Leidos Deploys Scan Van to Cut Rare Cancer Diagnosis Times to 10 Days
Ai4
Analog slide logistics delay rare cancer treatments by months, prompting federal healthcare networks to digitize pathology archives and train AI on sub-200 slide datasets.
## Background
Published by Ai4 on YouTube on August 20, 2026, this presentation features Todd Minnigh, Chief of Medical Imaging Innovation at Leidos and steering committee member for CPACE at the University of Pittsburgh. According to the page description, the talk explores strategies for gathering scarce medical data, developing computational pathology AI for rare cancers, and translating these models into clinical practice.
## TL;DR
The presentation highlights the challenge of diagnosing rare cancers, which account for 25% of all cancer cases but lack sufficient digital tissue slide data for training standard AI models. Minnigh illustrates the human cost of diagnostic delay through his niece Katie's 109-day journey with alveolar soft part sarcoma, where physical slide logistics delayed treatment. To solve data fragmentation, Leidos and the VA created the 'Scan Van' to digitize physical archives on-site at federal health institutions. This pipeline aims to accelerate diagnostic timelines from months to under ten days by building robust AI models from sub-200 slide datasets.
## Atmosphere & style
The session carries a blend of deeply personal urgency and authoritative technical expertise, set against the backdrop of a large 700-seat auditorium. Minnigh balances solemn, emotional storytelling about his niece's diagnostic delays with a pragmatically optimistic vision for federal healthcare technology. The delivery feels engaging, field-tested, and collaborative, shifting smoothly from personal impact to technical hardware deployment and interactively resolving practical audience questions about data governance.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) One in four cancer cases (25%) is classified as a rare cancer by the National Cancer Institute, making rare strains far more common collectively than realized.
- [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) Most practicing physicians and pathologists have rarely or never encountered specific rare cancer types, leading to significant diagnostic uncertainty.
- [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) Advancements in computational pathology are reducing the data requirement for AI models, shifting from 5,000 slides to under 1,000, with active efforts aiming for sub-200 slide models.
- [00:33](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=33s) Analog logistics—shipping physical glass slides across institutions—cause massive delays in rare cancer treatment, as demonstrated by a 109-day diagnostic odyssey.
- [00:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=53s) Outside the 50 common cancer types, there are over 500 distinct rare cancer subtypes scattered sparsely across thousands of hospitals nationwide.
- [00:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=53s) To overcome severe data fragmentation in local hospitals, data collection leverages federal healthcare networks like the VA, DoD, and the Joint Pathology Center's 55 million slide repository.
- [00:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=53s) A mobile 'Scan Van' with robotic slide sorters was developed to digitize physical glass slide archives directly on site at regional health centers.
- [01:55](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=115s) Internal slide digitization within federal health systems functions as an internal data migration, avoiding initial IRB delays until external model training occurs.
- [21:15](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1275s) While common cancers like breast and prostate already have massive digitized datasets and FDA-cleared AI models, rare cancers suffer from extreme data sparsity.
- [21:15](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1275s) Isolating rare cancer cases from the dominant volume of common cancer cases within data ingestion pipelines is a critical step for developing effective diagnostic AI.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) Introduction & The Reality of Rare Cancers
Addressing an auditorium with roughly 700 seats, Todd Minnigh opens by asking the audience if anyone knows someone diagnosed with a rare cancer. Pointing out the surprising statistics, he notes that the National Cancer Institute (NCI) classifies one in four cancer instances as "rare"—meaning a quarter of all individuals with cancer are fighting a rare strain. The focus of the presentation is how artificial intelligence can be created to aid physicians in identifying these elusive conditions when clinical experience is scarce.
Minnigh introduces himself and his background as Chief of Medical Imaging Innovation at Leidos, which he describes as "the 17 billion dollar company you've never heard of." As a major government contractor, Leidos handles everything from autonomous ships and hypersonic missile testing to deploying the electronic medical record (EMR) for the Department of Defense and completing 1.5 million veteran medical exams in the past year alone. Alongside his work at Leidos and the VA, Minnigh serves on the steering committee for the Computational Pathology and AI Center of Excellence (CPACE) at the University of Pittsburgh.
When diagnosing rare cancers, the root issue is that most practicing physicians have never encountered the specific disease—or have only seen it once or twice. AI decision support could bridge this gap, but building models with severely limited datasets is traditionally difficult. Minnigh highlights ongoing advancements with his colleague Dr. Hooman Rashidi at the University of Pittsburgh, noting how computational pathology has rapidly progressed.
- **1 in 4 (25%)**: Proportion of all cancer cases categorized as rare by the National Cancer Institute.
- **5,000 to 1,000 slides**: Dr. Hooman Rashidi's benchmark shift in the number of slide images required to build an effective pathology AI model.
- **Sub-200 slide challenge**: Minnigh's active challenge to research teams to construct reliable AI models with fewer than 200 tissue slides.
- **Reference papers**: Research work by Dr. Liron Pantanowitz (Head of Pathology at UPitt) and Jay Pantanowitz on low-data AI model development.
---
[00:33](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=33s) Katie’s Story: The Diagnostic Odyssey of Alveolar Soft Part Sarcoma
Sharing a personal story, Minnigh speaks about his niece, Katie, who inspired his focus on rare cancer workflows. Katie is a healthy, young woman close to his family who went to a suburban Chicago hospital assuming she had a routine hernia. During surgery at Good Shepherd Hospital, doctors discovered a mass instead. Unsure of what it was, they referred her to Northwestern Memorial in downtown Chicago. Further research revealed her condition was an extremely rare cancer called **alveolar soft part sarcoma**, which occurs only about 80 times a year across the entire US population of 330 million. Because Northwestern lacked extensive case data, her records had to be sent to Dana-Farber Cancer Institute in Boston, which held one of the few historical cohorts with around 12 cases.
```
[Oct 24, 2022] Initial Hernia Surgery (Good Shepherd Hospital)
│
▼
[Biopsy & Unsure Mass Found] ──► Referral to Northwestern Memorial
│
▼
[Analog Specimen Shipped] ──► Sent to Dana-Farber (Boston)
│
▼
[Feb 10, 2023] First Targeted Treatment Commenced (109 Days Later)
```
The diagnostic process took from October 24, 2022, until February 10, 2023, for Katie to receive her first treatment. Minnigh emphasizes that the actual diagnostic analysis took less than a week; the remaining three and a half months were lost entirely to physical logistics and analog delays. Pathology remains 90% analog, requiring glass slides to be packaged, shipped, unwrapped, re-sliced, and reviewed by distant tumor boards.
Had the operating surgeon performed a frozen section during the initial procedure and received instant AI-assisted feedback via a cloud database, Katie’s outcome could have been vastly different. Without knowing the tumor type, the surgeon removed the mass immediately, causing severe abdominal wall tissue loss and leading to metastasis. Standard protocol at specialized centers like Dana-Farber would have called for initial radiation treatment to shrink the tumor safely before surgical removal.
- **80 cases/year**: National occurrence rate of alveolar soft part sarcoma across 330 million people.
- **109 days**: Total delay between initial surgery and proper targeted therapy for Katie due to physical slide shipping logistics.
- **12 cases**: The total case archive at Dana-Farber, representing one of the largest single-institution datasets in the country for this sarcoma.
