Software Read 7,056 Routine Cervical Slides and Missed None of the 45 Cancers It Was Built to Catch

experiment · The journal of pathology. Clinical research · la publicación, 1 sep 2026 · gratis

Un programa leyó las placas del cuello del útero y encontró los 45 casos de un cáncer que casi siempre se pasa por alto

El estudio se hizo en China, con placas ya guardadas y no dentro del trabajo diario de un laboratorio. Nadie puede prometer todavía la misma precisión en América Latina ni en Estados Unidos.

Versión breve · la versión detallada sigue, unos 7 min

Pregunte al chat de weeklyAI

Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.

Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.

El estudio, de un vistazo
Quiénes: Placas de tejido del cuello uterino, incluidas 45 de un cáncer agresivo llamado GAS
Cuántos: 7,056 placas
Dónde: China
Cuándo: Entre marzo de 2024 y abril de 2025
Tipo de estudio: analysis of what people did
Quién lo hizo: Tercer Hospital de la Universidad de Pekín y Universidad de Tsinghua
El límite que importa: Se evaluó fuera del trabajo diario del laboratorio y solo con pacientes y equipos de China

Cuánto vive una paciente según el tipo de cáncer de cuello uterino

Adenocarcinoma gástrico de cuello uterino (GAS)frente aAdenocarcinoma relacionado con el virus del papiloma humano

La supervivencia a cinco años es mucho menor en el GAS: 42% frente a 91%

Lo que sí muestra es que un programa barato, que lee las mismas placas que ya se hacen, puede avisar de un cáncer que hoy se escapa.
Lectura de weeklyAI

Así podría verse · ilustración generada por weeklyAI.watch, no es una fotografía

Un equipo del Tercer Hospital de la Universidad de Pekín y de la Universidad de Tsinghua, en China, construyó un programa llamado GASPath. El programa lee las placas de tejido del cuello del útero teñidas con la coloración habitual, esa que casi todos los laboratorios ya usan.

Lo entrenaron con 309 placas de 96 casos confirmados de adenocarcinoma gástrico de cuello uterino, un tipo de cáncer agresivo que suele confundirse con otros. Sumaron 1,320 placas de cuatro enfermedades parecidas.

En la primera prueba interna, el programa acertó en 0,980 de los casos, casi todos. En las biopsias, que son muestras muy pequeñas y difíciles de leer, también le fue bien.

Después lo probaron con muestras de 12 hospitales distintos. En promedio detectó poco más de 9 de cada 10 casos, aunque el resultado varió de un hospital a otro. Con un ajuste en el umbral de decisión, llegó a casi 97 de cada 100, y las falsas alarmas casi no aumentaron.

Entre marzo de 2024 y abril de 2025 lo probaron con 7,056 placas seguidas del trabajo rutinario de un hospital. El cáncer buscado era apenas el 0.6% de esos casos. Encontró los 45 casos que había.

Ese resultado tiene un borde. El programa da un aviso temprano cuando sospecha de este cáncer. No reemplaza al patólogo ni a las pruebas de confirmación. En el estudio, tres patólogos con experiencia revisaron cada caso y confirmaron el diagnóstico con pruebas de laboratorio adicionales.

Tampoco puede decirse que sirva igual en otras partes. El estudio se hizo en China, con pacientes y equipos de allí; los autores escriben que falta comprobar cómo funciona en otras poblaciones y con otros escáneres. Y la prueba grande se hizo con placas recogidas durante el trabajo diario, pero la evaluación se hizo fuera del flujo de trabajo normal, así que todavía no se sabe qué pasa cuando un patólogo lo usa en vivo.

Lo que sí muestra es que un programa barato, que lee las mismas placas que ya se hacen, puede avisar de un cáncer que hoy se escapa. Para un lector de América Latina o de América del Norte, esto no es algo para pedir en su clínica todavía. Es una señal de hacia dónde va.

Los autores están empezando un estudio en varios hospitales, con pacientes asignados al azar, para ver si la herramienta cambia la rapidez del diagnóstico, las decisiones de tratamiento y la supervivencia.

