Conversational Pathology Copilot
把显微镜
变成一场对话Turn the microscope
into a conversation.
Loupe 是一款会话式数字病理副驾驶：在全切片影像上一边阅片一边提问。它调用病理视觉模型与文献工具，给出可核查的推理、结构化量化与一键可采纳的标注草稿——病理医生始终掌握最终决定权。Loupe is a chat-first copilot for whole-slide imaging. Ask while you read; it calls pathology vision models and literature tools, returning verifiable reasoning, structured quantification, and one-tap annotation drafts — pathologist always in control.
Loupe slide › manuscript
从切片到投稿，AI 嵌进每一步 / one clean workflow — slide to submission
✓ 诚信优先 · 不编造数据 / Integrity-first — never fabricates data图表带 ILLUSTRATIVE 水印 · 引用核验 · AI 披露 / figures watermarked · citations verified · AI disclosed
亮点 / Why Loupe
四个让病理医生留下来的理由Four reasons pathologists stay
会话即工作流Chat is the workflow
用自然语言提问，得到带工具调用痕迹的可核查回答。Ask in plain language; get answers backed by a visible tool trace.
可核查的量化Verifiable quantification
IHC 阳性率、H-score、鉴别诊断与 PubMed 引用，就地呈现。IHC positivity, H-score, differential & PubMed — rendered inline.
智能体 + 人审Agents, human-approved
智能体批量产出草稿，你逐条接受/驳回，预算与审计可控。Agents batch-draft; you accept/reject, with budgets & audit.
可治理Governable
采纳率遥测、成本与审批率，部署可被量化与优化。Adoption telemetry, cost & approval rate — measurable, tunable.
业务流 / How a case moves through Loupe
从切片接入到出报告，一条贯穿的工作流One workflow, from slide intake to sign-out
Loupe 不是一堆零散功能，而是顺着真实诊断流程把 AI 嵌进每一步——而决定权始终在病理医生手里。Loupe isn’t a bag of features — it threads AI through the real diagnostic flow, with the pathologist deciding at every step.
01 接入与预读分诊 / Ingest & triage
切片入库，即被 AI 预读一遍Every slide is pre-read the moment it lands
上传或从 DICOMweb 导入全切片影像。Loupe 自动算出质控评分、写出一段“镜下预读”，并按器官归类——让你从最该看的切片开始。Upload or import WSIs via DICOMweb. Loupe scores QC, writes a scan-time pre-read, and groups by organ — so you start with the slide that matters.
02 会话式阅片 / Conversational read-out
问一句，得到可核查的答案Ask in plain language, get an auditable answer
选区直接提问。Loupe 调用工具读取该区域、做组织/细胞统计、量化 IHC（如 ER 92%、H-score 285）、检索 PubMed，并摆出它考虑过的鉴别诊断——不是只给结论。Select a region and ask. Loupe reads the tile, computes stats, quantifies IHC (e.g. ER 92%, H-score 285), cites PubMed, and shows the differential it weighed — not just a verdict.
03 自主智能体批处理 / Autonomous agents
一句话下达任务，智能体自己跑One sentence in, drafts out
“标出所有侵袭性肿瘤前沿并标记出芽灶”——智能体在迭代/时长/Token 预算内自主规划执行，逐张切片产出标注草稿，可对整组切片一键派发。“Map every invasive front and flag tumor budding.” The agent plans and runs within iteration / time / token budgets, drafting annotations per slide — dispatch across a whole case at once.
04 人审采纳 / Human review & adopt
AI 提草稿，你来定夺AI proposes, you decide
任务详情展示分步计划、预算用量与活动轨迹；草稿队列可逐条或批量接受/驳回，校验器把低置信项顶到队首；驳回理由进入下一轮提示。The task page shows the plan, budgets and activity trace; accept/reject drafts singly or in bulk, with a verifier surfacing low-confidence ones — and your reasons feed the next run.