---
[00:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=53s) Scaling Up Data Collection: The Federal Advantage and the "Scan Van"
To prevent these delays, Minnigh's team set a goal to compress the timeline from diagnosis to treatment down to less than 10 days using digital infrastructure and AI inference. Outside of approximately 50 common cancers (such as breast, lung, prostate, and stomach cancers), there are over 500 distinct rare cancer subtypes. Across the 5,331 acute care hospitals in the United States, an average facility might see the most common "rare" cancer roughly 7.5 times a year—yielding only 75 archival slides over a 10-year retention period. For ultra-rare cancers like Katie's, individual community hospitals rarely have a single sample.
To solve this data fragmentation without waiting for hundreds of independent university hospitals to share proprietary datasets, Minnigh turned to the largest integrated healthcare network in the country: the US Federal Government. The Department of Veterans Affairs (VA) and Department of Defense (DoD) operate over 200 hospital sites (172 VA hospitals and 42 military facilities) and host the Joint Pathology Center (JPC)—the largest pathology reference repository in the world, holding over 55 million glass slides spanning more than a century.
To harvest these dispersed physical specimens, Leidos developed an innovative solution: a mobile, high-speed digital scanning facility housed inside a custom truck—the "Scan Van." Equipping this vehicle with slide-sorting robotics and scanning hardware allows field teams to drive directly to regional health centers, digitize physical glass slide archives on site, and upload high-resolution images into a centralized cloud database for model training.
- **50 vs. 500**: Roughly 50 common cancer types versus 500 distinct rare cancer types.
- **5,331**: Total number of acute care hospitals across the United States.
- **55 million**: Microscope slides stored at the Joint Pathology Center (JPC), dating back over 100 years.
- **PACT Act Connection**: Legislative framework granting automatic benefits to veterans exposed to burn pits, Agent Orange, and jet fuel, driving federal interest in identifying rare toxicities and cancers.
- **Hardware Integration**: Built in California, refined at the University of Pittsburgh to develop AI slide-sorting algorithms, and deployed to the Tampa VA hospital.
---
[01:55](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=115s) Q&A: Institutional Review Boards (IRB) & Scope of Data Collection
Opening the floor to questions, an audience member asks how the team handles Institutional Review Board (IRB) approvals and patient data-sharing restrictions when collecting slides from various clinics and cancer centers, noting that IRBs often take months to process approvals.
Minnigh clarifies that during the initial digitizing phase within federal facilities, the operation functions purely as an internal data migration process—moving existing medical records from analog glass to digital bits entirely inside the VA or DoD systems. Because the data remains within the institution's firewall, no IRB is required for the collection itself. An IRB process is only triggered later when external pharmaceutical partners or commercial developers seek access to train third-party AI models or conduct formal clinical studies.
A second audience member asks whether the mobile unit collects slides strictly for rare cancers or if they are building a universal cancer database across all tissue types.
- "What took me 5,000 slides to make a good model a few years ago, now I can do with a thousand or less."
- "The actual diagnosis took less than a week. Everything else was logistics."
- "What if the answer to rare cancer research, diagnosis, and treatment was sitting in the cardboard box of a hospital a thousand miles away from you?"
<ElicitationsGroup message="Explore related topics or next steps:">
<Elicitation label="Discuss AI computational pathology approaches" query="What are the current computational pathology methods used to train AI models on small datasets like rare cancers?"/>
<Elicitation label="Examine the PACT Act and VA digital pathology" query="How is the PACT Act accelerating digital pathology adoption across Veterans Affairs hospitals?"/>
</ElicitationsGroup>
[21:15](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1275s) Distinguishing Common vs. Rare Cancers in Data Pipeline Strategy
The discussion highlights the distinct challenges of building diagnostic models for common versus rare cancers. Diseases such as prostate cancer and breast cancer benefit from being widely studied, heavily documented, and already greatly digitized across major healthcare systems. Because of this rich historical data, several FDA-cleared algorithms are already actively deployed in clinical environments for detecting common malignancies, including breast and prostate cancers.
However, the real frontier for medical imaging AI lies in capturing data for rare cancers—types that most general pathologists encounter infrequently and cannot diagnose with high confidence. While rare cancers form the primary target dataset for training specialized AI systems, filtering out common pathologies remains a critical technical hurdle. The speaker emphasizes that properly separating common cancers from rare ones within the data ingestion pipeline is a complex but essential task for ensuring directional diagnostic accuracy.
- **Data Availability Contrast:** Common cancers (breast, prostate) have massive, digitized datasets, whereas rare cancers suffer from severe data sparsity.
- **Regulatory Status:** FDA-cleared algorithms already exist for identifying well-studied conditions like prostate and breast cancer.
- **Core Challenge:** Isolating rare cancer cases from the dominant volume of common cancer cases is a primary technical target for building effective computational pathology tools.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) the 17 billion dollar company you've never heard of. — Todd Minnigh
> [01:55](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=115s) What took me 5,000 slides to make a good model a few years ago, now I can do with a thousand or less. — Todd Minnigh
> [01:55](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=115s) The actual diagnosis took less than a week. Everything else was logistics. — Todd Minnigh
> [01:55](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=115s) What if the answer to rare cancer research, diagnosis, and treatment was sitting in the cardboard box of a hospital a thousand miles away from you? — Todd Minnigh
## Mentioned
Todd Minnigh, Leidos, University of Pittsburgh, Computational Pathology and AI Center of Excellence (CPACE), MediConsult Capital, National Cancer Institute (NCI), National Institutes of Health (NIH), FDA 510k, Department of Veterans Affairs (VA), Department of Defense (DoD), Dr. Hooman Rashidi, Dr. Liron Pantanowitz, Jay Pantanowitz, Good Shepherd Hospital, Northwestern Memorial Hospital, Dana-Farber Cancer Institute, Joint Pathology Center (JPC), PACT Act, Scan Van
## Video description
Todd Minnigh - Chief of Medical Imaging Innovation, Leidos. University of Pittsburgh, Steering Committee - Computational Pathology and Ai Center of Excellence (CPACE), MediConsult Capital, Advisory board
How can we build AI that will help identify cancers so rare that most pathologists cannot diagnose with confidence?
25% of all cancer instances are of a 'rare' type by the National Cancer Institute of NIH. For many of these the data to build AI is extremely limited. Creation is a serious challenge, then validation for FDA 510k is an even bigger hurtle. Imagine if your medical team could diagnose something that they have never seen before with enough directional confidence to take the next step. This talk will explore strategies and tactics for gathering data, creating AI, and methods to get it into clinical hands. We will touch on what drives Innovation in teams and explore the foundation of innovation in business and products.
With 7 patents in medical imaging and a few more pending, steering committee position with University of Pittsburgh Computational Pathology and Ai Center of excellence, and 30+ years of experience in medical imaging and Ai, I hope to give you the inspiration to find the next advancement as you work here on the cutting edge of Ai in healthcare.
Todd Minnigh of Leidos argues that the key obstacle in treating rare cancers is not computational inability, but the analog bottlenecks of physical pathology archives that delay diagnosis and starve diagnostic AI models of training data.
The Diagnostic Crisis of Rare Cancers (0:04)
According to the National Cancer Institute, one in four cancer cases is classified as rare, spanning over 500 distinct subtypes outside of roughly 50 common cancers. Most practicing community physicians rarely or never encounter these conditions in clinical practice. Minnigh illustrates this with the case of his niece, who was diagnosed with alveolar soft part sarcoma—a disease occurring roughly 80 times per year across the United States. While Dana-Farber held one of the nation's largest cohorts with approximately 12 cases, the diagnostic journey took 109 days from initial surgery on October 24, 2022, to the start of targeted therapy on February 10, 2023.