Qué significa para usted

Si en su consulta le hablan de este cáncer de cuello uterino poco frecuente, pregunte si el patólogo ya lo tuvo entre sus sospechas, porque suele confundirse con otras lesiones. El programa todavía no está disponible en América Latina ni en Estados Unidos, así que lo que puede seguir es el estudio aleatorizado que los autores iniciaron para ver si cambia el diagnóstico.

Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113

Quién pagó: El trabajo fue financiado por la Fundación Nacional de Ciencias Naturales de China, Proyectos Clínicos Clave del Tercer Hospital de la Universidad de Pekín, el Programa Nacional Clave de I+D de China, el Centro de Investigación de Ingeniería de Shenzhen, un Programa Gubernamental de Formación para Talentos Clínicos Sobresalientes, el Fondo de Investigación e Innovación Interdisciplinaria de la Escuela de Posgrado Internacional Tsinghua Shenzhen, una subvención de inicio de la Universidad de Fuzhou, la Fundación Provincial de Ciencias Naturales de Fujian y la Fundación de Ciencias Naturales de Tianjin; el artículo declara no tener conflictos de interés y no indica que los financiadores tuvieran algún papel en el estudio.

No tome esto como consejo médico profesional.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1372 palabras · unos 7 minLeerla →Cerrar

Una prueba de laboratorio detectó los 45 casos de un cáncer de cuello uterino especialmente letal, en un programa chino de 7,056 placas

Un sistema de inteligencia artificial leyó las mismas placas que ya se preparan en el laboratorio, sin análisis adicionales. El estudio se hizo en China y midió diagnósticos, no vidas salvadas.

Hay un tipo de cáncer de cuello uterino que engaña a quien lo mira al microscopio. Se llama adenocarcinoma gástrico de endocérvix, o GAS por sus siglas en inglés, y comparte rasgos con tejidos benignos y con otros tumores de la misma zona. El estudio que le contamos reunió 309 placas de 96 casos confirmados de GAS, junto con 1,320 placas de otras cuatro categorías parecidas, en el Hospital Tercero de la Universidad de Pekín, entre enero de 2018 y enero de 2025. Sobre ese material se entrenó y se probó GASPath, un programa que solo mira imágenes de placas teñidas con la técnica habitual de hematoxilina y eosina¹.

En la prueba interna, el programa acertó el diagnóstico en el 98 por ciento de los casos². Cuando se lo llevó a 12 hospitales distintos, detectó el 90 por ciento de los casos de GAS, y con ajustes en el umbral de decisión llegó al 97 por ciento³. En muestras de biopsia, esas tomas pequeñas que tanto cuesta interpretar, logró una puntuación de 0,990 sobre 1⁴. Y en el despliegue grande, con 7,056 placas recogidas durante un año de consultas reales, identificó los 45 casos de GAS que había entre ellas: 45 de 45, sin excepción⁵.

Para entender por qué esto importa, hay que ver la magnitud del problema. El GAS es un subtipo muy agresivo: su supervivencia a cinco años es del 42 por ciento, frente al 91 por ciento del adenocarcinoma relacionado con el virus del papiloma humano⁶. Y el problema no es solo su agresividad, sino que se detecta tarde: el 59 por ciento de los casos se diagnostica en etapa II o posterior⁷. Es, además, el subtipo más común de adenocarcinoma cervical no asociado al VPH, y representa alrededor del 10 por ciento de todos los adenocarcinomas en países occidentales y hasta el 25 por ciento en poblaciones asiáticas⁸. Con la vacunación contra el VPH extendiéndose, se espera que su incidencia relativa aumente⁹.

¿Cómo funciona el programa? No es un microscopio nuevo ni una prueba de laboratorio distinta. Es un programa que recibe una placa ya escaneada, la divide en miles de fragmentos pequeños, y va evaluando cada fragmento en busca de señales sutiles: glándulas irregulares y anguladas, pérdida leve de la orientación del núcleo, atipia celular discreta¹⁰. Luego combina esas señales para dar una probabilidad de que la placa corresponda a GAS. Todo con las mismas placas que ya se preparan de rutina¹.

