Research map / Detection and segmentation / Open-set and few-shot detection

Tyche: Stochastic In-Context Learning for Medical Image Segmentation

CVPR 2024 · Detection and segmentation › Open-set and few-shot detection · approach: Interactive and referring segmentation · 16 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Tyche, a stochastic in-context learning framework for medical image segmentation that generates multiple plausible masks for unseen tasks without retraining, via a SetBlock architecture (Tyche-TS) or in-context test-time augmentation on existing ICL models (Tyche-IS).

Problem

Medical image segmentation models need retraining for each new task and usually output a single deterministic mask, ignoring annotator disagreement and ambiguity.

Approach

Condition a UNet-like network on a context set and Gaussian noise channels, with SetBlock layers letting candidate predictions interact and a best-candidate Dice loss to encourage diversity; alternatively, apply in-context test-time augmentation to target and context of a deterministic ICL model to get stochastic outputs.

Key results

• On multi-annotator out-of-domain tasks, Tyche-TS reaches LIDC-IDRI max Dice 0.91 and Hippocampus 0.88 versus UniverSeg 0.84 on Hippocampus and 0.67 on LIDC-IDRI.
• Tyche-TS achieves GED of 0.40 on LIDC-IDRI, comparable to stochastic upper bounds such as PhiSeg (0.33) and CIDM (0.42), while using 1.7M parameters.
• Tyche-TS runs in 18.09 ms for 8 predictions versus 2,857 ms for SegGPT, and SetBlock ablation shows it is crucial to best-candidate Dice.

Measured effects reported in the paper

Surprising findings

• Tyche-IS, a test-time augmentation wrapper with no extra training, is competitive with the trained Tyche-TS on single-annotator best candidate Dice.
• SegGPT performs very poorly on STARE (max Dice 0.02) despite being a large ICL model.

Limitations

• Tyche-TS is slower than some task-specific stochastic models and Tyche-IS needs additional inference time for test-time augmentation.
• GED rewards excessive prediction diversity, so it is an imperfect metric.
• Stochastic upper bounds were trained on the out-of-distribution tasks, which Tyche does not use.

Experimental setup

Datasets: MegaMedical QUBIQ LIDC-IDRI STARE Hippocampus PanDental WBC SCD ACDC SpineWeb

Benchmarks: LIDC-IDRI Hippocampus Prostate QUBIQ STARE PanDental WBC SCD ACDC SpineWeb

Metrics: best candidate Dice Generalized Energy Distance inference time parameter count

Baselines: UniverSeg SegGPT SENet SAM SAM-Med2D Probabilistic UNet PhiSeg CIDM

Builds on: UniverSeg CrossBlock U-Net test-time augmentation

Closest papers by meaning

• Training Like a Medical Resident: Context-Prior Learning Toward Universal Medical Image Segmentation (CVPR 2024 · similarity 0.87)
• Generative Medical Segmentation (AAAI 2025 · similarity 0.86)
• S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.86)
• Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline (CVPR 2025 · similarity 0.86)
• ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement (AAAI 2025 · similarity 0.85)
• Show and Segment: Universal Medical Image Segmentation via In-Context Learning (CVPR 2025 · similarity 0.85)