Minnigh highlights that the diagnostic analysis itself took less than a week, while more than three months were lost because pathology remains roughly 90% analog. The physical shipping of glass slides delayed proper identification. Had instant decision support been accessible during the initial procedure, standard protocols like preoperative radiation could have preceded resection, rather than immediate surgery that caused severe tissue loss and subsequent metastasis.
Shrinking Data Thresholds in Computational Pathology (0:04)
Building artificial intelligence models for rare diseases has historically been hindered by the scarcity of training data. However, computational pathology requirements are dropping rapidly. Work by Dr. Hooman Rashidi at the University of Pittsburgh shows that the benchmark volume required for an effective pathology model has decreased from 5,000 slides to 1,000 or fewer. Minnigh has pushed research teams toward a sub-200 slide challenge to train reliable models on severely limited cohorts, referencing low-data framework research by Dr. Liron Pantanowitz and Jay Pantanowitz.
Mobilizing Federal Repositories and the Scan Van (0:53)
To compress the diagnostic and treatment initiation timeline down to under 10 days, Minnigh’s team bypasses fragmented private systems by leveraging federal healthcare infrastructure. Across the country's 5,331 acute care hospitals, individual facilities might only see a specific rare cancer several times per decade. In contrast, the Department of Veterans Affairs and Department of Defense operate over 200 hospital sites and oversee the Joint Pathology Center, which houses more than 55 million physical slides accumulated over a century.
To digitize these dispersed physical archives without risking specimen loss, Leidos developed the 'Scan Van,' a custom vehicle outfitted with high-speed slide scanners and robotic slide sorters. The unit was engineered in California, calibrated with sorting algorithms at the University of Pittsburgh, and deployed to the Tampa VA hospital. This setup allows technicians to digitize physical glass slides directly on site and upload high-resolution images into a centralized cloud repository.
Regulatory Pathways and Ingestion Pipelines (1:55)
Addressing administrative hurdles, Minnigh states that on-site scanning across federal facilities operates as an internal data migration from analog glass to digital formats behind institutional firewalls. Consequently, Institutional Review Board (IRB) approval is not required during the collection phase, though an IRB process is triggered later if external commercial or pharmaceutical entities request access to train third-party models.
A central operational hurdle in this pipeline is distinguishing common malignancies from rare targets. While common cancers such as breast and prostate cancer already have extensive digital archives and several FDA-cleared diagnostic algorithms, rare cancers suffer from severe data sparsity. Technical pipelines must effectively filter out high-volume, well-documented common pathologies to isolate the rare specimens required for targeted model training.
Minnigh notes that federal interest in rare cancer identification is currently accelerated by the PACT Act, which provides benefits to veterans exposed to environmental toxins. However, the operational challenge remains scaling physical slide scanning quickly enough to turn archival specimens scattered across nationwide facilities into functional diagnostic algorithms.
## Background
本视频由科技会议频道 Ai4 于 2026 年 8 月 20 日在 YouTube 平台发布。视频内容为 Leidos 医学影像创新主管、匹兹堡大学计算病理学与 AI 卓越中心指导委员会成员 Todd Minnigh 的主题演讲，探讨如何在极小样本数据下构建医疗影像 AI，以攻克占全美 25% 却极度缺乏诊断经验的罕见癌症难题，并介绍推动其临床转化的策略与实践。
## TL;DR
演讲指出全美四分之一的癌症属于罕见病种，但绝大多数病理医生因“见所未见”且缺乏训练数据而面临严重的误诊与诊断延误。主讲人通过侄女 Katie 罹患极罕见肉瘤且因传统玻片物理转运延误三个半月的切肤之痛，阐明了病理数字化与小样本 AI 辅助决策的紧迫性。针对单家医院样本极度匮乏的数据孤岛困境，Leidos 联合退伍军人事务部（VA）与联合病理中心（JPC），利用拥有 5,500 万张历史切片的联邦医疗宝库，并创新性地研发部署了搭载高容错机械臂与质控算法的“移动扫描车”（Scan Van），深入各大医疗机构现场将沉睡的模拟切片转化为高价值数字资产，以不到 200 张切片的极限小样本攻坚罕见癌症 AI 建模。
## Atmosphere & style
演讲开场幽默风趣且极具现场互动感，随后迅速切入侄女患病求医的真实悲剧，叙事沉痛而饱含深情，极具情感冲击力与感染力。主讲人将直击人心的个人痛点与严谨的数学推算、联邦宏观数据战略以及前沿工程实践紧密交织，展现出从悲痛转化为推动系统性变革的坚定力量。整场分享既有对传统低效医疗流转体制的深刻反思，又有以“卡车装载扫描仪”物理破局的硬核极客务实感，令听众在深受触动的同时对医疗 AI 的落地路径充满启发。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) 全美四分之一（25%）的癌症病例属于罕见癌症，但临床医生往往极度缺乏罕见病种的诊断经验，形成'见所未见'的诊疗困境。
- [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) 针对罕见癌症训练数据极度匮乏的问题，计算病理学建模正向极限小样本挑战，尝试仅用不到 200 张数字切片构建高精度 AI 模型。
- [03:30](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=210s) 传统病理诊断严重依赖物理玻璃切片的跨机构邮寄与流转，导致罕见癌症确诊周期长达数月，延误最佳治疗窗口。