Lo que el estudio no midió es igual de importante. No midió si las pacientes vivieron más ni si el tratamiento cambió. Midió diagnósticos. Tampoco se probó dentro de un flujo de trabajo real: la evaluación con las 7,056 placas se hizo sobre material almacenado, no en tiempo real junto al patólogo¹¹. Y el programa se entrenó y validó solo con pacientes y escáneres de China; la generalización a otras poblaciones, etnias y equipos de escaneo queda por verificarse¹². Los propios autores anuncian que están iniciando un estudio aleatorizado multicéntrico para evaluar si el sistema influye en la rapidez del diagnóstico, en las decisiones terapéuticas y en la supervivencia¹³.

El interés de fondo es claro: si funciona, podría servir como herramienta de apoyo sobre todo en lugares con pocos recursos, donde las pruebas complementarias que hoy confirman el GAS, como la inmunohistoquímica o la hibridación del VPH, son caras y no siempre están disponibles¹⁴. En hospitales no terciarios, el sistema podría orientar a derivar a tiempo a la paciente y a pedir las pruebas de confirmación¹⁵.

Así lo leemos nosotros. Cuando una tecnología decide quién es sospechosa y quién no, los errores no se reparten por igual. Si un programa así se usara en clínicas públicas de América Latina, Estados Unidos o Canadá, los casos que se le escapen o las falsas alarmas podrían concentrarse en mujeres indígenas, afrodescendientes, migrantes o de zonas rurales, justo las que menos acceso tienen a confirmar un diagnóstico con pruebas adicionales. No sabemos si eso pasará con este programa en particular, porque el estudio no lo midió. Pero es lo que la experiencia con otras tecnologías de decisión nos lleva a esperar. Usted podría saber que nos equivocamos si, cuando llegue un programa así a su comunidad, los resultados se publican separados por grupo de mujeres y el programa no falla más en las que ya están postergadas.

Hay otra cosa que conviene tener presente. La gente no abandona el cuidado de su salud por ignorancia ni por falta de conciencia: cuando los estudios preventivos cuestan demasiado o quedan lejos, millones de personas se conforman con lo que tienen a mano. Un programa que lea placas de cuello uterino sin costo adicional en un hospital público podría hacer que más mujeres acepten hacerse la prueba y que los casos sospechosos lleguen antes a un centro donde puedan confirmarlos. Pero eso es una expectativa, no un resultado del estudio. Lo que el estudio muestra es que el programa acierta al leer las placas; no muestra que más mujeres vayan a hacerse la prueba ni que lleguen antes. Usted podría comprobarlo preguntando en su centro de salud cuántas mujeres se hacen la prueba antes y después de que llegue un programa así.

Y una tercera cosa, que a nosotros nos parece la más práctica. Un programa puede acertar en un hospital grande y quedar sin usarse en un puesto de salud rural si nadie capacita al personal, si no hay internet estable o si la alerta no llega a quien puede actuar. El beneficio medido en un estudio no se traduce solo. Lo que usted puede hacer es concreto: cuando en su clínica o en su comunidad se anuncie un programa de inteligencia artificial para detectar cáncer, pregunte quién revisa las alertas, si esa persona fue capacitada, y si hay un camino claro desde la alerta hasta la confirmación del diagnóstico. Si no lo hay, pídalo por escrito a la autoridad de salud. Y no deje de hacerse sus pruebas por esto.

También leímos otros trabajos que apuntan en la misma dirección. Uno, del que solo pudimos leer el resumen, encontró que un programa de seguimiento asistido por inteligencia artificial mejoró la adherencia y la calidad de vida de pacientes en quimioterapia, aunque sus autores piden estudios aleatorizados para confirmarlo. Otro, también leído solo en resumen, usó aprendizaje automático sobre análisis de sangre de rutina para estimar el riesgo de cáncer en personas derivadas por sospecha urgente, y encontró que priorizar al 10 por ciento de mayor riesgo reduciría a menos de la mitad el número de personas que hay que investigar para encontrar un caso. Son señales distintas, en enfermedades distintas, pero apuntan a lo mismo: la inteligencia artificial puede ayudar a ordenar la sospecha cuando el volumen de casos supera lo que el personal alcanza a revisar con la misma atención.

