People talk about AI as if it were a subject. It isn’t, really. It’s a pile of methods that only become interesting once you point them at something that already has its own language, its own failure modes, and people who will notice if you get it wrong. Fei-Fei Li has been saying versions of this for years. I started taking it seriously when I had to.
We just published a review of deep learning for axillary lymph node assessment in breast cancer, in Virchows Archiv. That’s Springer’s pathology journal: old, specialist, Q1 in the category, and not particularly interested in whether your architecture is fashionable. If you work on this problem, this is one of the places the people who look at slides will see the paper.
The authors are Vedad Dedic, myself, Nejra Selak and Timur Ceric. The literature splits, a bit messily, into two tasks that get conflated in talks. Detection: the node is already out, the tissue is on the slide, and you are asking whether a metastasis is in it. That one is getting used. Prediction: you only have the primary tumour biopsy, and you are trying to guess whether the axilla would come back positive if you sampled it. That one is much less settled, which is not surprising. You are asking the image for something that isn’t in the image.
I did the software and ML side. We were inspired by Xu et al.: you can try to predict nodal status from the primary-tumour biopsy itself. I ran a version of that on our data, tuned it a bit, and landed in the same ballpark. I was not trying to beat the paper. I wanted to know whether the result survived a different set of slides. It did, more or less, which is the sort of unglamorous outcome you actually want before you write about a method.
I keep coming back to that when someone asks me to put a model on their operational data. The useful question is rarely whether we can train something. It’s whether a method that looked good in someone else’s setting still makes sense in yours - and whether you are willing to say so if it doesn’t.
Dedic V, Kadric M, Selak N, Ceric T. Deep learning in breast cancer histopathology: predicting and detecting axillary lymph node metastasis. Virchows Archiv, 4 September 2026. DOI