An AI Caught Every One of 45 Cervical Cancers That Pathologists Often Miss

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

Una IA encontró los 45 casos de un cáncer de cuello uterino que los patólogos suelen pasar por alto

El estudio se hizo en China, con pacientes y equipos chinos, y la prueba en el mundo real fue fuera de línea: nadie ha medido todavía si cambia un diagnóstico o la suerte de una paciente.

Versión breve · la versión detallada sigue, unos 8 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: Muestras de tejido de cuello uterino, incluidas 45 de un cáncer agresivo llamado GAS
Cuántos: 7,056 muestras
Dónde: China
Cuándo: entre marzo de 2024 y abril de 2025
Tipo de estudio: analysis of what people did
Quién lo hizo: Hospitales y universidades de China
El límite que importa: La prueba real fue fuera de línea y no se sabe si cambia el diagnóstico de una paciente

Supervivencia a cinco años: cáncer gástrico de cuello uterino frente al relacionado con el virus del papiloma

Cáncer gástrico de cuello uterinofrente aCáncer relacionado con el virus del papiloma humano

42 de cada 100 mujeres siguen vivas a los cinco años, frente a 91 de cada 100 en el otro tipo

Una alerta es una señal para que un patólogo mire con más atención, no un diagnóstico.
Lectura de weeklyAI

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

El cáncer de cuello uterino del tipo gástrico es muy agresivo. El artículo describe un estudio en el que unas 42 de cada 100 mujeres siguen vivas cinco años después del diagnóstico. El artículo describe un estudio en el que, en el tipo relacionado con el virus del papiloma humano, unas 91 de cada 100 mujeres siguen vivas cinco años después. Y el artículo describe un estudio en el que, en 59 de cada 100 casos, el cáncer gástrico se descubre cuando ya está en etapa II o más avanzada.

Investigadores de varios hospitales y universidades de China construyeron un programa de inteligencia artificial llamado GASPath. El programa lee las mismas láminas teñidas con hematoxilina y eosina que usa cualquier laboratorio de patología. Lo entrenaron con 309 láminas de 96 casos confirmados de cáncer gástrico, el grupo más grande reunido hasta ahora para este tipo de estudio. También usaron 1,320 láminas de otras cuatro clases de tejido parecido, para que el programa aprendiera a distinguirlas.

Después vino la prueba difícil. En consultas de rutina juntaron 7,056 láminas de cuello uterino, tomadas entre marzo de 2024 y abril de 2025. De todas ellas, sólo 45 tenían cáncer gástrico, el 0.6%. El programa encontró los 45 casos. También marcó como sospechoso el 11.5% del tejido normal.

Lo hizo sin las pruebas complementarias que suelen acompañar este diagnóstico, como la inmunohistoquímica, la hibridación del virus del papiloma o los estudios de genes. Sólo usó la lámina de rutina.

Ahora las limitaciones, que importan. El estudio se hizo en China, con pacientes y escáneres chinos. No se sabe cómo funcionaría con otras poblaciones ni con otros equipos. La prueba en el mundo real se hizo fuera de línea, no dentro de un flujo de trabajo clínico, así que nadie ha medido si el aviso llega a tiempo o si cambia lo que hace el patólogo. Y el programa también marcó como sospechoso el 11.5% del tejido normal. Una alerta es una señal para que un patólogo mire con más atención, no un diagnóstico.

Lo que puede buscar donde usted vive: si su hospital o su clínica empiezan a usar lectura de láminas de biopsia de cuello uterino asistida por inteligencia artificial, y si esa alerta llega a manos de un patólogo a tiempo.

Qué significa para usted

Si en su hospital o su clínica empiezan a usar lectura de láminas de biopsia de cuello uterino asistida por inteligencia artificial, pregunte si esa alerta llega a manos de un patólogo. El programa marcó como sospechoso el 11.5% del tejido normal, así que una alerta es una señal para mirar con más atención, no un diagnóstico. Nadie ha medido todavía si el aviso cambia lo que hace el patólogo.