Volviendo al estudio que nos ocupa, hay un dato que merece subrayarse porque muestra la dificultad del problema: entre las cuatro categorías restantes, el 11.5 por ciento de las muestras de mucosa normal se clasificaron como GAS⁵. Es decir, el programa también genera falsas alarmas en tejido que no tiene cáncer. Eso no invalida el resultado: en un sistema pensado para no dejar pasar ningún caso, cierta tasa de falsas alarmas es el precio. Pero significa que una alerta del programa no es un diagnóstico. Es una señal para que el patólogo mire con más cuidado.

Lo que este estudio hace posible, para usted, es una pregunta distinta de la que solía hacer. Antes, la conversación era si la inteligencia artificial podía leer una placa. Ahora la pregunta es si el programa que lee esa placa llegó a su laboratorio, si alguien revisa las alertas, y si hay un camino desde esa alerta hasta la confirmación. Pregunte eso la próxima vez que oiga hablar de inteligencia artificial en su centro de salud. Es una pregunta que puede hacer hoy, y cuya respuesta le importa.

De dónde sale cada dato de contexto, y cuánto leímos de cada documento

Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This study therefore proposes a highly sensitive artificial intelligence (AI)–assisted diagnostic system for GAS based exclusively on H&E‐stained histopathological images."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In internal validation (Stage I), it achieved an accuracy of 0.980 (95% CI 0.977–0.983) and an ROC‐AUC of 0.995 (95% CI 0.994–0.997)."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In external validation (Stage II), the sensitivity reached 0.902 and improved to 0.968 with proposed strategies."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "For biopsy samples, GASPath achieved an ROC‐AUC of 0.990 (95% CI 0.984–0.997)."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In large‐scale real‐world deployment (Stage III, n = 7,056), GASPath achieved a balanced accuracy of 0.953, with 100% sensitivity for GAS (45/45 cases correctly identified)."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Endocervical gastric‐type adenocarcinoma (GAS) is a highly aggressive subtype of cervical adenocarcinoma, with a 5‐year survival rate markedly lower than that of HPV‐related adenocarcinoma (42% versus 91%; Figure 1A1 ) [ 1 , 2 , 3 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Diagnostic ambiguity is the central factor underlying GAS's clinical stealth (Figure 1A3 ), with 59% cases detected at stage II or later, leading to a substantial gap between its potential and actual clinical detection rate (Figure 1A4 ) [ 8 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "GAS represents the most common subtype of non‐HPV‐associated cervical adenocarcinoma and the second most common cervical adenocarcinoma overall, accounting for approximately 10% of all adenocarcinomas in Western countries and up to 25% in Asian populations [ 4 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "With the widespread adoption of HPV vaccination, the relative incidence of GAS is expected to increase (Figure 1A2 ) [ 5 , 6 , 7 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The heatmaps highlight morphological features of GAS that are easily underestimated, such as irregular, angulated glands, subtle loss of nuclear polarity, and mild cytologic atypia, which show substantial morphological overlap with other diagnostic categories."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Fourth, although Stage III involved large‐scale real‐world data, the evaluation was conducted offline rather than within prospective diagnostic workflows; the impact of pathologist‐in‐the‐loop interaction, alert fatigue, and laboratory information system integration on diagnostic behavior and outcomes remains uncharacterized."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Second, GASPath was trained and validated using multicenter cohorts from China, and its generalizability across populations, ethnicities, and scanning protocols remains to be verified."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "We are initiating a multicenter randomized study to assess GASPath's influence on diagnostic timeliness, therapeutic decision‐making, and survival outcomes."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Its implementation could enable cost‐effective, scalable AI‐assisted diagnostics, potentially transforming the early detection and management of this aggressive cancer subtype."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In non‐tertiary hospitals, it can guide timely referral and confirmatory testing, improving diagnostic accessibility and equity."

No tome esto como consejo médico profesional.

experiment · The journal of pathology. Clinical research · the paper, 1 Sep 2026 · free

Software Read 7,056 Routine Cervical Slides and Missed None of the 45 Cancers It Was Built to Catch

The test was run in China on slides already on file. The authors say it is not yet known whether the tool helps inside a working laboratory.

