What if a model could segment medical images without retraining?

Tyche is a new model that learns to segment medical images without having to retrain a new algorithm each time. Instead of creating a single mask, it produces possible solutions, like an artist drawing various ways to outline an organ. The secret is the 'context': a set of already seen examples that guides the model in tasks never faced before. A new convolutional block architecture allows different results to 'talk' to each other, improving the final quality. During testing, Tyche uses a trick called 'in-context augmentation', which varies the image and combines the results, generating predictions. This means doctors and researchers can obtain plausible options for the same scan, better understanding areas of uncertainty. The advantages are clear: fewer resources, less time, and more flexibility. No supercomputer is needed to retrain the model each time a new pathology appears. Tyche opens the way to more adaptable artificial intelligence, which learns from few examples and offers multiple answers instead of just one. The future of diagnostics could be more probabilistic than deterministic, and Tyche is an important step in this direction.

Tyche relies on a support set (context set) containing a small number of annotated examples. During inference, the model extracts features through its backbone and combines them with context tokens to generate predictions for new images. The novelty lies in the convolutional block, which introduces residual connections between segmentation candidates, allowing mutual refinement. This block carefully computes weight sums to avoid gradient explosions and maintains spatial resolution. In‑context augmentation works by adding small perturbations (rotations, scaling, noise) to test inputs and aggregating outputs through a stochastic pooling function, producing a distribution of plausible masks. The loss is a combination of Dice loss for overlap and a cross‑entropy term that rewards diversity among candidates, preventing the model from collapsing to a single deterministic solution. Experiments show Tyche matches fine‑tuning performance on datasets like ISIC and ChestX‑ray while reducing adaptation time from hours to seconds. Uncertainty analysis reveals that stochastic predictions better reflect annotator ambiguities, making it valuable for radiologists. In summary, Tyche merges few‑shot learning, in‑context augmentation, and an innovative architecture to deliver multiple, realistic segmentations without retraining.