- [08:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=533s) 罕见肉瘤等病种若在未确诊情况下盲目手术切除，极易造成严重组织损伤并加速癌细胞扩散，亟需 AI 提供术中即时决策支持与诊疗指南建议。
- [08:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=533s) 罕见癌症研发的最大瓶颈是严重的数据孤岛，各顶尖癌症中心由于病例分散及商业学术壁垒，单家机构均无法凑齐足够的训练与验证数据。
- [13:10](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=790s) 美国联邦退伍军人事务部（VA）与国防部拥有封闭庞大的医疗系统及联合病理中心（JPC）保存的 5,500 万张百年切片，是攻克数据稀缺的天然富矿。
- [15:54](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=954s) Leidos 研发部署了搭载机械臂与高速扫描仪的'移动扫描车'（Scan Van），深入各医院档案室直接将沉睡的物理切片转化为数字资产。
- [15:54](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=954s) 车内专门开发了切片质控 AI 算法，可前置识别老旧切片的破损、污渍与墨水标记，实现智能分拣与高容错数字化扫描。
- [19:09](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1149s) 在合规路径上，机构内部的切片扫描属于'数据迁移工程'，初期无需 IRB 审批；仅在后续训练外部 AI 或跨机构共享时再履行伦理审查。
- [21:15](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1275s) 相较于已获 FDA 批准且数据充沛的常见癌症算法，当前构建医学影像 AI 的核心挑战在于如何从海量数据中精准甄别并分离出罕见癌症样本。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) 罕见癌症的挑战与 Leidos 的背景
演讲开场时，主讲人 Todd Minnigh 用轻松幽默的语气调侃现场近 700 个座位座无虚席的场面，随后向全场抛出一个问题：“在座有谁认识患有罕见癌症的人吗？”现场举手的人寥寥无几。然而 Todd 指出，根据美国国家癌症研究所（NCI）的统计，全美有整整四分之一（25%）的癌症病例属于“罕见癌症”。这意味着只要你认识癌症患者，其中就有四分之一的概率是罕见病种。本次演讲的核心议题，正是探讨如何利用人工智能（AI）帮助病理学家和临床医生准确识别这些极度缺乏诊断经验的罕见癌症。
Todd 介绍了自己的背景与供职机构——Leidos。他将其戏称为“一家你从未听说过的 170 亿美元巨头”，因为作为美国政府最大的主承包商之一，Leidos 的业务涵盖了从无人自主舰艇、高超音速导弹测试技术，到为美国国防部（DoD）部署电子病历系统（EMR），去年仅退伍军人医疗检查就完成了 150 万人次。Todd 本人深耕医学影像领域，长期与美国退伍军人事务部（VA）合作，并担任匹兹堡大学计算病理学与 AI 卓越中心（CPACE）指导委员会成员。
在罕见癌症的临床诊断中，医生面临的最大困境是“见所未见”：绝大多数病理医生职业生涯中可能仅见过一两次，甚至从未遇见过某些亚型。AI 辅助决策在此类场景下本应发挥关键作用，但行业痛点在于“极度缺乏训练数据”。针对小样本建模，Todd 提到匹兹堡大学计算病理学领域的权威专家 Houman Rashidi 曾对他说：“几年前我需要 5,000 张切片才能训练出一个优秀的模型，现在只需 1,000 张甚至更少。”而 Todd 正在向团队发起更激进的挑战——尝试仅用不到 200 张数字病理切片来构建高精度 AI 模型。
- **罕见癌症占比**：占全美所有癌症病例的 25%（每 4 例癌症中即有 1 例为罕见型）。
- **Leidos 业务体量**：年营收 170 亿美元的联邦承包商，涉足国防、自主技术及退伍军人医疗服务（去年完成约 150 万次体检）。
- **学术合作**：依托匹兹堡大学计算病理学与 AI 卓越中心（由 Liron Pantanowitz 主持，其子 Jay Pantanowitz 亦为研究员），推进极限小样本模型构建。
---
### [03:30](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=210s) 侄女 Katie 的故事：漫长而致命的诊断延迟
Todd 话锋一转，分享了一段极具情感冲击力的亲身经历。如果不是因为他的亲侄女 Katie，他今天绝不会站在这里投身于这个领域。Katie 是他妻子姐姐的女儿，两家人关系极为亲密，过去 20 年里有 15 个圣诞节和几乎每个感恩节都在一起度过。Katie 身材高挑、美丽、健康，是 Todd 大女儿一直仰慕的榜样。
然而在去年，年轻健康的 Katie 感觉腹部不适，起初以为只是普通的疝气，便前往芝加哥郊区的一家社区医院（Good Shepherd Hospital）接受常规疝气修复手术。但在手术中，外科医生意外发现了一个肿块。由于无法确认性质，医院将样本转诊至芝加哥顶尖的西北纪念医院（Northwestern Memorial Hospital）。然而即使是顶级医学中心也无法给出确切结论。家属多方查阅文献后发现，该组织疑似一种极度罕见的肉瘤——腺泡状软组织肉瘤（Alveolar Soft Part Sarcoma, ASPS）。全美 3.3 亿人口中，该病每年仅新发约 80 例，全美唯一对其有深入研究的只有波士顿的丹娜—法伯癌症研究所（Dana-Farber Cancer Institute），而该机构历史累计也仅有约 12 例相关病例。
从 2022 年 10 月 24 日 Katie 接受首次探查手术，到 2023 年 2 月 10 日终于获得针对性方案开始首次治疗，整整耗费了三个半月（超过 100 天）。Todd 沉痛地反问：为什么确诊需要这么久？病理诊断本身只需要不到一周，其余所有时间全被消耗在了原始、低效的物理物流与机构周转上——由于 90% 的病理工作仍依赖传统模拟玻片，标本必须在 Good Shepherd、西北纪念医院与丹娜—法伯之间反复打包、邮寄、切片、染色、组织肿瘤委员会（Tumor Board）会诊并安排异地就医。
```mermaid
flowchart TD
accTitle: 传统罕见癌症转诊流程与理想AI辅助流程对比
subgraph Traditional["传统模拟转诊流程（历时 3.5 个月）"]
A["社区医院初次手术<br/>(发现不明肿块/冷冻切片)"] --> B["标本物理邮寄至顶级综合医院<br/>(耗时周转/无法确诊)"]
B --> C["标本转运至专科研究所 Dana-Farber<br/>(重新切片染色/肿瘤委员会讨论)"]
C --> D["最终确诊并制定方案<br/>(病情延误/已有转移)"]
end
subgraph Proposed["理想数字化与AI驱动流程（预期 < 10 天）"]
E["手术室现场扫描玻片"] --> F["云端病理大模型推理<br/>(跨机构聚合知识库)"]
F --> G["即时反馈高置信度诊断<br/>与推荐安全干预指南"]
G --> H["术中正确决策/精准靶向治疗"]
end
```
---
### [08:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=533s) 改变诊疗范式：AI 如何打破恶性循环
Todd 强调，缩短从发病确诊到开始治疗的时间窗口是当务之急。他们的目标是将这一周期压缩到 10 天以内，而在医疗资源分布不均的现状下，AI 是实现这一目标的唯一路径。即便退一步，单纯实现全流程数字化切片远程流转，也能极大改善会诊效率。
更严重的后果在于误诊误治带来的不可逆伤害。Katie 最初的手术医生是一位普通外科医生，术中进行了快速冷冻切片（Frozen Section）检查。医生在显微镜下初步判断“看起来像肉瘤”，随即凭经验决定将其切除。然而当样本辗转送到丹娜—法伯时，顶级肉瘤专家却表示这是极其不幸的处置——对于腺泡状软组织肉瘤，最安全的临床路径是“不要立即切除”，而是先行缝合切口，通过放疗使肿瘤缩小并稳定边界后再行手术切除。直接切除不仅导致 Katie 失去了大面积的腹壁肌肉组织，更直接造成了癌细胞的扩散转移。
如果当时手术室里的医生能够直接将切片图像上传至云端模型，AI 哪怕仅给出“90% 概率为极罕见肉瘤，建议遵循先放疗后切除指南”的提示，Katie 的预后就会截然不同。现有技术完全能够支撑这种云端推理，真正的瓶颈在于数据孤岛：全美前 50 种常见癌症（如乳腺癌、肺癌、前列腺癌等）占据了绝大多数临床与科研资源，而剩下的 500 种罕见癌症呈现高度长尾分布；丹娜—法伯有 12 例、其他医院有 1 例或 3 例，没有任何一家机构能够单独凑齐 200 到 1,000 张切片用于模型训练与验证。药企在开发靶向药物时，也因极小样本量缺乏足够的生物标志物（Biomarkers）形态学数据，即便有《罕见病药物法案》（Orphan Drug Act）的支持，研发动力依然严重不足。
---
### [13:10](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=790s) 联邦医疗数据宝库与全美样本分布
解决数据极度匮乏的唯一办法就是“抽干大海”（Boil the Ocean）——将分散在全美各机构的历史切片彻底汇集起来。那么，全美最大的未开发医疗数据富矿究竟在哪里？答案是美国联邦政府体系。退伍军人事务部（VA）与国防部（DoD）构成了美国最大的封闭式公立医疗系统，涵盖 172 家 VA 医院、42 家军区医院，拥有自己的医生、实验室与全量数据。
更为关键的是，该体系下设有全球最大的病理参考实验室——联合病理中心（Joint Pathology Center, JPC）。JPC 完整保存了过去 100 多年来的 5,500 万张显微镜玻璃切片。Todd 分享了一个历史细节：当年科学家之所以能够成功测序 1918 年西班牙大流感（H1N1）的病毒基因组，正是得益于 JPC 库房中妥善封存的百年前人体组织切片。
Todd 通过数学推算展示了全美病理数据的稀缺性与集中度：
- **急救医院基数**：全美共有 5,331 家急救医院（Acute Care Hospitals）。
