cs.CVNov 17, 2026

Reference-Guided Data Curation for 3D Medical Vision-Language Pretraining

Authors: Faidra Patsatzi, Ashwin Kumar, Chong Wang, Curtis P. Langlotz, Reinhard Heckel, Akshay S. Chaudhari

Organizations: Stanford University · Technical University of Munich

Abstract

Training 3D vision-language models (VLMs) on large medical imaging datasets collected across institutions is constrained by data quality. Vast, uncurated pools of image-report pairs are characterized by redundancy, class imbalance, and cross-institutional heterogeneity. We address this challenge of data curation for improving 3D VLM pretraining by proposing an offline, training-free, reference-guided data curation recipe. Our approach leverages a small, high-quality reference set to independently extract quality signals for each modality of a large, uncurated pool and merges them through a rank-aggregation step into a single ranking that guides sample selection. We validate this method on datasets comprising 35k 3D abdominal CT volume-report pairs from two institutions and 58k 2D chest X-ray image-report pairs. On 3D CT, using our reference-guided data curation strategy, our curated model matches the zero-shot disease classification performance of a state-of-the-art baseline (Merlin) while using 3.3×3.3\times less training data and a reference set of 5k samples. On 2D chest X-rays, our method yields a curated model that matches or exceeds the zero-shot disease classification performance of a model trained on the full pool while using up to 10×10\times less training data and a reference set of only 74 samples. Ablations show that the selection is stable under resampling of the reference, that smaller reference sets still outperform random sampling, and that label rebalancing does not improve performance. To support reproducibility, we release our code.

Figures & tables

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

May 12, 2026

20/20 Vision Language Models: A Prescription for Better VLMs through Data Curation Alone

Data curation has shifted the quality-compute frontier for language-model and contrastive image-text pretraining, but its role for vision-language models (VLMs) is far less established. We ask how far data curation alone can take VLM performance, holding architecture, training recipe, and compute fixed and varying only the training data. Our pipeline, applied to the MAmmoTH-VL single-image subset, lifts performance by +11.7pp on average across 20 public VLM benchmarks (spanning grounding, VQA, OCR/documents, captioning, spatial/3D, counting, charts, math, brand-ID, and multi-image reasoning) and by +11.3pp on average across all nine capability axes of DatBench, our high-fidelity VLM eval suite. At 2B, our curated model surpasses InternVL3.5-2B by 9.9pp at ~17x less training compute and closes the gap to Qwen3-VL-2B to within 1.8pp at ~87x less compute, from pretraining alone. Beyond accuracy, curation delivers four further properties: (1) Reliability: per-capability std across training seeds drops by ~67% and the lift survives a 4k-to-16k context-length sweep; (2) OOD generalization: the 9-eval OOD average rises by +7.2pp, and multi-image BLINK rises by +3.09pp despite single-image-only training, with Visual Correspondence gaining +11.8pp; (3) Behavioral gains beyond benchmarks: across ~1,100 open-ended queries the curated 2B is more honest and more specific than the matched-compute baseline, and more concise and less refusal-prone than a frontier 2B reference; (4) Pareto-dominance on inference cost: at every scale (1B, 2B, 4B) the curated model raises accuracy while lowering response FLOPs vs. the matched-compute baseline, and the curated 4B matches near-frontier accuracy at 3.3x lower response FLOPs than Qwen3-VL-4B. Data curation is a high-leverage tool for building better VLMs, reaching near-frontier accuracy at up to ~150x less training compute.
Mar 2, 2026

Learning to Read Where to Look: Disease-Aware Vision-Language Pretraining for 3D CT

Recent 3D CT vision-language models align volumes with reports via contrastive pretraining, but typically rely on limited public data and provide only coarse global supervision. We train a 3D CT vision-language model on 98k report-volume pairs (50k patients) collected at a single hospital, combined with public datasets, using SigLIP-style contrastive pretraining together with prompt-based disease supervision in the shared vision-text embedding space. On CT-RATE, our model achieves state-of-the-art text-to-image retrieval (R@10 31.5 vs. 22.2) and competitive disease classification (AUC 83.8 vs. 83.8), with consistent results on Rad-ChestCT (AUC 77.0 vs. 77.3). We further observe that radiologists routinely reference specific images within their reports (e.g., ``series X, image Y''), linking textual descriptions to precise axial locations. We automatically mine 262k such snippet-slice pairs and introduce the task of intra-scan snippet localization -- predicting the axial depth referred to by a text snippet -- reducing mean absolute error to 36.3 mm at 12 mm feature resolution, compared with 67.0 mm for the best baseline. Adding this localization objective leaves retrieval and classification broadly unchanged within confidence bounds, yielding a single unified model for retrieval, classification, and intra-scan grounding.
Apr 15, 2026

CLIP Architecture for Abdominal CT Image-Text Alignment and Zero-Shot Learning: Investigating Batch Composition and Data Scaling

Vision-language models trained with contrastive learning on paired medical images and reports show strong zero-shot diagnostic capabilities, yet the effect of training batch composition on learned representations remains unexplored for 3D medical imaging. We reproduce Merlin, a dual-encoder model that aligns 3D abdominal CT volumes with radiology reports using symmetric InfoNCE loss, achieving a zero-shot macro F1 of 74.45% across 30 findings (original: 73.00%). We then investigate two axes of variation. First, we control the normal-to-abnormal ratio within training batches at 25:75, 50:50, and 75:25 using section-level balanced sampling on the full dataset. All three configurations underperform the unbalanced baseline by 2.4 to 2.8 points, with 75:25 achieving the best result (72.02%) among balanced variants. Second, we conduct data scaling ablations on a 4,362-study subset, training with 20%, 40%, and 100% of the data. Performance scales sub-linearly from 65.26% to 71.88%, with individual findings varying dramatically in data sensitivity. Enforcing 50:50 balanced sampling on the same subset further degrades performance to 68.01%, confirming that explicit class balancing hurts regardless of dataset or balancing granularity. Our results indicate that the stochastic diversity of random sampling, combined with Merlin's alternating batching over anatomical subsections, provides more effective regularization than engineered class ratios at the small batch sizes required by 3D medical volumes.