One-shot synthesis of rare gastrointestinal lesions improves diagnostic accuracy and clinical training

TL;DR

EndoRare tackles the scarcity of rare GI lesion data by a retraining-free, one-shot diffusion framework that learns from routine image--text pairs and a single rare-lesion exemplar. It combines knowledge-enhanced diffusion pretraining with cross-modal visual concept learning and a Prototype-Specific Embedding to synthesize diverse, clinically faithful images of rare lesions. The approach yields substantial improvements in AI-assisted diagnosis and novice clinician performance, validated by objective metrics and blinded expert ratings across multiple diseases. This work offers a data-efficient pathway to bridge rare-disease gaps in diagnostics and education, while addressing privacy concerns and outlining directions for broader clinical deployment.

Abstract

Rare gastrointestinal lesions are infrequently encountered in routine endoscopy, restricting the data available for developing reliable artificial intelligence (AI) models and training novice clinicians. Here we present EndoRare, a one-shot, retraining-free generative framework that synthesizes diverse, high-fidelity lesion exemplars from a single reference image. By leveraging language-guided concept disentanglement, EndoRare separates pathognomonic lesion features from non-diagnostic attributes, encoding the former into a learnable prototype embedding while varying the latter to ensure diversity. We validated the framework across four rare pathologies (calcifying fibrous tumor, juvenile polyposis syndrome, familial adenomatous polyposis, and Peutz-Jeghers syndrome). Synthetic images were judged clinically plausible by experts and, when used for data augmentation, significantly enhanced downstream AI classifiers, improving the true positive rate at low false-positive rates. Crucially, a blinded reader study demonstrated that novice endoscopists exposed to EndoRare-generated cases achieved a 0.400 increase in recall and a 0.267 increase in precision. These results establish a practical, data-efficient pathway to bridge the rare-disease gap in both computer-aided diagnostics and clinical education.

Figures (9)

• Figure 1: Overview of the EndoRare framework.a, Clinical motivation. Long-tail rarity in gastrointestinal diseases causes extreme data scarcity, limiting both clinician exposure and AI training. b, Framework. Three stages transfer knowledge from routine to rare lesions: diffusion pretraining on routine image--text pairs, attribute alignment between images and text, and controllable synthesis of rare cases using learned embeddings and disentangled attributes. c, Evaluation. We assess (1) image quality and clinical faithfulness, (2) gains in AI-assisted diagnosis when training data are augmented with synthetic rare cases, and (3) clinical relevance via improved diagnostic performance of novice trainees.
• Figure 2: Visual realism and class faithfulness.a, Representative generated samples (rows 1–4) alongside ground-truth exemplars (bottom row) for four diseases. b, Image fidelity vs. diversity: class-wise points and per-method means (circles with s.e.m. bars) in the FID–LPIPS plane. The y-axis is inverted (lower FID is better) and higher LPIPS indicates greater diversity; better performance is towards the upper-right. c, Image–image consistency vs. diversity: CLIP-I–LPIPS scatter with per-method means (s.e.m. bars). Higher CLIP-I and higher LPIPS are desirable; better performance is towards the upper-right. d, Clinical evaluation: expert gastroenterologists score generated images on a 3-point scale (1 = unrealistic; 3 = realistic and class-faithful without exact replication).
• Figure 3: Classification performance with generated data. Eight strategies are compared across four rare-lesion categories and their macro-average (Overall). a, Precision–recall area under the curve (PR-AUC). b, Receiver operating characteristic area under the curve (ROC-AUC). c, $\mathrm{TPR}@20\%\ \mathrm{FPR}$ (true positive rate at the operating point with the false-positive rate fixed at $20\%$). Bars in a–c denote means across runs; error bars indicate standard deviation.
• Figure 4: Low-FPR detection performance (0–20% FPR).a, Per-entity ROC curves in the clinically relevant low-FPR range (JPS, CFT, FAP, PJS). b, Partial AUC (pAUC) computed over the same 0–20% FPR interval, summarized per entity and as a macro-average (Overall). Higher pAUC indicates greater cumulative sensitivity under low-false-positive operating constraints.
• Figure 5: Impact of synthetic exposure on novice diagnostic performance.a, Study workflow. Conventional training suffers from data scarcity (top), limiting the ability of novices to recognize rare patterns. In contrast, EndoRare enrichment (bottom) provides diverse synthetic exemplars to bridge this experience gap. b, Quantitative assessment. Pairwise comparison of diagnostic metrics (Recall, Precision, F1) before (Pre) and after (Post) synthetic training. Significant improvements are observed for JPS, CFT, and overall performance, while FAP and PJS exhibit ceiling effects due to high baseline distinctiveness.