- **发病率最高的罕见癌症**（年发病 40,000 例）：平均每家医院每年接诊约 7.5 例；若按医院常规保存 10 年切片档案计算，单院库存仅约 75 例。
- **极度罕见癌症**（如 ASPS，年发病 80 例）：80 例分散在 5,331 家医院，绝大多数机构几十年内都难遇一例。
- **顶尖转诊中心**：全美约有 300 家大学附属医院与 NCI 认证癌症中心，疑难患者往往自发流向丹娜—法伯、MD 安德森（MD Anderson）或梅奥诊所（Mayo Clinic）。但这些顶级中心彼此存在学术与商业竞争，切片数据共享阻力极大；而 VA 系统则完全不存在这种竞争壁垒。
---
### [15:54](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=954s) “移动扫描车”：打破数据孤岛的创新实践
为了将沉睡在各大医院库房纸箱中的模拟切片转化为数字资产，Leidos 团队提出并落实了一个大胆的工程方案：“癌症治疗的突破口可能是一辆卡车”——一辆装载高通量高速数字化病理扫描仪的移动扫描车（Scan Van）。
这一行动恰逢立法与政策的强力支持。美国国会通过的《尊严退伍军人全面效益法案》（PACT Act）规定，退伍军人若暴露于露天垃圾焚烧坑（Burn Pits）、橙剂（Agent Orange）或航空燃油尾气，一旦罹患特定罕见癌症可自动获批医疗福利。目前美国参议院正在审议的《军事建设与退伍军人事务拨款法案》（MilCon-VA Bill）中，也明确加入了支持罕见病理数字化的专项条款。
移动扫描车项目目前已完成原型研发并投入实地部署：
- **研发与验证轨迹**：车辆在加利福尼亚州完成主体改装与集成，随后运往匹兹堡大学优化切片分拣工作流，目前已正式进驻坦帕退伍军人医院（Tampa VA Hospital）。
- **车内硬件与 AI 算法**：
- **工业级机器人分拣系统**：位于车内核心区域，运行速度相对平缓但容错率极高，可稳定处理存在墨水标记、指纹污渍、破损等历史老旧切片。
- **高通量高速扫描单元**：用于快速处理标准切片。
- **切片质控 AI**：团队专门开发了前置 AI 算法，用于自动识别切片质量瑕疵并智能调度不同的分拣与扫描硬件。
```mermaid
flowchart LR
accTitle: 移动病理扫描车工作流程
subgraph DataCollection["现场数据数字化 (移动扫描车)"]
A1["各医院档案室<br/>(沉睡玻璃切片)"] --> A2["AI 视觉预检<br/>(墨迹/污损分类)"]
A2 --> A3["高容错机械臂 / 高速扫描矩阵"]
end
subgraph CentralRepository["云端与科研赋能"]
A3 --> B1["罕见癌症数字化图像库"]
B1 --> B2["药企靶向药物研发<br/>与生物标志物分析"]
B1 --> B3["小样本诊断 AI 模型研发"]
end
```
---
### [19:09](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1149s) 现场问答：IRB 伦理审查与样本采集范围
Todd 在演讲尾声总结道：“如果攻克罕见癌症诊断与治疗的关键线索，此刻正静静躺在距离你一千英里外某家医院地下室的硬纸盒里呢？”移动扫描车正是为了通过物理介入打破这种数据割裂。随后，演讲进入现场简短问答环节。
现场观众首先提出了关于机构伦理审查委员会（IRB）合规性的关键问题：移动扫描车穿梭于不同医院和癌症中心采集病理切片时，即便进行了严格的去标识化（De-identification），申请 IRB 审批通常也极其耗时，项目如何应对这一合规阻碍？
Todd 回应解释了其法律与合规路径的巧妙分层：
- **数据数字化阶段**：在 VA 或国防部体系内部进行切片扫描与图像归档，在法理上属于机构内部的“数据迁移工程”（Data Migration Process），即纯粹将模拟资产转为数字资产，因此在这一阶段**不需要**申请 IRB 审批。
- **AI 研发与跨机构共享阶段**：后续当需要利用这些脱敏数据训练外部 AI 模型，或与制药企业开展联合科研时，再针对具体的科研项目依法履行 IRB 申请流程。
在被问及扫描车采集的数据范围是仅针对罕见癌症还是涵盖所有类型癌症切片时，Todd 开始进一步阐述建库的顶层设计理念与长远目标（演讲在此处进入下一部分）。
### [21:15](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1275s) 区分常见癌症与罕见癌症的数据采集与算法挑战
在问答互动环节中，讨论聚焦于如何精准捕获罕见癌症数据。像前列腺癌（prostate cancer）和乳腺癌（breast cancer）这类常见病种已经被研究得非常透彻，并且拥有庞大的数字化病理与影像数据积累。因此，构建医疗影像 AI 的真正难点与核心需求，在于设法从数据极度稀缺的环境中获取那些罕见癌症的样本。
演讲者进一步说明，针对常见癌症领域，医疗行业目前已经具备相当成熟的技术储备，甚至市场上已有针对乳腺癌、前列腺癌以及巴氏涂片（Pap smear）筛查等获得 FDA 审批认证（FDA-cleared）的成熟算法。然而，即便核心数据源的目标是罕见癌症，如何构建一套行之有效的机制，将这些常见的癌症与罕见癌症精准鉴别和剥离出来，在实践中依然是一大考验。讲者坦言“这种甄别剥离本身就是一项挑战”（"the separating is a challenge"），而攻克这种区分与筛选能力，正是当前研发团队所锁定的关键目标（"That is the target"）。
- **成熟领域的数字化与算法现状**：乳腺癌、前列腺癌及巴氏涂片检测由于病例丰富、数字化程度高，市面上已有获得 FDA 认证的成熟算法提供辅助诊断支持。
- **罕见癌症的核心攻坚点**：必须解决如何在数据源中有效甄别（discriminating）并分离常见癌症与罕见癌症的难题，为罕见癌症模型积累高纯度的训练与验证数据。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) 在座有谁认识患有罕见癌症的人吗？ — Todd Minnigh
> [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) 一家你从未听说过的 170 亿美元巨头 — Todd Minnigh
> [00:04](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=4s) 几年前我需要 5,000 张切片才能训练出一个优秀的模型，现在只需 1,000 张甚至更少。 — Houman Rashidi
> [08:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=533s) 看起来像肉瘤 — 外科医生
> [08:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=533s) 不要立即切除 — 丹娜—法伯肉瘤专家
> [08:53](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=533s) 90% 概率为极罕见肉瘤，建议遵循先放疗后切除指南 — Todd Minnigh
> [13:10](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=790s) 抽干大海 — Todd Minnigh
> [15:54](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=954s) 癌症治疗的突破口可能是一辆卡车 — Todd Minnigh
> [19:09](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1149s) 如果攻克罕见癌症诊断与治疗的关键线索，此刻正静静躺在距离你一千英里外某家医院地下室的硬纸盒里呢？ — Todd Minnigh
> [21:15](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1275s) the separating is a challenge — Todd Minnigh
> [21:15](https://www.youtube.com/watch?v=_OwPQjZbkxk&t=1275s) That is the target — Todd Minnigh
## Mentioned
Todd Minnigh, Leidos, 美国国家癌症研究所 (NCI), 美国国立卫生研究院 (NIH), 匹兹堡大学计算病理学与 AI 卓越中心 (CPACE), MediConsult Capital, Houman Rashidi, Liron Pantanowitz, Jay Pantanowitz, Katie, Good Shepherd Hospital, 西北纪念医院 (Northwestern Memorial Hospital), 丹娜—法伯癌症研究所 (Dana-Farber Cancer Institute), 腺泡状软组织肉瘤 (ASPS), 美国退伍军人事务部 (VA), 美国国防部 (DoD), 联合病理中心 (JPC), MD 安德森癌症中心 (MD Anderson), 梅奥诊所 (Mayo Clinic), 《尊严退伍军人全面效益法案》 (PACT Act), 《军事建设与退伍军人事务拨款法案》 (MilCon-VA Bill), 坦帕退伍军人医院 (Tampa VA Hospital), 《罕见病药物法案》 (Orphan Drug Act), FDA 510(k)
## Video description
Todd Minnigh - Chief of Medical Imaging Innovation, Leidos. University of Pittsburgh, Steering Committee - Computational Pathology and Ai Center of Excellence (CPACE), MediConsult Capital, Advisory board
How can we build AI that will help identify cancers so rare that most pathologists cannot diagnose with confidence?