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 estudio 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, el Programa Gubernamental de Formación para Talentos Clínicos Sobresalientes, el Fondo de Investigación e Innovación Interdisciplinaria de la Escuela de Posgrado Internacional Tsinghua Shenzhen, la Subvención de Inicio para Talentos Introducidos de la Universidad de Fuzhou, la Fundación Provincial de Ciencias Naturales de Fujian y el Proyecto de la Fundación de Ciencias Naturales de Tianjin; el artículo no indica si los financiadores tuvieron alguna influencia, y no se declararon conflictos de interés.

No tome esto como consejo médico profesional.

Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1682 palabras · unos 8 minLeerla →Cerrar

La IA que encontró un cáncer de cuello uterino que los patólogos suelen pasar por alto

Un estudio chino probó un sistema que lee placas de laboratorio de rutina y detectó los 45 casos reales de un subtipo agresivo. Todavía nadie ha medido si eso cambia el desenlace de una paciente.

Hay un tipo de cáncer de cuello uterino que se esconde. Se llama adenocarcinoma gástrico de tipo endocervical, o GAS por sus siglas en inglés. La revisión recoge un estudio en el que su tasa de supervivencia a cinco años es muy inferior a la del adenocarcinoma asociado al virus del papiloma humano: 42% frente a 91%¹. La revisión recoge un estudio en el que la ambigüedad diagnóstica es el factor central detrás de ese sigilo clínico, y el 59% de los casos se detecta en etapa II o más tarde².

Un equipo de investigación en China desarrolló un sistema de inteligencia artificial llamado GASPath, construido sobre un marco de aprendizaje por instancias múltiples que captura variaciones morfológicas finas en imágenes teñidas con hematoxilina y eosina³. En su despliegue a gran escala en el mundo real, con 7,056 muestras, GASPath alcanzó una precisión equilibrada de 0,953 y una sensibilidad del 100% para GAS: identificó correctamente los 45 casos⁴. En todos los umbrales probados, encontró los 45 casos entre las 7,056 muestras⁵. En muestras de biopsia, alcanzó un ROC-AUC de 0,990⁶. En la validación externa, la sensibilidad llegó a 0,902 y subió a 0,968 con las estrategias propuestas⁷.

Los mapas de atención del sistema resaltan rasgos morfológicos de GAS que se subestiman con facilidad, como glándulas irregulares y anguladas, pérdida sutil de la polaridad nuclear y atipia citológica leve, que se superponen sustancialmente con otras categorías diagnósticas⁸. El sistema permite detección de alta sensibilidad en placas de rutina, sin necesidad de pruebas auxiliares extensas⁹.

¿Cómo funciona, en palabras llanas? El sistema no mira la placa entera de una vez. La divide en pequeños cuadros, extrae de cada uno un resumen numérico de lo que ve y luego pesa cuáles de esos cuadros importan más para la decisión final³. Los cuadros más importantes quedan marcados en un mapa de calor sobre la imagen original, de modo que un patólogo puede ver dónde se fijó la máquina⁸. Todo esto ocurre sobre placas teñidas con hematoxilina y eosina, el tinte corriente de cualquier laboratorio de patología, sin necesidad de pruebas auxiliares extensas⁹.

El estudio comparó GASPath con otros tres sistemas de inteligencia artificial ya establecidos: CLAM, TransMIL y AB-MIL⁷. En la validación externa, GASPath alcanzó la mayor sensibilidad media, con 0,902, superando a CLAM, TransMIL y AB-MIL por 5.41%, 2.17% y 0.90% respectivamente, y quedó primero en 8 de 12 centros⁷. La ventaja se mantuvo en el despliegue real: GASPath logró la mayor precisión equilibrada, 0,953, superando al segundo mejor, CLAM, por 3.3%⁴.

Vale la pena detenerse en lo que esto significaría para una familia. Si el sistema funciona como sugiere este estudio, una alerta temprana en una biopsia podría llevar a una histerectomía o a una linfadenectomía a tiempo, en lugar de esperar a que el cáncer se haya extendido². En hospitales de tercer nivel, el sistema podría aumentar la eficiencia diagnóstica y reducir la carga de trabajo mediante el marcado confiable de casos y el apoyo de telepatología; en hospitales que no son de tercer nivel, podría guiar la derivación oportuna y las pruebas de confirmación, mejorando la accesibilidad y la equidad diagnóstica¹⁰. Predicciones positivas también podrían fomentar pruebas moleculares adicionales, como la amplificación de HER2 o la expresión de Claudin18.2, que orientan terapias individualizadas con agentes como el trastuzumab o el zolbetuximab¹¹.