Short version · the longer version follows, about 6 min

Ask the weeklyAI chat

Ask me about this study: who was studied, what it found, and what it does not say.

Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.

The study at a glance
Who: women whose cervical tissue slides were read by software
How many: 7,056 slides, 45 of them this cancer
Where: China
When: March 2024 to April 2025
Kind of study: analysis of what people did
Who did it: Peking University Third Hospital and Tsinghua University
The limit that matters: All patients and scanners were in China; the real-world run was offline, not in a live lab.

GASPath against three other machine-learning approaches, tested on the same slides

GASPathagainstCLAM

GASPath's average sensitivity was higher by 5.41%

GASPathagainstTransMIL

GASPath's average sensitivity was higher by 2.17%

GASPathagainstAB-MIL

GASPath's average sensitivity was higher by 0.90%

The 100% figure applies to 45 GAS cases inside a 7,056-slide real-world cohort — not to all cervical lesions, and not to any lab outside the study.
weeklyAI's reading

How it could look · illustration generated by weeklyAI.watch, not a photograph

A cervical cancer called gastric-type adenocarcinoma is easy to miss. It hides inside the same pink-and-purple slides pathologists already make every day. The article describes a study in which 59 percent of cases are at stage II or later by the time they are found. The article describes a study in which 42 percent of these women are alive five years after diagnosis, compared with 91 percent for the HPV-related kind.

A team at Peking University Third Hospital and Tsinghua University built software called GASPath to read those slides. They trained it on 309 slides from 96 confirmed cases, plus 1,320 slides showing four conditions that look like it.

On their own stored slides, it sorted the five groups correctly 0.980 of the time. On biopsies, the small and difficult samples, it sorted the groups correctly 0.963 of the time.

Then they tested it on slides gathered from 12 other hospitals in China, where it caught a little more than 9 of every 10 cancer cases. Adjusting one cutoff raised that to almost 97 of every 100.

The real test came next. From March 2024 to April 2025, they ran GASPath over 7,056 cervical slides from their consultation files. Only 45 held this cancer, about six in a thousand. The software found all 45.

Three things limit what this means. Every patient and every scanner was in China; the authors say it is unproven in other populations and labs. The real-world run was done offline on saved slides, not inside a live diagnostic workflow.

The authors are now starting a randomized study of whether GASPath changes how fast this cancer is diagnosed and treated. That result is not in.

What this means for you

The software found every one of the 45 cancers it was built to catch, which is a real result, but it was tested on slides already on file in China, not in a working laboratory. So there is nothing to ask for yet. If this cancer enters your family's care, the pathologist is the person to ask.

Who paid: The work was funded by the National Natural Science Foundation of China, Key Clinical Projects of Peking University Third Hospital, the National Key R&D Program of China, Shenzhen Engineering Research Centre, a Government-funded Training Program for Outstanding Clinical Medicine Talents, the Tsinghua Shenzhen International Graduate School Cross-disciplinary Research and Innovation Fund, a Fuzhou University start-up grant, the Fujian Provincial Natural Science Foundation and the Tianjin Natural Science Foundation; the article declares no conflicts of interest and does not state that funders had any role in the study.

Do not take this as professional medical advice.

The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1232 words · about 6 minRead it →Close

AI Caught Every One of 45 Rare Cervical Cancers on Slides a Pathologist Already Had

A Chinese study tested software that reads ordinary tissue slides. It found all the hard-to-spot cases in a real clinic's files — and left the harder questions open.

Endocervical gastric-type adenocarcinoma — GAS — is a cervical cancer that behaves very differently from the kind most people have heard about. The article describes a study in which its five-year survival is 42%, against 91% for the HPV-related adenocarcinoma it is often confused with¹. The article describes a study in which it is the most common non-HPV cervical adenocarcinoma and the second most common cervical adenocarcinoma overall, about 10% of cervical adenocarcinomas in Western countries and up to 25% in Asian populations². And the article describes a study in which, as HPV vaccination becomes widespread, its relative share is expected to rise³. The problem is not only that it is aggressive. It is that it hides: the article describes a study in which 59% of cases are found at stage II or later⁴.