05 多切片对比 / Compare in sync
H&E 与 IHC，并排同步阅读H&E and IHC, side by side, in sync
把同一病例的多张切片（H&E + MMR PMS2/MSH6，或 H&E + ER）放进同步缩放平移的对比视图——例如本例 MSH6 缺失提示错配修复缺陷、需 Lynch 筛查。Put a case’s slides (H&E + MMR PMS2/MSH6, or H&E + ER) into a pan/zoom-synced comparison — here MSH6 loss flags mismatch-repair deficiency for Lynch workup.
06 遥测与治理 / Telemetry & governance
“医生到底信不信 AI”——用数据回答Quantify whether clinicians actually trust the AI
跟踪分类器采纳率随置信度与时间的走势、按标签的接受情况，以及智能体的成本与审批率——让整个部署可被治理、可被优化；并用你已审核过的预测做 split-conformal 校准，得到一个可信阈值，只有达到该置信度的预测才标记为“可信”，覆盖率如实标注、绝不夸大。Track classifier adoption by confidence and over time, per-label acceptance, and agent cost & approval rate — so the whole deployment can be governed and tuned. Split-conformal calibration then turns your own reviewed predictions into a trust threshold: only predictions above it are badged "trusted", with the coverage stated honestly — never inflated.
界面速览 / A quick tour
三张截图，快速看懂 LoupeThree screenshots, one quick look
全部来自真实运行的产品界面——点击可放大。All from the live product — click to enlarge.
细节一览 / In detail
那些“看得见的细节”，放大给你看The load-bearing detail, up close
统一尺寸的特写——每一处关键读数都清晰可读。Uniform close-ups — every load-bearing number, clearly legible.
科研写作闭环 / From cohort to manuscript
从一张切片，一直走到投稿From a slide, all the way to submission
Loupe 把诊断台延伸成科研台：冻结队列、用真实数据算统计与出版级图表、起草手稿与基金，并在投稿前做诚信核查——全程不编造数据。Loupe extends the diagnostic desk into a research desk — freeze a cohort, compute real statistics and publication-grade figures, draft the manuscript and grant, and run an integrity check before submission. It never fabricates data.
真实统计与出图Real stats & figures
队列冻结后由 biostats 计算 ROC（含 DeLong 95% CI）、校准（Brier/ECE）、Kaplan–Meier + log-rank 并出版级渲染；数据不足则留空而非编造。From a frozen cohort: ROC (DeLong 95% CI), calibration (Brier/ECE), Kaplan–Meier + log-rank, publication-rendered — insufficient data is left blank, not faked.
模型结果成图Predictions to figures
把真实模型预测一键变成评测图与指标，登记到下载中心，可复现。Turn real model predictions into evaluation figures + metrics, filed in the download center, reproducibly.
投稿预审Pre-submission review
逐个数字核对能否溯源到队列真实数；无结果支撑的性能声称被标为过早；可叠加 reviewer-2 式批判。Every number is traced to the cohort’s real facts; unbacked performance claims are flagged as premature; an optional reviewer-2 critique on top.
样本量与功效Power & sample size
两组率/均数、log-rank 生存、单 AUC 等标准公式——投稿与基金都要求的样本量论证。Standard formulas for two proportions/means, log-rank survival, single AUC — the sample-size justification reviewers and funders require.
面上合规检查NSFC format check
按面上项目撰写提纲逐项检查必需章节、参考文献时效、篇幅与 AI 使用声明。Audits an NSFC General-Program proposal against the official outline — sections, reference recency, length, AI-use disclosure.
写作与投稿Write & submit
文献综述综合、目标导向改写（含“未引入新数字”校验）、参考文献 .bib/CSL 导出、期刊匹配、Markdown→LaTeX / Word。Related-work synthesis, goal-directed rewrite (with an “introduced-number” check), .bib/CSL export, journal matching, Markdown→LaTeX / Word.
真正的差异化是诚信：示意图带 “ILLUSTRATIVE” 水印、引用经核验、AI 使用全程披露——正好对接 NIH 禁 AI 代笔、NSF 与基金委强制披露的新规。The real differentiator is integrity: illustrative figures are watermarked, citations are verified, and AI use is disclosed — aligned with the new NIH / NSF rules on AI in papers and grants.