Pero hay límites que el lector debe conocer. El sistema fue entrenado y validado con cohortes multicéntricas de China, y su generalización a otras poblaciones, etnias y protocolos de escaneo aún no se ha verificado¹². Aunque la etapa III involucró datos del mundo real a gran escala, la evaluación se realizó fuera de línea, no dentro de flujos de trabajo diagnósticos prospectivos; el impacto de la interacción con el patólogo, la fatiga de alertas y la integración con el sistema de información del laboratorio sigue sin caracterizarse¹³. En el cohorte real, el 11.5% de las muestras de mucosa cervical normal se clasificaron erróneamente como GAS, lo que significa que una alerta es una señal para que un patólogo mire con más atención, no un diagnóstico⁴.

Así lo leemos nosotros. La promesa de esta herramienta es grande, pero conviene mirarla con la misma cautela con que miraríamos cualquier tecnología que decide sobre la vida de las personas. El patrón que vemos es este: cuando una máquina se presenta como limpia, incansable e imparcial, las instituciones tienden a delegarle no solo lo pesado y repetitivo sino también lo delicado, como si su falta de cansancio garantizara que no arrastra los sesgos de quien la diseñó. Y los errores de esas máquinas no se reparten por igual: recaen sobre quienes ya cargan con desventajas, y casi nunca se sabe cómo fue entrenada ni quién responde por la equivocación.

¿Qué esperamos, entonces? Esperamos que un sistema que acierta en todos los casos de una enfermedad rara en un estudio retrospectivo sea presentado como listo para usarse en clínicas y hospitales sin una evaluación prospectiva que mida su efecto real en el trabajo diario, en los costos y en los diagnósticos que se adelantan o se retrasan. También esperamos que, en lugares donde los servicios de salud ya están saturados, la herramienta se quede en los hospitales grandes y no llegue a la clínica pública donde la mayoría se atiende. Nos equivocaríamos si el sistema mantiene su sensibilidad en cohortes independientes de otros países, con equipos y protocolos distintos, y si existen reglas claras de responsabilidad y revisión humana cuando la alerta falla. También nos equivocaríamos si vemos que la herramienta llega a la red pública sin costo adicional para la paciente y con un patólogo que confirme cada resultado.

¿Qué puede hacer usted con esto? Puede preguntar en su centro de salud si la herramienta que se usa fue probada con personas parecidas a usted y si hay un patólogo que revise cada alerta. Puede pedir que le expliquen quién responde si el sistema falla. Puede averiguar si en su región existe algún programa que use esta clase de herramientas en la red pública y, si no existe, preguntar a las autoridades de salud si hay planes de llevarla a los centros donde usted se atiende. Y puede preguntar si la herramienta que le ofrecen ya pasó por un estudio con pacientes reales y no solo con diapositivas guardadas.

Hay un detalle que ayuda a dimensionar el problema. La revisión recoge un estudio en el que el adenocarcinoma gástrico de tipo endocervical representa el subtipo más común de adenocarcinoma cervical no asociado al VPH y el segundo adenocarcinoma cervical más común en general, con aproximadamente el 10% de todos los adenocarcinomas en países occidentales y hasta el 25% en poblaciones asiáticas². La revisión recoge un estudio en el que, con la adopción generalizada de la vacuna contra el VPH, se espera que la incidencia relativa de GAS aumente². Es decir, el problema que este sistema intenta resolver no va a desaparecer; probablemente crezca.

Este estudio se suma a una familia de investigaciones que buscan usar la inteligencia artificial para detectar enfermedades antes de que sea tarde. Otros trabajos que leímos, y de los que solo pudimos leer el resumen porque el texto completo está detrás de una suscripción, exploran caminos parecidos. Un estudio evaluó pruebas de sangre con modelos de aprendizaje automático para estimar el riesgo de cáncer en adultos derivados por sospecha urgente en el sistema de salud británico; cinco pruebas mostraron posible utilidad clínica, con puntuaciones que iban de 0,86 para cáncer gastrointestinal superior a 0,72 para cáncer colorrectal bajo¹⁴. Otro trabajo probó una herramienta de imagen aprobada por la FDA para el diagnóstico de enfermedad pulmonar intersticial en 209 pacientes de 47 centros médicos, y encontró que el porcentaje de pacientes que recibieron terapia farmacológica específica aumentó significativamente tras su uso¹⁵. Un tercer estudio, un análisis interino de un ensayo de implementación con 200 pacientes, evaluó el impacto de un biomarcador de inteligencia artificial en decisiones sobre terapia hormonal en cáncer de próstata de riesgo intermedio, y encontró que las pruebas del biomarcador alteraron el 27.5% de las decisiones finales¹⁶.

