Original title: Pix2Rep-v2: Data-Efficient Representation Learning for Dense Medical Imaging Applications
S. Sifaoui, E. Angelini, S. Toupin, T. Pezel +1 · September 2026
Medical images usually need experts to label every pixel, which takes enormous time and effort. Researchers created Pix2Rep-v2 to learn from scans without human labels. The system learns tiny features for individual 2D pixels and 3D voxels, which are 3D pixels. It works by reducing repetitive information at the pixel level while maintaining equivarianceA property where the representation changes predictably when the input image is transformed, like rotated or shifted., meaning learned features properly track changes when an image shifts or transforms. This design lets it scale efficiently to 3D and wide-view scans. The researchers also introduced a way to test models using reference examples without any extra training. Pix2Rep-v2 needs far less labeled data than fully supervised setups and matches top methods, gaining +9.3 Dice points in one-shot segmentation on the M&Ms-2 dataset.
Imagine sorting a large pile of photos into categories without knowing what the categories are. Normally, you'd look for similarities and differences, but if many photos are similar, you might waste time on duplicates. Here, the system learns to pick out the unique details in medical images by automatically reducing the emphasis on repetitive elements, so it can work efficiently with few examples.
This could help doctors get accurate diagnostic tools faster, even when labeled medical images are scarce, which is common in rare diseases or new technologies.
The work does not establish performance on all medical imaging tasks or modalities beyond those tested, as it was evaluated on four datasets. It also does not compare with every possible baseline method or guarantee generalizability to real-world clinical settings.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Pix2Rep-v2: Data-Efficient Representation Learning for Dense Medical Imaging Applications — S. Sifaoui, E. Angelini, S. Toupin et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.