25% of all cancer instances are of a 'rare' type by the National Cancer Institute of NIH. For many of these the data to build AI is extremely limited. Creation is a serious challenge, then validation for FDA 510k is an even bigger hurtle. Imagine if your medical team could diagnose something that they have never seen before with enough directional confidence to take the next step. This talk will explore strategies and tactics for gathering data, creating AI, and methods to get it into clinical hands. We will touch on what drives Innovation in teams and explore the foundation of innovation in business and products.
With 7 patents in medical imaging and a few more pending, steering committee position with University of Pittsburgh Computational Pathology and Ai Center of excellence, and 30+ years of experience in medical imaging and Ai, I hope to give you the inspiration to find the next advancement as you work here on the cutting edge of Ai in healthcare.
[Source](https://www.youtube.com/watch?v=_OwPQjZbkxk)
Source quotes
[0:04] Todd Minnigh: 在座有谁认识患有罕见癌症的人吗？
[0:04] Todd Minnigh: 一家你从未听说过的 170 亿美元巨头
[0:04] Houman Rashidi: 几年前我需要 5,000 张切片才能训练出一个优秀的模型，现在只需 1,000 张甚至更少。
[8:53] 外科医生: 看起来像肉瘤
[8:53] 丹娜—法伯肉瘤专家: 不要立即切除
[8:53] Todd Minnigh: 90% 概率为极罕见肉瘤，建议遵循先放疗后切除指南
[13:10] Todd Minnigh: 抽干大海
[15:54] Todd Minnigh: 癌症治疗的突破口可能是一辆卡车
[19:09] Todd Minnigh: 如果攻克罕见癌症诊断与治疗的关键线索，此刻正静静躺在距离你一千英里外某家医院地下室的硬纸盒里呢？
[21:15] Todd Minnigh: the separating is a challenge
[21:15] Todd Minnigh: That is the target
Source quotes
[0:04] Todd Minnigh: the 17 billion dollar company you've never heard of.
[1:55] Todd Minnigh: What took me 5,000 slides to make a good model a few years ago, now I can do with a thousand or less.
[1:55] Todd Minnigh: The actual diagnosis took less than a week. Everything else was logistics.
[1:55] Todd Minnigh: What if the answer to rare cancer research, diagnosis, and treatment was sitting in the cardboard box of a hospital a thousand miles away from you?
Original transcript
[0:04] Thanks for coming. Glad y'all could find a
[0:06] seat. [laughter] I'm I'm struck by how
the there's like 700 seats in here. So,
[0:11] um so I have a question. Does anybody in
the room know somebody that has rare cancer?
[0:19] No. Uh it might surprise you to find out that uh
the national national cancer institute has said
[0:28] that uh one in four cancers are rare. So if you
know somebody with cancer, there's a one in four
[0:34] chance that that is classified as a rare cancer.
Uh and this talk is going to be about those
[0:39] particular types of cancer. Uh and h how we can
create AI to help physicians identify them. So,
[0:50] uh, good slides are up. So, this is me. Uh, I'm
Todd Men. I work for a company called Lidos. Does
[0:56] anybody know who Lidos is? What? Yeah, you
either know it or you don't. Uh, I like to
[1:02] describe Lidos as the 17 billion company you've
never heard of. And if you haven't heard of us,
[1:07] it's very unlikely you'll ever buy anything from
us. We are the one of the largest government
[1:12] contractors. We make everything from autonomous
ships to hypersonic missile testing technology. We
[1:19] deployed the EMR for the Department of Defense. We
did a million and a half veteran examinations last
[1:24] year like medical examinations. Uh and I work
in the medical imaging space uh predominantly
[1:30] with the VA uh and also in cooperation with the
University of Pittsburgh where I sit on their
[1:35] uh computational pathology and AI center of
excellence um steering committee. Uh so LIDOS is
[1:44] doing projects with the VA and this this relates
directly to one of the projects that we're doing.
[1:52] When it comes to diagnosing rare cancer, the
problem is it's rare and odds are the physician
[1:59] that's looking at it has never seen it before or
maybe they've seen it once or twice but they don't
[2:03] have a lot of experience with it. And so what
would be most helpful would be if they had AI to
[2:10] support their decision-m what's the problem with
that? How many people have tried to build AI with
[2:17] insufficient amount of data? It's really hard.
It's hard to get it to be accurate. Uh and I'm not
[2:26] going to give you a bunch of magic about synthetic
data and hyperparameter adjustment. That's not
[2:32] what this is about. This is about looking at
it from a completely different perspective. So,
[2:38] one of the leaders at the uh University
of Pittsburgh is a friend of mine named
[2:42] Human Rashidi. And he is actually one of the
foremost developers of artificial intelligence
[2:47] in the medical imaging community. And we were
having a conversation the other day and I said,
[2:52] "Can you give me a quote that I can add to my
deck?" and basically said this, "What took me
[2:57] 5,000 slides to make a good model a few years
ago, now I can do with a thousand or less." And
[3:04] I'll tell you in a few minutes how I've challenged
them to make a model with uh less than 200 slides.
[3:13] There's an article on this if you have more
uh if you want more information. Uh uh Leon
[3:19] Pantanowitz is the head of pathology
at the University of Pittsburgh and
[3:22] uh uh Jay Pantanowitz is his
son who's also a researcher.
[3:30] So I'm going to tell you a story about somebody
very close to me. This is a story of a little
[3:35] girl. Her name's Katie. She's my niece. If
it weren't for Katie, I would probably not be
[3:45] standing here right now. I would not be giving
this talk and I probably wouldn't be in this
[3:49] space. Uh here's Katie. She's the kind of niece
uh and you guys probably have families like this
[4:00] that you're really close to. So she is my wife's
sister's daughter. Uh they have four kids. We
[4:06] spent oh 15 of the last 20 Christmases together.