Lo que estos trabajos comparten es una promesa y una advertencia. La promesa: la inteligencia artificial puede ver patrones que a simple vista se escapan, y puede hacerlo sobre materiales que ya existen en cualquier laboratorio. La advertencia: casi todos estos estudios miden si la herramienta acierta, no si la paciente vive más o mejor. Para saber eso hace falta un ensayo prospectivo, con pacientes reales y patólogos en el circuito, que mida si la herramienta cambia el tiempo hasta el diagnóstico, el número de casos omitidos y el gasto, y que documente los errores que comete en ese uso cotidiano.

Los autores del estudio lo dicen con claridad: su implementación podría permitir diagnósticos asistidos por inteligencia artificial rentables y escalables, con el potencial de transformar la detección temprana y el manejo de este subtipo agresivo de cáncer¹⁷. Pero también reconocen que el impacto clínico real —en la eficiencia del flujo de trabajo, la rentabilidad y los resultados para los pacientes— requiere cuantificación prospectiva¹³. Están iniciando un estudio aleatorizado multicéntrico para evaluar la influencia de GASPath en la oportunidad del diagnóstico, la toma de decisiones terapéuticas y los resultados de supervivencia¹³.

Aquí está lo que usted puede llevarse. La próxima vez que le hagan una citología o una biopsia de cuello uterino, pregunte si el laboratorio usa alguna herramienta de inteligencia artificial para revisar las placas y, si la usa, pregunte quién revisa las alertas. No para desconfiar del patólogo, sino para entender el circuito completo: quién mira, quién confirma y quién responde si algo se pasa por alto. La tecnología solo importa si llega al lugar donde usted se atiende, sin costo adicional y con una persona capacitada del otro lado. Esa es la pregunta que vale la pena hacer.

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: "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: "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 - 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: "Across all thresholds, it successfully identified all 45 GAS cases among 7,056 samples, achieving 100% sensitivity (Figure 5A–D )."
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 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: "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: "GASPath enables high‐sensitivity detection of GAS in routine H&E‐stained slides, obviating the need for extensive auxiliary testing while preventing underdiagnosis and misdiagnosis."
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 tertiary hospitals, it enhances diagnostic efficiency and reduces workload through reliable case flagging and telepathology support. 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 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Positive predictions from GASPath may also encourage ancillary molecular testing (e.g. HER2 amplification, Claudin18.2 expression), aiding individualized therapy with agents such as trastuzumab or zolbetuximab."
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: "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."
Neal M, Dean M, Duffy S, Ferguson RE, Horan P, Johnston C, et al. (2026). Real-World Validation of PinPoint Blood Tests in the NHS: Multivariable Machine Learning to Predict Cancer Risk in Primary Care Urgent Referrals. Mayo Clinic Proceedings: Digital Health. 10.1016/j.mcpdig.2026.100382 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Five tests have performance indicating potential clinical utility. Receiver operating characteristic area-under-curve scores (95% CI) for these were: Upper gastrointestinal=0.86 (0.81-0.90), Gynecological=0.81 (0.77-0.85), Lung=0.79 (0.74-0.84), Head & Neck=0.73 (0.68-0.78), and Lower gastrointestinal=0.72 (0.67-0.78),"
Tambe AP, Boente RD, George G, Kheir F, Tahamtani Omran O, Selvan KC. (2026). Clinical Utility of an FDA-Authorized Artificial Intelligence Imaging Platform in Interstitial Lung Disease Diagnosis. Diagnostics. 10.3390/diagnostics16152445 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The percentage of patients receiving ILD-specific pharmacotherapy increased significantly following Fibresolve utilization (12.5% to 44.4%, p p"
Wegener E, Ng M, Guerrieri M, Showalter TN, de Leon J, Ramani S, et al. (2026). Artificial Intelligence impact on decision-making regarding short-term ADT with prostate radiotherapy: ASTuTE clinical trial interim results. Clinical and Translational Radiation Oncology. 10.1016/j.ctro.2026.101257 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Biomarker testing altered 27.5% (CI: 21.4-34.2%) of final decisions."
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."