The reason it hides is on the slide. Well-differentiated GAS can look like normal glandular tissue; poorly differentiated GAS can look like an HPV-related cancer. Pathologists already have the slides. The authors built a system, GASPath, to read those same slides — H&E-stained tissue, the standard pink-and-blue stain used in every pathology lab — and nothing else⁵. It is built on a type of machine learning called multiple instance learning, which lets software learn from images labeled only at the level of the whole slide rather than from thousands of individually marked cells, and it is designed to pick up fine-grained differences in shape and structure⁶.

The evidence came in three stages. First, an internal cohort at Peking University Third Hospital: 309 slides from 96 confirmed GAS cases, plus 1,320 slides from four look-alike categories, collected January 2018 to January 2025. There, the system reached an accuracy of 0.980 and an ROC-AUC of 0.995⁷ — the first meaning it got about 98 in 100 slides right, the second a measure of how well it separates the categories. Second, external validation: 249 GAS slides from 12 independent centers, where sensitivity reached 0.902 and improved to 0.968 under the authors' adjusted thresholds⁸. On biopsy samples — the small snippets taken in clinic, where the tissue is often too scant for a confident read — the ROC-AUC was 0.990⁹. Third, the real-world test: 7,056 cervical slides from the hospital's consultation workflow between March 2024 and April 2025, of which only 45 were GAS. The system found all 45, with a balanced accuracy of 0.953¹⁰.

Two things about that third stage matter as much as the number. GAS was about 0.6% of those 7,056 slides — the natural rarity of the disease in a busy lab, not a curated test set. And the system's heatmaps pointed at the features pathologists themselves find easiest to wave past: irregular, angulated glands, subtle loss of nuclear polarity, mild cytologic atypia¹¹. Those are exactly the findings that overlap with other diagnoses, which is why the cancer is missed.

The comparison the authors ran is worth understanding, because it tells you what the software is and is not. They tested GASPath against three established machine-learning approaches — CLAM, TransMIL, and AB-MIL — under the same data split and protocol. GASPath had the highest average sensitivity externally, 0.902, ahead of CLAM by 5.41%, TransMIL by 2.17%, and AB-MIL by 0.90%, and ranked first in 8 of 12 centers.

Now the limits, which the authors state plainly. The system was trained and validated on cohorts from China; its generalizability across populations, ethnicities, and scanning protocols remains to be verified¹². The real-world stage was run offline on stored slides, not inside a live diagnostic workflow, and the effect of a pathologist working alongside it — including alert fatigue and integration with lab systems — is uncharacterized¹³. Among the non-GAS categories, 11.5% of normal samples were flagged as GAS. And both the internal and external cohorts were assembled retrospectively, with consultation cases that may skew toward difficult or ambiguous slides rather than routine screening.

A route to the reader's own clinic does exist in the authors' framing: in hospitals without specialist pathology, they suggest, a tool like this could guide timely referral and confirmatory testing, improving access and equity¹⁴. That is a proposal, not a result. The authors are now initiating a multicenter randomized study to test whether the system affects how quickly diagnoses are made, what treatment decisions follow, and survival¹⁵. Until that reports, the honest summary is that this software reads slides well. Whether patients do better because of it is unproven.

Here is how we read it. This belongs to a family of stories we keep returning to: software trained on records that already exist, pointed at a problem where the record was always there and the eyes were the bottleneck. The pattern to watch is not whether the accuracy number is high. It is who supplied the slides, and whether anyone tested the tool on patients who look nothing like them. When a health service near you announces an AI screening tool, the question to ask is not "how accurate is it?" but "which hospitals did the test slides come from, and were they checked on patients like the ones in this waiting room?" We also read, in other work on tools of this kind, a standing worry: when a confident machine disagrees with a person's own eyes, the person often defers. You would know we were wrong about this if clinics using such tools routinely logged and reviewed the cases where staff overruled the software — and published them alongside the successes.

One more thing we would want to see, and it is the thing that decides whether any of this reaches anyone: who keeps the record. A tool's alerts only change care if they land in front of the person who decides, if that person can see why the case was flagged, and if the misses are written down somewhere. So the practical question for a reader is not technical at all. Ask whether the clinic can show you the reasoning behind a flag, and whether the cases it got wrong are ever counted.