Uh almost every Thanksgiving were together.
[4:16] Uh, the girl in the middle is my oldest daughter.
The girl with the sunglasses on her face is
[4:23] Katie's older sister. And the one with the one
sunglasses on top is is Katie's uh Katie herself.
[4:34] She's uh tall. My my daughters looks up to her.
She's a couple years older than my daughter. Uh
[4:41] that's Katie with her dad Dan. He's probably my
best friend. Um, there's the Three Musketeers
[4:50] again and again. There's Katie at Notre Dame
a couple of years ago. Um, more than maybe
[4:57] five years ago now. Here's Katie last year.
She's young. She's beautiful. She's healthy.
[5:08] She thought she had a hernia. She went to
a small hospital. uh in Chicago, actually
[5:15] in the suburbs of Chicago, and had an operation
trying to fix the hernia. But as you can imagine,
[5:23] it wasn't a hernia or I wouldn't be talking
about it here. Uh what they discovered was
[5:28] a mass and they weren't sure what it was. So, uh
the good people at Good Shepard Hospital referred
[5:36] her to Northwestern in downtown Chicago.
Sure. Certainly Northwestern Memorial in
[5:41] Chicago can do almost anything, right? They
weren't sure. Family did some research and
[5:48] discovered the very type of saroma that they
discovered when they were doing the surgery,
[5:55] if it is that type of saroma, has really only
been studied by Dana Farber in Boston. She has
[6:02] something called alvear soft part saroma.
It occurs 80 times a year across all 330
[6:08] million of us. So nobody has any experience
with it. Dana Farber had like 12 cases.
[6:20] The thing that I would trying to solve for because
the business unit that I ran at Lidos before I
[6:26] started doing this project solved for things like
this was how do I shorten the timeline from the
[6:31] time I know somebody has an issue until they're
being treated. How do I compress that timeline?
[6:36] In her case, her first operation was on October
24th of 2022. And the time she was being treated
[6:44] for what she had was February 10th, the first
treatment of 2023. Why did it take so long? Well,
[6:52] for one thing, pathology is 90% analog.
For another thing, nobody could diagnose
[6:59] this confidently until the specimen was
actually shipped to Dana Farber. First,
[7:05] first to Northwestern, then back to Good
Shepard, then out to Dana Farber. Then they
[7:09] unwrap it. They make their own slides. They do
their diagnosis. They have their tumor board.
[7:12] They make a discussion. They call her out. She
goes out. They finally make a final diagnosis
[7:16] and schedule it for treatment. It doesn't have
to be this way. Does anybody have a relative that
[7:23] it took a long time for them to get treated
for something? One, two, three. Yeah. Four,
[7:29] five. Yeah. It doesn't have to be that way. The
actual diagnosis took less than a week. Everything
[7:37] else was logistics. And so Lidos runs a business
unit called QTC. And QTC does millions of veteran
[7:44] examinations for what's called compensation
and pension examinations. That I'll give you
[7:50] an example. Somebody was a paratrooper. Uh they
landed wrong and hurt their knee. Uh and then
[7:55] later in life they're a mailman and they discover
that the knee is now really a problem. They can't
[7:59] walk the route anymore. And so now they go to
their VA man and their VA man says, "You know
[8:05] what? you should go get an examination. They would
refer them to us and we would do the examination.
[8:09] That's why we do a million examinations a year
as a federal contractor. Actually, we 1.5 last
[8:16] year also soft heart saroma really rare any rare
cancer as I said less than 40,000 times a year.
[8:25] uh National Cancer Institute says that that
25% of all cancers, you know, there's like the
[8:32] really common stuff, breast cancer, lung cancer,
prostate cancer, stomach cancer, there's like
[8:40] 50 of them. And then in this other quadrant,
there's 500. It like fractionalizes into all
[8:45] of these tiny little things that nobody has any
experience with. That's where we really need AI.
[8:53] So this reduced time for treatment higher
purpose we said to ourselves how can we make
[8:59] it less than 10 days and obviously we're
in an AI conference the only way to make
[9:04] that happen is is through AI there's also
a notion that if you simply had it digital
[9:11] and you could get it get the digital image
in front of the right pathologist you could
[9:15] accelerate the diagnosis and I will say
yes lacking AI that would be the next best
[9:22] So what if we could literally make smart people
smarter? What if we could take the doctors that
[9:28] we have and give them the aggregate knowledge
of many doctors? What if we could build AI to
[9:35] diagnose sooner? See, Katie's surgeon didn't
know what he found. He was a general surgeon
[9:43] doing a hernia surgery. He found something. I
mean, when in surgery when they find something,
[9:48] they do what's called a frozen section. Anybody
know what a frozen section is? No. All right. So,
[9:54] they do they make a they make a slide right there
in the O. They take a little piece of tissue,
[9:59] they freeze it, they slice it, they put it
on a glass slide, just 25 by 75 millimeter
[10:05] piece of glass like you used in biology class
in high school. They put in a microscope like
[10:09] the one you had in biology class, probably a
better one. Uh, and they look at it and they
[10:12] try to make a decision. What is this?
His decision was looks like a saroma.
[10:21] His thought was I should take it out. By the time
she got to Dana Farber, Dana Farber said that was
[10:29] unfortunate. If he known what he found, the safest
path is to do nothing. Close radiation treatment,
[10:41] shrink it, then remove it. That does two things.
It prevents it from metastasizing and you take
[10:49] less tissue. As it is, she lost a huge piece
of her abdominal wall and it metastasized.
[10:56] So if that doctor in that O or in that operating
room, yes, O could have just put that up on the
[11:07] cloud database and gotten an inference and it
came back and it said 90 90% confident that
[11:12] it's this really rare thing over here. If it
is, this is the safest protocol to follow. Her
[11:17] outcome would be completely different. There
is nothing that's stopping us from doing this
[11:22] with the technology that exists today. except that
the data is fragmented. Dana Farber has 12 cases,
[11:31] another hospital has one, another hospital
has one, another hospital has three or four.
[11:36] Nobody has 200, 500, a thousand so that I
can build an AI model from. And it's not
[11:42] just this one cancer. There's 500 cancers. So
in this case, boiling the ocean is the answer.
[11:54] So if you had boiled the ocean and you had all of
the data from all of the hospitals in one place,
[12:02] the far pharmaceutical companies who build cancer
drugs, they identify whether or not a a cancer is
[12:09] likely to respond to their drug by looking for
what they call biomarkers. They literally look
[12:14] at the morphology of the cell and try to make a
decision. Does this cell have the properties of
[12:19] cells that typically respond to this? But
if something happens 80 times, 100 times,
[12:25] a thousand times a year, it's not worth their
effort. I mean, financially, it doesn't make any
[12:29] sense for them to do that. Even with the Orphan
Drug Act, it's still hard to get energy around
[12:35] doing that. But what if he could make it easy for
them? What if Katie's doctors knew what they had
[12:42] before they tried to proceed with treatment? So AI
can do that for us. I don't think there's anybody
[12:51] in this room that needs to even read this slide
to know what it says. I need lots of samples. I
[12:55] need to curate them. I need to annotate them.
I need to say that's it. That's not it. Train
[13:00] the model. I think if you're here, you probably
know this part. Where am I going to get them?
