NeCS-Flow: Necrosis-Guided Contrast-Enhanced MRI Synthesis with Rectified-Flow Refinement

Alzaga Goni, E.; Zhao, W.; Adams, R.; Moinian, S.; Xiong, Y.; Wu, T.; Yan, K.; Shpanskaya, Y.; Vasireddi, A.; Houk, J.; Calabrese, E.; Ma, D.

doi:10.64898/2026.10.21.760821

bioRxiv · 10.64898/2026.10.21.760821

NeCS-Flow: Necrosis-Guided Contrast-Enhanced MRI Synthesis with Rectified-Flow Refinement

Alzaga Goni, E.·Zhao, W.·Adams, R.·Moinian, S.·Xiong, Y.·Wu, T.·Yan, K.·Shpanskaya, Y.·Vasireddi, A.·Houk, J.·Calabrese, E.·Ma, D.

View original source ↗Export metadata ↓

Abstract

Background: Gadolinium-based contrast agents raise concerns about tissue deposition, adverse reactions, and increased costs. Purpose: To develop and validate a deep learning pipeline for gadolinium-free post-contrast T1w brain MRI synthesis incorporating necrotic scaling and Posterior-Mean Rectified Flow (PMRF) refinement, and to evaluate diagnostic utility through a blinded multi-reader study. Study Type: Retrospective. Population: 454 patients with adult-type gliomas (grade 2-4) from four publicly available datasets. Training used 374 patients; internal testing used 60 held-out patients; external validation used 20 patients. Field Strength/Sequence: 1.5-3.0 T; pre-contrast T1w, T2w, and FLAIR as inputs; post-contrast T1w as synthesis target. Assessment: A three-stage pipeline comprising nnU-Net Necrotic core segmentation, necrotic-scaling-guided U-Net Contrast Synthesis, and PMRF perceptual refinement (NeCS-Flow) was evaluated using whole-brain and tumor-region similarity metrics, enhancing tumor Dice scores, perceptual quality metrics, and a blinded three-reader study assessing diagnostic accuracy, enhancement detection, image quality, and real-versus-synthetic discriminability. Statistical Tests: Paired t-tests and two one-sided tests for model comparison; Fisher's exact test and two-sample t-tests for reader study comparisons, with p < 0.05 considered significant. Results: Necrotic scaling consistently improved real versus synthetic enhancing tumor Dice internally (HGG: 0.793 to 0.816) and externally (HGG: 0.465 to 0.612). NeCS-Flow outperformed two benchmark architectures on all quantitative metrics in internal and external testing, although absolute enhancing tumor accuracy declined for all models on the external cohort. PMRF refinement improved perceptual realism across both tumor types, reducing both FID and KID. Readers distinguished synthetic from real images only modestly above chance (overall accuracy 61.7%), and no significant differences between real and synthetic images were found for any clinical metric. Data Conclusion: NeCS-Flow synthesizes perceptually realistic and diagnostically equivalent post-contrast T1w brain MRI, with necrotic scaling improving enhancing tumor delineation and PMRF refinement improving perceptual realism across gliomas. Performance on an external cohort was preserved relative to benchmark methods, though reduced in absolute terms.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Alzaga Goni, E., Zhao, W., Adams, R., Moinian, S., Xiong, Y., Wu, T., Yan, K., Shpanskaya, Y., Vasireddi, A., Houk, J., Calabrese, E., Ma, D.. 2026-10-26. NeCS-Flow: Necrosis-Guided Contrast-Enhanced MRI Synthesis with Rectified-Flow Refinement. https://doi.org/10.64898/2026.10.21.760821

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Single-cell foundation models transfer across deep evolutionary distances

Single-cell foundation models are trained mostly on human and other vertebrate cells, yet are proposed as general-purpose models. Whether they transfer far beyond their training species is unknown. We evaluated zero-shot embeddings from Geneformer, scGPT and UCE against PCA and scVI in developmental atlases of Dictyostelium discoideum and Caenorhabditis elegans, separated from humans by 1.2 billion and 600 million years. Without retraining, the models recovered developmental order and preserved organism-specific cell identities. UCE encodes genes by protein sequence and could use 97% of measured genes, whereas Geneformer and scGPT, after ortholog mapping, used under one-fifth. UCE performed most consistently, often exceeding species-native baselines. Restricted to ortholog-mappable genes, it still scored above vocabulary-based models on five of six annotation tasks, some narrowly. Foundation models can thus yield informative cell embeddings across more than a billion years of evolution, and protein-sequence gene representations make these embeddings more complete and often more accurate.

bioinformatics↗

bioRxiv2026

EMBER: Contextualized gene representations enable compositional prediction across cellular perturbations

Predicting the effects of individual and combined perturbations is central to identifying context-specific vulnerabilities and designing targeted interventions. Yet learning gene representations that integrate functional relationships with cellular state and transfer across perturbation tasks remains challenging. Here, we present EMBER, a gene-centric foundation model that learns context-dependent functional representations of genes. EMBER combines interaction-guided contrastive learning with masked expression reconstruction during pretraining on a single-cell corpus curated for diversity and balanced sampling across cell types and experimental batches. EMBER decouples gene representation learning from task-specific prediction through a composable embedding framework: the pretrained encoder remains frozen, while lightweight probes combine contextualized gene embeddings to predict individual and combined perturbation effects. The same pretrained representations support single-gene perturbation-response prediction, digenic interaction prediction, and multi-target drug-response prediction. EMBER improves cross-context perturbation-response prediction and genetic-interaction ranking relative to evaluated baselines. EMBER also prioritized SCFD1, a mediator of autophagosome-lysosome fusion, as a candidate dependency during mTOR inhibition. Together, these results establish contextualized gene embeddings as reusable building blocks for modeling genetic and pharmacological perturbations, supporting the prioritization of genetic interactions and context-specific interventions for experimental testing.

bioinformatics↗

bioRxiv2026

Small molecule structure elucidation from NMR spectra with generative and reasoning models

Elucidating the structures of unknown small molecules remains a major challenge in chemistry, pharmacology, and drug discovery. While nuclear magnetic resonance (NMR) spectroscopy is the gold standard for structure elucidation, interpreting spectra remains labor-intensive and requires expert knowledge to navigate the vast space of possible structures. Here, we present an AI system for rapid, automated de novo structure elucidation that combines a molecular generative model with a scientific reasoning agent. Its generative component, ChefNMR2, is an atomic flow-matching model that predicts small-molecule structures from molecular formulae and one-dimensional 1H and 13C NMR spectra. ChefNMR2-Agent augments our generative model with large language models (LLMs) to reason over and refine candidate structures. Across diverse benchmarks, ChefNMR2 substantially outperforms existing tools, offering calibrated confidence estimates and two to three orders of magnitude faster inference than frontier LLMs. We use ChefNMR2-Agent to elucidate the structures of four novel natural products, matching human-chemist accuracy in a fraction of the time. Our work demonstrates how specialized generative models and general-purpose reasoning agents can unite to solve challenging scientific inverse problems and accelerate NMR structure elucidation, thus opening new possibilities for AI-driven small-molecule discovery across diverse disciplines.

bioinformatics↗