No tome esto como consejo médico profesional.

Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.

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

An AI Caught Every One of 45 Cervical Cancers That Pathologists Often Miss

In a test on 7,056 slides from one Chinese hospital, the software found all 45 cancers of a hard-to-spot type. The test was done offline, and it flagged some healthy tissue too.

Short version · the longer version follows, about 7 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 with a hard-to-spot cervical cancer (gastric-type adenocarcinoma) and look-alike tissue samples
How many: 7,056 slides, including 45 cancer cases
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, with 12 other Chinese hospitals
The limit that matters: Tested offline in China only; no proof yet it changes diagnosis or survival.

GASPath against three other computer programs of its kind

GASPathagainstCLAM, TransMIL, and AB-MIL

GASPath came out ahead on the measures they used.

An alert from a tool like this means a pathologist should look closer. It does not mean cancer is there.
weeklyAI's reading

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

The article describes a study in which 42 women in a hundred are alive five years after a diagnosis of gastric-type adenocarcinoma of the cervix, against 91 in a hundred for the HPV-related kind. It is one of the deadliest cervical cancers. The article describes a study in which 59 out of 100 cases are found at stage II or later.

Researchers at Peking University Third Hospital and Tsinghua University in China built a computer program called GASPath to catch it. The program reads the ordinary pink-and-purple slides that labs already make from cervical tissue, with no extra tests. It learned from 309 slides belonging to 96 patients with this cancer, the largest set gathered for this purpose so far, plus slides from four other tissue types that look similar under a microscope.

Then came the real test. The team ran GASPath over 7,056 slides from routine consultations at one hospital, collected between March 2024 and April 2025. This cancer made up six cases in a thousand. The program flagged all 45 of the cancer cases. Nothing was missed.

The article describes a study in which the program managed this without extra tests such as staining tests or gene sequencing, which are expensive and often unavailable. It used only the slide a lab prepares anyway.

Two things hold the result in place. The work was done in China, on Chinese patients and Chinese scanners, so how well it would do in other countries is not yet known. And the test was run offline, on slides pulled from the record. No one has measured yet whether the program changes a diagnosis or helps a woman live longer.

The program also raised false alarms. It misclassified 11.5 percent of normal tissue samples as cancer. The researchers write that this fits the tissue's appearance, which can look like the cancer. An alert from a tool like this means a pathologist should look closer. It does not mean cancer is there.

The same team compared GASPath with three other computer programs of its kind. GASPath came out ahead on the measures they used.

What to look for: whether your hospital or clinic starts using software to read cervical biopsy slides, and whether a suspicious result reaches a pathologist quickly.

What this means for you

If your clinic starts using software to read cervical slides, know that in one Chinese study it caught all 45 cancers of this hard-to-spot type, though it also flagged some healthy tissue. No one has measured yet whether such a tool helps a woman live longer, and it has not been tested where you live. Ask whether a suspicious result reaches a pathologist quickly.

Who paid: The study 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, Government-funded Training Program for Outstanding Clinical Medicine Talents, Tsinghua Shenzhen International Graduate School Cross-disciplinary Research and Innovation Fund, Fuzhou University Research Start-up Grant, Fujian Provincial Natural Science Foundation, and Tianjin Natural Science Foundation; the article does not state whether funders had any say, and no conflicts of interest were declared.

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 · 1430 words · about 7 minRead it →Close

AI Caught Every One of 45 Cervical Cancers on Routine Slides — Inside One Chinese Hospital System

A computer program read ordinary tissue slides and flagged a hard-to-spot cancer that pathologists often miss. The test was real and large. It was also offline, and all of it happened in China.

