RuiPath 2.0

A pathologist signing out a resection specimen does not answer one question. They decide whether tumour is present, then which histological subtype it is, then how deep it invades, then how many mitotic figures sit in ten high-power fields — and each answer occupies a different line of the report. Computational pathology has largely answered those questions one at a time, which is why putting AI into a working laboratory has meant assembling and maintaining a shelf of single-task, single-cancer classifiers.

RuiPath 2.0 — 瑞智病理大模型 in Chinese — is the second generation of a clinical pathology model built by Ruijin Hospital, Shanghai Jiao Tong University School of Medicine together with Huawei Cloud, announced on 30 August 2026 at a medical AI forum in Shanghai. The 7-billion-parameter model spans 19 common cancer types and 205 diagnostic tasks, which its developers put at more than 90% of a pathologist's routine caseload. Those 205 tasks are the determinations Ruijin's own pathologists record, not a research benchmark taxonomy, and the count is up from 104 in the February 2025 first generation.

The design premise of the line is that one fixed checkpoint should serve many hospitals rather than being rebuilt at each site. The RuiPath 2.0 base model is evaluated on 59 downstream diagnostic tasks and reported to lead on 42 of them, measured against the pathology encoders the field benchmarks on — UNI2-h and Virchow2. Hospitals that want a local variant fine-tune on top of that shared base rather than pretraining their own.

#Key Features

• Broad diagnostic coverage in one model: 19 common cancer types and 205 diagnostic tasks are served by a single checkpoint, replacing the per-cancer, per-task classifier shelf that earlier deployments required.
• Structured interpretability outputs: alongside the diagnosis the model emits explicit histological subtype, depth of invasion and mitotic figure annotations, so a reviewing pathologist can trace which findings produced the call.
• Synthetic training data for rare entities: rare presentations are chronically under-sampled in any single archive, so a pathology multimodal generative model was used to synthesise additional lymphoma training material; accuracy on lymphoma presence-or-absence is reported at 96.48%.
• Low-data site adaptation: hospitals can run reinforcement learning against their own slides on top of the released base model using under 10% of a conventional training set, producing a site-specific variant tuned to local staining, sectioning and scanning.
• A lightweight companion for local hardware: a separate 2-million-parameter variant, RuiPath 2.0 Edge, was announced at the same event to run on an ordinary hospital PC in cloud-edge coordination with the main model.

#Technical Details

RuiPath 2.0 is a 7-billion-parameter model operating on H&E-stained whole-slide images. Its base model reaches the best reported result on 42 of 59 downstream diagnostic tasks, against 7 of 14 for the first generation's base model, and the developers report roughly 90% accuracy in routine use. No preprint, technical report or peer-reviewed paper describes RuiPath 2.0; its architecture, pretraining corpus, evaluation protocol and the identity of the 59 benchmark tasks have not been published, and every public technical claim traces to the launch announcement rather than to a reviewable source. The one architectural anchor in the family is the predecessor: the RuiPath vision foundation model released in June 2025 is a ViT-L/16 encoder trained with DINOv2 self-supervision on Ruijin's million-slide archive. Nothing comparable has been stated for this generation.

#Applications

The model is aimed at cancer diagnosis in hospitals that lack a deep pathology bench, a gap that is acute in China, where experienced pathologists concentrate in tertiary centres. It is delivered through Huawei Cloud's smart healthcare zone as a hosted service, and the developers report it in routine use at more than 90 hospitals, among them Chongqing University Three Gorges Hospital, Pu'er People's Hospital, Rui'an People's Hospital and Shijiazhuang People's Hospital. The intended workflow is triage and second-read support rather than autonomous sign-out: the model proposes a structured set of findings and a pathologist reviews them.

#Impact

RuiPath 2.0 is notable less for a methodological advance than for the scale at which a pathology foundation model has been placed inside routine diagnostic workflow, and for treating the interpretable intermediate findings a report actually needs as first-class outputs rather than post-hoc attribution maps. That said, the evidence base is thin: all performance figures are self-reported through a launch announcement, no independent or peer-reviewed evaluation exists, and the benchmark suite cannot be reproduced from what has been published. Weights for this generation are not publicly downloadable — only the first generation's vision encoder and a 700-slide pan-cancer evaluation set were released, under non-commercial licenses (CC-BY-NC-SA and CC-BY-NC-ND) and an application-and-approval process that forbids redistribution.