TheseusSyn: Unified Medical Image Synthesis via Context-Conditioned Masked Autoregression

Abstract

High-fidelity medical image synthesis is promising for data-efficient learning and privacy-preserving model development. However, most existing approaches rely on symbolic or label-based conditioning, which constrains generative diversity and limits case-specific contextual fidelity. Inspired by the Ship of Theseus paradox, we propose TheseusSyn, a novel context-conditioned synthesis paradigm that generates images through iterative content replacement. Our framework employs a multi-codebook discrete tokenizer to convert medical images into sequences of token indices, extracted with a global representation token that serves as a contextual constraint. A masked autoregressive generative model then iteratively masks and predicts tokens conditioned on both local neighborhoods and the global context. By progressively replacing image content until the entire token sequence is renewed, TheseusSyn produces synthetic samples that preserve anatomical coherence and semantic consistency while introducing controlled diversity. We evaluate the proposed framework on multiple datasets from MedMNIST by constructing synthetic datasets with varying masking strategies and training ResNet-18 classifiers exclusively on either real or synthetic data. Experimental results demonstrate that classifiers trained on TheseusSyn-generated images achieve competitive performance on official test sets, indicating that the synthesized data preserve diagnostically relevant features and semantic consistency. Systematic analysis of masking strategies reveals their impact on structural fidelity and downstream classification accuracy. These results suggest that context-conditioned masked autoregression provides an effective and generalizable solution for label-free medical image synthesis, offering a scalable alternative to conventional data augmentation and class-conditioned generative models in medical imaging applications.