A woman's cervical cancer that pathologists frequently overlook was identified in every one of 45 real cases by an artificial intelligence system that read only routine tissue slides, according to a study published in the *Journal of Pathology: Clinical Research*. The system, called GASPath, was built by a team led by researchers at Peking University Third Hospital and Tsinghua University, with collaborators at a dozen other Chinese hospitals. In a run of 7,056 slides taken from the hospital's ordinary consultation workflow, it caught all 45 cases of endocervical gastric-type adenocarcinoma, or GAS¹². On the same slides it reached a balanced accuracy of 0.953 — a measure that weighs how well it did on each kind of slide, not just the common ones¹.

GAS is not the cervical cancer most people have heard of. It is a highly aggressive subtype, and its five-year survival is markedly lower than that of the HPV-related kind — 42 percent versus 91 percent³. It is also the most silent. Because its cells look deceptively ordinary under a microscope, 59 percent of cases are found at stage II or later⁴. The study's own numbers show how badly the deck is stacked: in the real-world run, GAS made up only about 0.6 percent of all cases. A pathologist reading slides all day is looking for a needle in a very large haystack.

This was the largest GAS group ever assembled for AI research — 309 slides from 96 confirmed cases, collected from January 2018 to January 2025⁵. Around them the researchers placed four look-alike conditions, 1,320 slides in all, so the model would have to tell GAS apart from normal cervix, benign glandular lesions, HPV-related adenocarcinoma, and a mucinous form of endometrioid cancer. In that first internal test it reached an accuracy of 0.980 and an ROC-AUC of 0.995. Across 12 outside hospitals it caught 90.2 percent of cases, rising to 96.8 percent under settings tuned to miss nothing⁶. On biopsies — the small, shallow samples taken in a clinic — it reached an ROC-AUC of 0.990⁷.

To understand why this matters, it helps to know what the system actually does. It is not a chatbot and it does not reason. It is a pattern-recognition program trained on pictures. Each slide is cut into small squares, and a large pretrained model turns each square into a set of numbers describing what it sees⁵. Then a second stage — the study's real contribution — weighs those squares against one another and decides what the slide as a whole looks like. The authors describe it as resembling a panel of subspecialists: several "expert agents" each attend to a different feature, while connections between them let each revise its view based on what its neighbors found. The system then produces both a diagnosis and a heatmap showing which parts of the slide pushed it toward that answer⁸.

The heatmaps are worth pausing over, because they show what the model is looking at. They point to features the study calls easily underestimated: irregular, angulated glands, a subtle loss of nuclear polarity, mild cytologic atypia⁸. These are exactly the features that overlap with other diagnostic categories — which is precisely why they get missed. The authors' claim is that GASPath detects the cancer on routine slides without extensive auxiliary testing⁹. That matters because the confirmatory tools — immunohistochemistry, HPV RNA in situ hybridization, next-generation sequencing — are expensive, labor-intensive and often unavailable in low-resource settings, and even then they are not always conclusive.

This study belongs to a family of attempts to put AI where a human is currently making a hard call. One recent real-world evaluation in England tested machine-learning blood tests on 16,481 patients referred urgently for suspected cancer across five hospital trusts and 170 family-doctor surgeries — the summary reports that five of the tests showed performance indicating potential clinical utility, and that prioritizing the highest-risk tenth of patients would cut the number who need investigating to find one cancer by a factor of 2.6 to 6.1¹⁰¹¹¹²¹³¹⁴. We could read only the summary of that work; the full paper is behind a subscription. Another, also summary-only, followed 209 patients across 47 medical centers using an FDA-approved imaging tool for interstitial lung disease, and reported that previously planned invasive diagnostic procedures were avoided in many cases¹⁵¹⁶¹⁷. A third, again summary-only, reported that an AI biomarker changed 27.5 percent of treatment decisions in intermediate-risk prostate cancer¹⁸¹⁹.

The limits here are not small, and the authors state them plainly. Both the internal and external groups were assembled after the fact, and because some came through a consultation service, they may skew toward the difficult and ambiguous cases rather than the ordinary screening population²⁰. The system was trained and validated on Chinese patients, Chinese scanners and Chinese staining, and whether it carries over to other populations, ethnicities and scanning protocols remains to be verified²⁰. Most importantly, the large real-world run was conducted offline, not inside a live diagnostic workflow — so the effect on pathologist behavior, on alert fatigue, and on laboratory computer systems is uncharacterized²¹. The authors also note that specificity could be improved by adding clinical, radiologic or immunohistochemical features. And the model was more likely to flag normal tissue as suspicious: 11.5 percent of normal samples were misclassified as GAS²¹.

