Generative Vision Atlas

core

BiomedJourney: Counterfactual Biomedical Image Generation by Instruction-Learning from Multimodal Patient Journeys

Yu Gu, Jianwei Yang, Naoto Usuyama · Microsoft, University of Washington · 2023-10 · arXiv:2310.10765

Why it matters

Learns disease-progression editing from pairs of longitudinal radiology reports, using a language model to turn two visits into an edit instruction. Latent diffusion, and one of the few genuine image-editing systems in the medical literature.

What this paper does

read: full text

Before this

InstructPix2Pix showed instruction-following image editing worked for natural images but carries no medical domain knowledge. RoentGen conditions chest X-ray generation on text descriptions alone, with no prior image as input, so it is not built to preserve patient-specific invariants such as anatomy, race or age when depicting a disease change.

The problem

Unconstrained counterfactual medical image generation, showing what a patient's chest X-ray would look like under a described disease progression while holding everything else about that patient fixed, was largely unexplored.

The idea

Mine real longitudinal patient records for supervision. Take two chest X-rays of the same patient taken at different times, use GPT-4 to turn the pair of radiology reports into a short natural-language description of the progression between them, and train an editing model on the resulting prior-image, progression-text, new-image triples.

How it works

Editing happens in VAE latent space, extending Stable Diffusion v1.5's architecture with a frozen VQGAN autoencoder and a frozen BiomedCLIP text encoder plus a trained linear projection layer, while the UNet denoiser is the component that gets trained. The prior image is concatenated with the noisy latent state, following InstructPix2Pix's conditioning scheme, while the progression text cross-attends inside the UNet. There is no mask, attention injection, or explicit locality loss protecting regions the description does not mention; region preservation is implicit, coming only from conditioning on the concatenated prior-image latent rather than from any spatial control mechanism, so editing is global in principle rather than locally constrained. Training uses a two-stage curriculum, first pretraining the UNet on 69,846 single image-report pairs from MIMIC-CXR with the prior image replaced by a constant dummy image for 200 epochs, then fine-tuning on roughly 9,354 to 10,000 real counterfactual triples with a genuine prior image for 128 more epochs, because the counterfactual triples are far too scarce to train from scratch without overfitting.

Evidence

On held-out MIMIC-CXR pairs, BiomedJourney reaches a composite CMIG score of 83.23 against RoentGen's 66.08, InstructPix2Pix's 42.12, and plain Stable Diffusion's 18.14. Race-preservation AUC rises to 97.22 from RoentGen's 84.71, and age-preservation Pearson correlation rises to 79.38 from RoentGen's 28.91. Mean pathology AUC across five common findings, atelectasis, cardiomegaly, edema, effusion and pneumothorax, is 80.54 against RoentGen's 79.61, measured with a DenseNet-121 classifier trained on CheXpert labels. Label-distribution KL divergence is 10.9 against RoentGen's 40.74, and a segmentation Dice score checking anatomical alignment is 81.05 against RoentGen's 67.38. These are all automated proxy metrics standing in for clinical judgment, not a radiologist reader study.

Limitations

The authors concede that image resolution is a likely cause of recurring errors including failure to generate very fine-grained changes, and that early versions of the model produced severe hallucinations, duplicated organs and ribs, traced to view mismatches and resolution misalignment in the training pairs before data cleaning fixed it. They concede MIMIC-CXR covers only emergency medicine, which limits the disease trajectories the model has seen, that they did not explore more powerful image or text encoders or full fine-tuning of all components, and that additional accuracy and feature-retention measurements could be incorporated into the evaluation. The training data itself skews toward older and white patients.

Why it matters

It shows instruction-based editing can be trained without hand-written edit instructions by mining the causal structure already present in longitudinal clinical records, and it demonstrates that preserving patient identity under a described disease change is achievable through conditioning alone, without any explicit spatial control mechanism.

Abstract, in the authors' own words

Rapid progress has been made in instruction-learning for image editing with natural-language instruction, as exemplified by InstructPix2Pix. In biomedicine, such methods can be applied to counterfactual image generation, which helps differentiate causal structure from spurious correlation and facilitate robust image interpretation for disease progression modeling. However, generic image-editing models are ill-suited for the biomedical domain, and counterfactual biomedical image generation is largely underexplored. In this paper, we present BiomedJourney, a novel method for counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. Given a patient with two biomedical images taken at different time points, we use GPT-4 to process the corresponding imaging reports and generate a natural language description of disease progression. The resulting triples (prior image, progression description, new image) are then used to train a latent diffusion model for counterfactual biomedical image generation. Given the relative scarcity of image time series data, we introduce a two-stage curriculum that first pretrains the denoising network using the much more abundant single image-report pairs (with dummy prior image), and then continues training using the counterfactual triples. Experiments using the standard MIMIC-CXR dataset demonstrate the promise of our method. In a comprehensive battery of tests on counterfactual medical image generation, BiomedJourney substantially outperforms prior state-of-the-art methods in instruction image editing and medical image generation such as InstructPix2Pix and RoentGen. To facilitate future study in counterfactual medical generation, we plan to release our instruction-learning code and pretrained models.

Research line

Medical: borrowed natural-image latentdominant

Method note — the shared flow-matching interpolation

Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:

xt=(1t)x0+tx1,vθ(xt,t)x1x0x_t = (1-t)\,x_0 + t\,x_1, \qquad v_\theta(x_t, t) \approx x_1 - x_0

Builds on

Nothing recorded yet.

Built on by

Nothing recorded yet.