A synthetic pathology world we control end to end, and the strongest generality signal the method produced.
Joeri’s thesis ends at this point, and I took up the work from here.
The TCGA-2k result was encouraging, but it couldn’t tell us how generally applicable our method was. I wanted to map out when adaptation works, and how reliably. It is difficult to take this up in a testing ground based on real-world data. You cannot obtain more data on demand, you cannot vary the shift or the task cleanly, and you constantly have to be careful to avoid data leakage issues. Therefore, I moved to a synthetic playground. There I could control everything: the data, the shift, the task, and the embedding model. I could also generate as much data as I wanted, with no leakage to worry about.
This section is about the data itself, the world I generate. The embedding model comes in the next section. The data is a synthetic histopathology dataset that I build from scratch. The images are cartoons inspired by real tissue: simple shapes standing in for cells, stroma, and background, arranged by rules I write. Because I generate the data, I set the class structure. I also set the shift, by changing the generation settings between the reference and deployment sets. That is a synthetic stand-in for a change of stain or scanner.
This data sits between two extremes. It is more realistic than the Gaussians of the first stage, which had no image content at all. On the other hand, it is far more controlled than real histopathology. I generate it in a structured, predictable way, so it has much less variability than real tissue. Real histopathology is a continuum, with no clean boundary between one tissue type and the next. Here the class structure is forced in strongly, so the clusters are much cleaner by construction. That is a feature for this experiment, but it limits what it can tell us. I further discuss this limitation at the end of this section.
The other piece of control is the embedding model. Instead of a pretrained foundation model, I train my own on the synthetic data, with self-supervised learning. I use two methods, SimCLR and I-JEPA, so the conclusions do not hang on a single pre-training recipe.
The real-data stages could not give me this much control. Now the embedding model, the shift, and the task are all things I set. If adaptation works here, I can say how consistently it works across these choices, instead of relying on a single number from a single combination.
Adaptation recovered the shift in every configuration I tried. It held across both embedding models and across several classification tasks defined on the synthetic images.
This was the strongest generality signal the method had produced up to this point. On real data I had one hopeful result. Here it worked across a whole grid of choices I controlled, and in some cases it worked spectacularly well. That raised my confidence that the approach was robust across a range of circumstances, and that it might transfer to harder settings.
This stage has clear limitations, and these limitations turn out to be important for what comes next.
The first limitation is the setting itself. Everything here is single-embedding, image-level classification on an input distribution I designed. One image gives one embedding, and the classifier acts on it directly. A real whole-slide problem is different in structure: a slide is thousands of tiles, each becomes an embedding, and a prediction comes from aggregating them, with the label attached to the whole slide rather than to any one tile. Section 4 takes that on.
The bigger limitation is the clean clusters. Separating and re-aligning the classes was easy here because I designed the classes to be separable. Real tissue is not built that way. The structure is not forced, the clusters are not clean, and, as I found out later, they do not always group themselves the way this experiment quietly assumes.
‹ Previous: Proof of concept · Up: FOMO-Shift · Next: Reality check ›