There is a trade-off buried in that last number, and it is worth naming. The system is deliberately tuned toward catching everything. A higher threshold raised specificity to 0.956 and 0.964, but the design choice throughout is sensitivity first. In plain terms: an alert is a prompt for a pathologist to look more closely, not a diagnosis. The authors also describe where such a tool could sit — in large hospitals, flagging cases and reducing workload; in smaller ones, guiding referrals and confirmatory testing²². They suggest positive predictions could encourage molecular testing for HER2 or Claudin18.2, which in turn could point toward targeted drugs²³. One stated hope is integration into China's national dual-cancer screening programs, alongside HPV testing and cytology²⁴.

Here is how we read it. The pattern in this family of stories is that a tool gets very good at finding something rare in the place where it was built, and the question that decides whether it helps anyone else is not accuracy but portability. We expect that in homes like the reader's, the first version of a system like this will arrive as a pilot in a large teaching hospital, not in the clinic down the road. We could be wrong: the way to tell would be slides from hospitals that never contributed training data, prepared with different stains and scanned on different machines, still being caught at the same rate — and heatmaps that keep pointing at tumor glands rather than at background or staining artifacts.

There is a second thing we would watch. If a tool like this is adopted in a public clinic and it flags a slide wrongly, or misses one, the patient and the local pathologist need a clear route to challenge it, understand why, and get it corrected. We would want to know who is accountable when an AI-assisted diagnosis is wrong, and whether a pathologist can override it — and have that override recorded. If no one at your clinic can answer that, it is worth raising before the tool arrives, not after.

What comes next is a prospective test. The authors say a multicenter randomized study is being initiated to assess the system's influence on diagnostic timeliness, treatment decisions and survival outcomes. Until that reports, no one can say that this tool saves lives — only that it found the right slides.

And the last thing we would ask, for a reader in Latin America or in the United States or Canada: ask whether a tool like this is reaching the places where women currently go undiagnosed — rural clinics, public screening programs — or only the large hospitals that already have pathologists. If it is only the latter, ask what would be needed to bring it to the former. That is the question this study makes it possible to ask, and it is a better one than whether the computer beat the doctor.

When you next hear that an AI caught every case, ask where the slides came from.

