able to do a holistic analysis on whole tissue samples.

2 replies

?
Ctrl + Enter to reply
Factor Miner

Heavy reliance on annotation and the difficulty of standardizing cross-center data directly block model generalization. If the sample size achieving an AUC of 0.77 hasn't undergone stratified validation, the figure of a 14% miss rate for young patients can't be explained. We need to see specifically how many age groups they divided into.

Da Wei
Da WeiAug 18

Annotated data is indeed a bottleneck. I had the same feeling when using WorkBuddy for student essays before—if image quality is poor, no tool will work... But this feature that shows regional density differences has some merit.