RuiPath

A histopathology encoder learns from the slides it is shown, and slides are not interchangeable between hospitals: fixation, section thickness, H&E staining protocol and scanner optics all vary, and a representation fitted to one archive gives ground on another. Ruijin Hospital's pathologists measured this directly — an encoder pretrained on three million American slides lost accuracy when tested against their own material — and it is the reason a Chinese tertiary centre built a pathology foundation model on its own corpus rather than adopting an existing one.

RuiPath — 瑞智病理大模型 in Chinese — is that model, developed by Ruijin Hospital, Shanghai Jiao Tong University School of Medicine with Huawei and announced on 18 February 2025 at a medical AI forum held at the hospital. It covers the 19 common cancer types that account for roughly 90% of China's annual cancer incidence, along with rarer entities such as pituitary neuroendocrine tumours, and supports over 100 auxiliary diagnostic tasks. Its core vision encoder, RuiPath-VisionFoundation-V1.0, was published separately on 30 June 2025, making it the first clinical large model released by a Chinese hospital.

The system is layered. A pan-cancer visual encoder supplies features; above it sits a cross-level visual-language alignment stage and a long-sequence reasoning stage, an arrangement the developers call PanVL-T1. That stack is what lets one model recognise tumour in an image, write out the microscopic findings for a region of interest, and hold a multi-turn diagnostic conversation, rather than returning a single label.

#Key Features

• A shared encoder rather than a per-hospital rebuild: the published configuration specifies num_classes: 0 and 1024-dimensional features — a headless representation extractor that downstream hospitals attach task heads to, not a classifier refitted at each site.
• Pretraining held apart from evaluation: the encoder was pretrained on Ruijin's single-centre archive only, and the twelve public computational pathology datasets were used exclusively for testing, so the reported transfer results measure genuine generalisation rather than in-domain recall.
• Text output alongside the diagnosis: the multimodal layer generates descriptive microscopic-findings text for a selected region and answers pathology knowledge questions, with common-question accuracy reported above 90%.
• Trained on domestic silicon: the million-slide pretraining run was completed on a 16-card cluster of Chinese accelerators over roughly two months.
• Released with a matched benchmark: the encoder shipped alongside the RuiPath Pan-Cancer Benchmark WSI Dataset V1.0 — 700 quality-controlled whole-slide images across seven cancer types with 21 downstream task labels — giving recipients clinical material to test against.

#Technical Details

RuiPath-VisionFoundation-V1.0 is a vit_large_patch16_224 encoder: ViT-Large, 16-pixel patches, 224-pixel input tiles, token pooling, no classification head, 1024-dimensional output features. It was pretrained by DINOv2 self-supervision on approximately one million whole-slide images from Ruijin's digital pathology archive. On 14 auxiliary diagnostic tasks drawn from 12 public datasets — among them BACH, BCNB, BRACS, CPTAC-COAD, LC25000, MHIST, PanNuke, PatchCamelyon and WSSS4LUAD — the encoder reaches the best reported result on 7, scored by AUC, accuracy and F1. Huawei reports distilling a general-purpose model into a 32-billion-parameter domain expert for the language half of the system. No preprint, technical report or peer-reviewed paper describes RuiPath, so the pretraining recipe, evaluation protocol and per-task numbers are known only through the developers' own publication and launch materials.

#Applications

The model targets cancer diagnosis in hospitals without a deep pathology bench. Around 5,000 Chinese hospitals have a pathology department, but fewer than 5% use digital slides in routine diagnosis, and pathologists concentrate in tertiary centres. In Ruijin's own department, per-slide AI review runs in seconds against the five to ten minutes a manual microscope pass takes, shifting the pathologist from reading slide by slide to reviewing the model's findings step by step. Hospitals elsewhere in China have downloaded the released encoder to train their own downstream task heads, and a global multi-centre programme with 12 institutions was launched alongside the open-source release to broaden the training corpus into cancers Ruijin sees rarely.

#Impact

RuiPath established a template other Chinese hospitals have followed: publish the encoder and a matched evaluation set, and let smaller centres build on a clinically grounded representation instead of assembling a corpus of their own. Its successor, RuiPath 2.0, extends the same line to 205 diagnostic tasks. The evidence base, however, rests entirely on self-reported figures, with no peer-reviewed or independent evaluation, and the release is narrower than "open source" suggests: the encoder is CC-BY-NC-SA and the benchmark set CC-BY-NC-ND, both obtainable only through an institutional application and approval process that forbids redistribution, and no training or inference code was published. Pathology encoders such as UNI2-h and Virchow remain the reference points the field benchmarks against.