NEVA

Treatment intensity in neuroblastoma turns on facts a slide alone does not give up: whether NMYC is amplified, whether 1p36 or 11q23 are deleted, and where the tumor falls on the Shimada favorable/unfavorable axis. The molecular half of that panel requires FISH or sequencing, out of reach for many centers treating children, and the histologic half carries known interobserver variability. NEVA (NEuroblastoma Vision–language AI) attacks both halves with the material that is always present — the H&E whole-slide image and the free-text pathology report — and outputs the eleven endpoints that drive management. It was developed at the Children's Hospital of Chongqing Medical University with collaborators at Stanford, Tsinghua Shenzhen, Harvard Medical School and Ant Group, and published in Nature Communications in July 2026.

Its mechanism is a two-stage, pathologist-inspired workflow that departs from the usual recipe of freezing a patch encoder and running multiple-instance learning over every tile. Stage 1 mimics low-power scanning: attention scores from a pretrained MIL slide encoder combine with morphology metrics — nuclear-to-cytoplasmic ratio and tissue quantity — to narrow tens of thousands of tiles to roughly 200. Stage 2 reprocesses only those tiles end to end, adapting the MUSK BEiT-3 backbone with low-rank adaptation while fusing the encoded pathology report. A slide of about 5,000 patches is processed in ~15.6 s on an RTX 4090 versus ~75.6 s for the frozen-encoder-plus-MIL alternative, and gradients still reach the vision encoder.

"Unified" here means one architecture and one recipe spanning diagnosis, biomarker prediction and prognosis, with a separate released checkpoint per endpoint. NEVA was benchmarked against ten pathology foundation models, including TITAN, CONCH v1.5, UNI, Virchow, CHIEF, Prov-GigaPath and CTransPath.

#Key Features

• Hierarchical tile screening: A hybrid score combining MIL attention with histologic quality metrics selects the diagnostically relevant ~200 patches per slide, cutting inference cost roughly fivefold while preserving fine-grained morphology.
• End-to-end adaptation, not frozen features: LoRA adapters on the query and value projections of the last three backbone blocks let the vision encoder specialize on a tumor barely represented in general pathology pretraining corpora.
• Image–report fusion: Slide features and the encoded report are combined with FiLM conditioning; the multimodal model beat both unimodal variants on 8 of 11 tasks, most sharply for ALK (0.764 versus 0.571 image-only and 0.648 report-only).
• Molecular readouts from H&E: Macro-average AUROCs of 0.924 for NMYC amplification, 0.830 for 1p36 deletion, 0.776 for 11q23 deletion and 0.703 for CMYC overexpression.
• Attention maps checked by pathologists: Blinded board-certified pathologists confirmed that top-1% attention patches carry the expected features — small round tumor cells in neuroblastoma, schwannian stroma in ganglioneuroma, karyorrhexis and "bull's-eye" nucleoli in NMYC-amplified cases.

#Technical Details

NEVA wraps the MUSK multiway transformer (BEiT-3) in a CLAM-style MIL aggregator with per-modality classifiers. Only the LoRA parameters and the head layers train: 196,608 adapter parameters out of 674,996,225 total, about 0.03%, with the pretrained backbone frozen. Inference takes up to 200 patches at 224 px alongside a 500-token report. The development cohort held 1,238 patients, 3,593 WSIs and 1,419 reports; the Chongqing cohort split into 711 training and 178 internal test cases, with evaluation on two temporally separated prospective series (n = 78 and n = 179) and three external centers (n = 46, 27 and 19). Diagnostic AUROCs reached 0.916 for subtype, 0.823 for Shimada classification, 0.806 for risk group and 0.791 for MKI, with C-indices of 0.717 for overall survival and 0.645 for progression-free survival. On integrated INRG pretreatment categories it was correct for 75 of 85 patients.

#Applications

The intended setting is routine pediatric pathology, especially where molecular assays are slow, costly or unavailable: a scanned H&E slide and the report already written for it yield a subtype call, Shimada and MKI grades, surrogate molecular status and a survival risk score in one pass. Eleven task-specific checkpoints are downloadable and run with no retraining, but each holds only the LoRA adapter and classifier head — about 23 MB — so essentially all the parameters come from the gated MUSK backbone users must obtain separately. The README and the paper's code-availability statement claim the MIT License, but the repository contains no LICENSE file and GitHub detects none, and the vendored TRIDENT and MUSK subtrees carry CC BY-NC-ND 4.0 notices that restrict much of the pipeline to non-commercial academic use. The underlying slides, reports and clinical annotations are withheld under patient-privacy and IRB constraints and reach researchers only by request under a data use agreement.

#Impact

NEVA shows that a rare pediatric cancer with no dedicated pretraining corpus can still be served by a general pathology backbone, provided the adaptation is end-to-end and the clinical text comes along. Its benchmark — eleven endpoints, six cohorts, ten competing foundation models — is the most complete public evaluation of computational pathology on neuroblastoma to date. The limits are real: every cohort is from Chinese centers, two external sets contain fewer than 30 patients, follow-up was too short to validate long-term survival endpoints, and for 1p36, 11q23 and CMYC the predicted status did not reproduce the prognostic separation seen with assayed status. Broader prospective validation remains the prerequisite for clinical use.