真实评测 · RUO / Measured, not claimed · RUO
不是宣称，是真跑出来的数Real numbers, from real runs
我们用现役最新的 Gemini（gemini-3.5-flash）直接读公开、带金标准标注的真实病理图——含真实 TCGA 全切片。真实模型、真实数据、金标准打分、如实报告，不编造、不挑数。We ran the newest Gemini (gemini-3.5-flash) directly on open, gold-labeled real pathology images — including real TCGA whole-slides. Real model, real data, gold-standard grading, reported as-is.
| 实验 / Experiment | 数据（公开） / Data (open) | 结果 / Result |
|---|---|---|
| 真实 TCGA 全切片 · 癌种识别（8 选 1）Real TCGA whole-slide · cancer-type ID (8-way) | GDC 开放 .svs · OpenSlide / GDC open-access .svs · OpenSlide | 7 / 8（随机 12.5%）7 / 8 (chance 12.5%) |
| 淋巴结转移检测（平衡，N=100）Lymph-node metastasis (balanced, N=100) | PatchCamelyon | 85% · 敏感度 0.9485% · sensitivity 0.94 |
| 病理是/否问答（N=100）Pathology yes/no VQA (N=100) | PathVQA | 79% |
能从真实切片认出大多数癌种、零拒答；但偏过度报阳、仍有漏诊——定位为提示 / 预筛，不替代病理诊断。更强的 pro 模型未见提升，flash 已够。It names most cancers from real slides with zero refusals — but over-calls positives and still misses some, so it is a triage / pre-read aid, not a replacement for diagnosis. A stronger pro model was no better; flash suffices.
⚠ 研究性评测（RUO）：样本量小、为封闭集选择题，不是临床验证，不可用于诊断决策。⚠ Research evaluation (RUO): small samples, closed-set multiple choice — not clinical validation, not for diagnostic decisions.
版本与部署 / Editions & deploy
同一套代码，两种版本，处处可部署One codebase, two editions, deploy anywhere
FastAPI + SQLAlchemy 异步后端、React 19 + OpenSeadragon 前端、OpenSlide 切片管线、可插拔多 LLM；提供 Docker / GPU / HTTPS 反代与 DICOMweb 导入。Async FastAPI + SQLAlchemy, React 19 + OpenSeadragon, OpenSlide pipeline, pluggable multi-LLM; Docker / GPU / HTTPS deploys and DICOMweb import.
| 能力 / Capability | Commercial 商业版 | Education 教育版 |
|---|---|---|
| 会话阅片 · 标注 · 报告草稿Chat read-out · annotations · report draft | ✓ | ✓ |
| 结构化 synoptic 报告（CAP：GI/乳腺/前列腺）+ HL7 v2 / FHIR R4 标准导出Structured synoptic reports (CAP: GI / breast / prostate) + HL7 v2 / FHIR R4 export | ✓ | ✓ |
| 自主智能体 · 遥测 · 多切片对比Agents · telemetry · comparison | ✓ | ✓ |
| 科研写作闭环 · 统计/出图 · 投稿预审 · 面上检查Research authoring · stats/figures · pre-submission review · NSFC check | ✓ | ✓ |
| 病理基础模型 UNI / CONCHPathology foundation models UNI / CONCH | — | ✓ (CC-BY-NC-ND) |
| 相似瓦片检索Similar-tile search | 颜色纹理指纹color/texture fingerprint | ✓ UNI embeddings |
| 镜像 / 适用场景Image / use | 最小镜像 · 云端可商用smallest · cloud-defensible | 含权重 · 仅非商业科研教学weights baked · non-commercial |
多 LLM 可插拔Pluggable LLMs
Gemini / OpenAI / MiniMax / Kimi 经统一兼容层接入。Gemini / OpenAI / MiniMax / Kimi via one compatible layer.
安全与审计Security & audit
JWT 鉴权、登录限流、全量审计事件、按病例的核查时间线、HTTPS 反代。JWT auth, rate-limit, full audit log, per-case audit timeline, TLS proxy.
开放切片与对接Open WSI & interop
OpenSlide 多厂商格式 · DZI 瓦片 · DICOMweb 导入 · 报告 FHIR/HL7 导出。OpenSlide formats · DZI tiles · DICOMweb import · FHIR/HL7 report export.