What the reader can hold onto is narrower and more useful than the headline number. The 100% figure applies to 45 GAS cases inside a 7,056-slide real-world cohort — not to all cervical lesions, and not to any lab outside the study. The diagnoses in this study were confirmed with immunohistochemistry and HPV in situ hybridization, the tests pathologists already order when something looks wrong. A flag from software like this is a prompt for a closer look, never a diagnosis, and nothing here says anyone should skip or delay the tests their clinician orders.

If the prospective study now being set up shows what the retrospective data suggest, the consequence is concrete: a woman whose biopsy looks unremarkable at a small clinic gets referred sooner, and a cancer that usually announces itself at stage II or later is caught while it is still only a pattern on a slide. That is the possibility. The next time you hear that a hospital is adopting software to read slides, ask which patients it was tested on — and whether anyone is keeping count of the ones it got wrong.

Where each piece of context comes from, and how much of it we read

Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "Endocervical gastric‐type adenocarcinoma (GAS) is a highly aggressive subtype of cervical adenocarcinoma, with a 5‐year survival rate markedly lower than that of HPV‐related adenocarcinoma (42% versus 91%; Figure 1A1 ) [ 1 , 2 , 3 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "GAS represents the most common subtype of non‐HPV‐associated cervical adenocarcinoma and the second most common cervical adenocarcinoma overall, accounting for approximately 10% of all adenocarcinomas in Western countries and up to 25% in Asian populations [ 4 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "With the widespread adoption of HPV vaccination, the relative incidence of GAS is expected to increase (Figure 1A2 ) [ 5 , 6 , 7 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "Diagnostic ambiguity is the central factor underlying GAS's clinical stealth (Figure 1A3 ), with 59% cases detected at stage II or later, leading to a substantial gap between its potential and actual clinical detection rate (Figure 1A4 ) [ 8 ]."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "This study therefore proposes a highly sensitive artificial intelligence (AI)–assisted diagnostic system for GAS based exclusively on H&E‐stained histopathological images."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "We developed GASPath, based on a novel multiple instance learning framework that efficiently captures fine‐grained morphological variations from H&E‐stained images."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "In internal validation (Stage I), it achieved an accuracy of 0.980 (95% CI 0.977–0.983) and an ROC‐AUC of 0.995 (95% CI 0.994–0.997)."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "In external validation (Stage II), the sensitivity reached 0.902 and improved to 0.968 with proposed strategies."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "For biopsy samples, GASPath achieved an ROC‐AUC of 0.990 (95% CI 0.984–0.997)."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "In large‐scale real‐world deployment (Stage III, n = 7,056), GASPath achieved a balanced accuracy of 0.953, with 100% sensitivity for GAS (45/45 cases correctly identified)."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "The heatmaps highlight morphological features of GAS that are easily underestimated, such as irregular, angulated glands, subtle loss of nuclear polarity, and mild cytologic atypia, which show substantial morphological overlap with other diagnostic categories."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "Second, GASPath was trained and validated using multicenter cohorts from China, and its generalizability across populations, ethnicities, and scanning protocols remains to be verified."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "Fourth, although Stage III involved large‐scale real‐world data, the evaluation was conducted offline rather than within prospective diagnostic workflows; the impact of pathologist‐in‐the‐loop interaction, alert fatigue, and laboratory information system integration on diagnostic behavior and outcomes remains uncharacterized."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "In non‐tertiary hospitals, it can guide timely referral and confirmatory testing, improving diagnostic accessibility and equity."
Yang, J., He, Q., Peng, J. et al. (2026). Artificial intelligence‐assisted histopathological diagnosis of endocervical gastric‐type adenocarcinoma: a multicenter model development and validation study. The Journal of Pathology: Clinical Research. https://doi.org/10.1002/2056-4538.70113 - the article this story is about — the whole article — the passage: "We are initiating a multicenter randomized study to assess GASPath's influence on diagnostic timeliness, therapeutic decision‐making, and survival outcomes."

Do not take this as professional medical advice.

The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.