Problem

Rare gastrointestinal lesions pose significant challenges in endoscopic practice due to extreme data scarcity. This limits both the development of robust artificial intelligence (AI) diagnostic models and the effective training of novice clinicians. Traditional data augmentation methods often fail to capture the necessary clinical diversity, while existing generative models typically require substantial data or exhibit limitations such as overfitting or lack of diversity when applied to extremely low-data settings, making them unsuitable for rare disease synthesis from a single example.

Method

The researchers propose EndoRare, a retraining-free, one-shot generative framework designed for synthesizing rare endoscopic lesions. EndoRare employs a three-stage pipeline: first, a medical-domain diffusion model is pretrained on common image-text pairs, leveraging language models to extract and structure key attributes like morphology, color, pathology, and location from endoscopy reports. Second, a cross-modal lesion-feature disentanglement module aligns visual factors with their textual counterparts. Finally, during personalization, the framework freezes the diffusion backbone and learns a compact Prototype-Specific Embedding (PSE) from a single rare lesion exemplar, which is then fused with attribute-aligned Tailored Lesion Embeddings (TLE) derived from language-aligned visual concepts. This approach guides the denoising trajectory to produce diverse and clinically credible synthetic images without further fine-tuning.

Results

EndoRare demonstrated superior performance across various metrics and evaluations. Quantitatively, it achieved a favorable balance between image fidelity (FID: 280.26±11.64) and diversity (LPIPS: 0.489±0.021), outperforming baselines that often sacrificed one for the other. Expert gastroenterologists, in blinded assessments, rated EndoRare-generated images highest for overall realism and class faithfulness (average rating 2.32), confirming their clinical plausibility. When used for data augmentation, EndoRare significantly improved the performance of downstream AI classifiers, raising the macro-average PR-AUC to 0.708 and ROC-AUC to 0.978, with notable gains in sensitivity at low false-positive rates. Crucially, a blinded reader study showed that novice endoscopists exposed to EndoRare-generated cases achieved a 0.400 increase in recall and a 0.267 increase in precision, particularly for morphologically ambiguous lesions like JPS and CFT, without compromising performance on distinctive categories.

Implications

EndoRare provides a practical and data-efficient solution to bridge the rare-disease gap in medical imaging, impacting both computer-aided diagnostics and clinical education. By generating diverse and high-fidelity synthetic images from a single example, it enhances AI model robustness and significantly accelerates the learning curve for novice clinicians in recognizing rare and challenging gastrointestinal lesions. The framework's design, which avoids gradient updates on patient pixels and enforces language-informed factorization, also helps mitigate privacy risks associated with instance memorization compared to heavy fine-tuning methods. The study highlights the importance of complementing generic image similarity metrics with expert clinical judgment and task-level utility for evaluating generative models in medical contexts, paving the way for improved diagnostic accuracy and training in low-prevalence settings.