[13:10] Who has this data? So, one of the largest healthc
care organizations in the United States is the US
[13:20] federal government. We don't think of ourselves
as socialized healthcare, but you know what? The
[13:26] VA and the DoD are exactly that. They own the
doctor. They own the They own the lab. They own
[13:31] the hospital. They own the data. And they have
over 200 sites. There's 172 VA hospitals and 42
[13:41] uh military hospitals and they have the largest
reference lab for pathology in the world. They
[13:47] have something called the joint pathology center.
It's got 55 million microscope slides going back
[13:51] to more than 100 years ago. Interesting story
about this joint pathology center. While I'm
[13:57] on the subject, the genome for the Spanish flu,
you know how they always say this one's H1N1,
[14:04] this one is this, this one is that? You know
how they know the genome for the Spanish flu?
[14:08] They had actual tissue samples from the
Spanish flu at the at the joint pathology
[14:12] center. They were able to extract some
of the tissue and do genetic tests on it.
[14:20] So there's 5331 acute care hospitals in the United
States where somebody might have discovered they
[14:26] have cancer. Good Shepard just happens to be one
of them. The most common of these rare cancers
[14:32] occurs 40,000 times a year. So, if you have the
very very most common rare cancer, the average
[14:38] hospital has seen it seven and a half times this
year. If they keep a 10-year archive, which is
[14:44] typical, they would have 75 examples of it, the
most common. And then you get to the other end
[14:49] of the scale where you have what what Katie has.
It happens 80 times a year. Well, you take 80 and
[14:57] divide it by 5,000, you can figure out how many
samples they have. Most places don't have any.
[15:03] So we need to think differently. And so the idea
came to us. Two things are happening. One thing is
[15:13] I need to go to those reference labs. I need to go
to those places that do a lot of patients. People
[15:18] like Katie migrate to places like Dana Farber.
So that's a really important point. So we need
[15:25] to aggregate data from Dana Farber. Those guys
tend to be kind of guarded on their data because
[15:29] they're competing with MD Anderson. They're
competing with Mayo Clinic. But the VA is not.
[15:38] But in the end, there's really only about 300
university hospitals and NCI cancer institutes.
[15:42] So if you wanted, you could go to all of them if
you could get them all to cooperate. I wasn't able
[15:48] to get them all to cooperate, but I was able
to get the VA to cooperate. So our innovation,
[15:54] believe it or not, the cure to cancer very well
might be a truck. It's a truck that's loaded with
[16:01] high-speed digital scanning equipment to turn
those samples into microscope SL those sample
[16:08] microscope slides into digital data. Once it's
digital data, then the pharmaceutical companies
[16:15] can actually test and see whether or not they're
can actually make an inference as to whether
[16:20] they're um pharmaceutical is going to work on this
and also they can uh build AI to help diagnose it.
[16:32] The VA is interested in this. There's something
called the PAC Act. The PAC Act basically tells
[16:37] you that uh if you have one of these things or any
of another any of a number of other ailments that
[16:43] you will automatically get benefits. You don't
even need to go to somebody like Lidos for a test
[16:49] because these things are so rare that you probably
got them from your exposure to a burn pit. You got
[16:54] them from your exposure to ancient orange. You got
them from your exposure to uh jet fuel exhaust.
[17:02] These are really rare. The number after each one
of these is how frequent they happen across all
[17:05] 330 million of us. So they're you can imagine if
it happens 800 times a year or 1300 times a year,
[17:12] you're just going to have a hard time finding
enough samples to to build an AI. So what we
[17:17] did was we worked with Congress uh and actually
right now in front of the Senate is what's called
[17:22] the Milcon VA bill, the military construction
uh bill. military and V8 together and and their
[17:33] construction is all in one appropriations bill.
And there's language in there to support doing
[17:37] exactly what I'm describing here. So, we could
scan for those, but that would leave another 486
[17:45] types of cancer to find. So, what we're proposing
is we bring the scan van to the hospitals that are
[17:52] likely to have the samples in large quantity. We
scan all of their rare data and then we go to the
[17:58] next place and then we go to the next place and
we build a cancer database that can be used by
[18:06] pharmaceutical companies and AI companies to
develop the AI to diagnose these things. So,
[18:11] uh it wasn't just an idea. We literally built one
of these things. It's currently sitting in Tampa.
[18:16] uh we built it out in California and then we
um moved it to University of Pittsburgh where
[18:21] we finalized some development of uh how it's
going to work uh how it sort slides. It has
[18:27] multiple kinds of technology inside to sort
the slides. Uh this is the inside of it. It's
[18:34] uh the one the machine on the bottom is
a robot. It sort slides in a way that's
[18:40] uh uh basically it it's slow but it's very
fault tolerant. the slides can have lots of
[18:46] problems like ink stains and finger stains and
stuff like that. The little boxes are just as
[18:50] fast as that giant thing in the bottom right
corner, but they're not fault tolerant. So,
[18:55] how to sort those is one of the AI
algorithms that we had to build.
[19:01] Uh, so this is it sitting outside the University
[19:03] of Pittsburgh before we moved it
to Tampa. It's now at Tampa VA.
[19:09] So the conclusion that I'm trying to bring you
guys to is what if the answer to rare cancer
[19:17] research diagnosis and treatment was sitting in
the cardboard box of a hospital a thousand miles
[19:24] away from you. It is talk earlier today talked
about how fragmented how siloed the data is.
[19:34] It is this is a physical attempt to take the data
out of the silos. Uh this is my number if you want
[19:43] to get in touch with me. I didn't leave much time
for questions. I only got 30 seconds left. But if
[19:48] anybody has something that they want to ask about,
I'm happy to take it on. Wonderful talk. Just one
[19:56] quick question. Um how do you deal with the IRB
issue? Let's say you I I know the idea is to
[20:04] move around to different you know clinics cancer
centers but usually um they will need some IRB to
[20:10] share the slides the patient. Yeah I know it can
be certainly will be deidentified but still they
[20:17] need you know and the IRB can take a long time.
Yes. So the IRB can take a long time when it's
[20:24] time to build the AI or to share with the with the
pharmaceutical companies. aggregating the data is
[20:30] just um it's a data migration progress process.
They're moving from analog to digital all staying
[20:35] within the VA. So for the stuff that stays within
the VA or within the DoD, you actually don't need
[20:41] an IRB. It's just data migration. When they want
to build the AI, then they have to build the then
[20:47] they have to do the the IRB. So how to how to deal
with that? I we'll have to cross that bridge. But
[20:54] they do they do work with the pharmaceutical
companies and other companies. Thank you. I'm
[20:58] sorry if this sounds if this is like a this
question is too simple but um when you're
[21:02] collecting these samples from hospitals um it's
all cancer or just rare cancers are you creating
[21:09] a comprehensive database for all forms of cancers
that have been cryostat or what the spirit is to
[21:15] capture the rare cancers because uh things like
prostate cancer and breast cancer are very well
[21:19] studied and already greatly digitized. So it's
really the rare cancers that you need to get the
[21:25] data from. Mhm. Um in fact we have uh algorithms
that can help though identify things like cancer
[21:31] and epsmeir cancer and uh breast cancer uh cancer
and prostate cancer already FDA cleared algorithms
[21:37] for some of those things. Got it. So you do have a
system for discriminating the common cancers from
[21:40] the rare cancers and include by even though you're
including the the your main data source as the
[21:45] rare cancers. Yeah, the separating is a challenge.
Got it. That is the target. Yes. Thank you.