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: "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: "Across all thresholds, it successfully identified all 45 GAS cases among 7,056 samples, achieving 100% sensitivity (Figure 5A–D )."
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: "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: "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 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: "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: "GASPath enables high‐sensitivity detection of GAS in routine H&E‐stained slides, obviating the need for extensive auxiliary testing while preventing underdiagnosis and misdiagnosis."
Neal M, Dean M, Duffy S, Ferguson RE, Horan P, Johnston C, et al. (2026). Real-World Validation of PinPoint Blood Tests in the NHS: Multivariable Machine Learning to Predict Cancer Risk in Primary Care Urgent Referrals. Mayo Clinic Proceedings: Digital Health. 10.1016/j.mcpdig.2026.100382 — only the abstract - the full text could not be fetched — the passage: "Objective To validate the United Kingdom Conformity Assessed-marked PinPoint blood tests, which use machine learning models and routinely available blood analytes to estimate cancer risk in adults referred on urgent suspected cancer pathways in National Health Service (NHS) England."
Neal M, Dean M, Duffy S, Ferguson RE, Horan P, Johnston C, et al. (2026). Real-World Validation of PinPoint Blood Tests in the NHS: Multivariable Machine Learning to Predict Cancer Risk in Primary Care Urgent Referrals. Mayo Clinic Proceedings: Digital Health. 10.1016/j.mcpdig.2026.100382 — only the abstract - the full text could not be fetched — the passage: "This work comprises a large-scale, prospective, observational, real-world NHS service evaluation of 9 blood tests, carried out from December 21, 2020 to July 31, 2025."
Neal M, Dean M, Duffy S, Ferguson RE, Horan P, Johnston C, et al. (2026). Real-World Validation of PinPoint Blood Tests in the NHS: Multivariable Machine Learning to Predict Cancer Risk in Primary Care Urgent Referrals. Mayo Clinic Proceedings: Digital Health. 10.1016/j.mcpdig.2026.100382 — only the abstract - the full text could not be fetched — the passage: "Total of 16,481 patients with urgent suspected cancer referrals were enrolled across 5 secondary care Trusts and 170 General Practitioner surgeries."
Neal M, Dean M, Duffy S, Ferguson RE, Horan P, Johnston C, et al. (2026). Real-World Validation of PinPoint Blood Tests in the NHS: Multivariable Machine Learning to Predict Cancer Risk in Primary Care Urgent Referrals. Mayo Clinic Proceedings: Digital Health. 10.1016/j.mcpdig.2026.100382 — only the abstract - the full text could not be fetched — the passage: "Five tests have performance indicating potential clinical utility. Receiver operating characteristic area-under-curve scores (95% CI) for these were: Upper gastrointestinal=0.86 (0.81-0.90), Gynecological=0.81 (0.77-0.85), Lung=0.79 (0.74-0.84), Head & Neck=0.73 (0.68-0.78), and Lower gastrointestinal=0.72 (0.67-0.78),"
Neal M, Dean M, Duffy S, Ferguson RE, Horan P, Johnston C, et al. (2026). Real-World Validation of PinPoint Blood Tests in the NHS: Multivariable Machine Learning to Predict Cancer Risk in Primary Care Urgent Referrals. Mayo Clinic Proceedings: Digital Health. 10.1016/j.mcpdig.2026.100382 — only the abstract - the full text could not be fetched — the passage: "Prioritization of the 10% of highest-risk patients would reduce the number needed to investigate to detect one cancer by a factor of 2.6-6.1."
Tambe AP, Boente RD, George G, Kheir F, Tahamtani Omran O, Selvan KC. (2026). Clinical Utility of an FDA-Authorized Artificial Intelligence Imaging Platform in Interstitial Lung Disease Diagnosis. Diagnostics. 10.3390/diagnostics16152445 — only the abstract - the full text could not be fetched — the passage: "Fibresolve is an imaging artificial intelligence (AI) tool recently approved by the Food and Drug Administration (FDA) for use in ILD diagnosis and made available to clinicians."
Tambe AP, Boente RD, George G, Kheir F, Tahamtani Omran O, Selvan KC. (2026). Clinical Utility of an FDA-Authorized Artificial Intelligence Imaging Platform in Interstitial Lung Disease Diagnosis. Diagnostics. 10.3390/diagnostics16152445 — only the abstract - the full text could not be fetched — the passage: "There were 209 patients that underwent evaluation with Fibresolve across 47 medical centers."
Tambe AP, Boente RD, George G, Kheir F, Tahamtani Omran O, Selvan KC. (2026). Clinical Utility of an FDA-Authorized Artificial Intelligence Imaging Platform in Interstitial Lung Disease Diagnosis. Diagnostics. 10.3390/diagnostics16152445 — only the abstract - the full text could not be fetched — the passage: "previously-planned invasive diagnostic procedures were avoided in many cases"
Wegener E, Ng M, Guerrieri M, Showalter TN, de Leon J, Ramani S, et al. (2026). Artificial Intelligence impact on decision-making regarding short-term ADT with prostate radiotherapy: ASTuTE clinical trial interim results. Clinical and Translational Radiation Oncology. 10.1016/j.ctro.2026.101257 — only the abstract - the full text could not be fetched — the passage: "This interim analysis of a multicentre implementation trial ( n = 200) evaluated a multimodal artificial intelligence biomarker's impact on shared short-term androgen deprivation therapy (ST-ADT) decisions in intermediate-risk prostate cancer (IR-PC)."
Wegener E, Ng M, Guerrieri M, Showalter TN, de Leon J, Ramani S, et al. (2026). Artificial Intelligence impact on decision-making regarding short-term ADT with prostate radiotherapy: ASTuTE clinical trial interim results. Clinical and Translational Radiation Oncology. 10.1016/j.ctro.2026.101257 — only the abstract - the full text could not be fetched — the passage: "Biomarker testing altered 27.5% (CI: 21.4-34.2%) of final decisions."
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 tertiary hospitals, it enhances diagnostic efficiency and reduces workload through reliable case flagging and telepathology support. 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: "Positive predictions from GASPath may also encourage ancillary molecular testing (e.g. HER2 amplification, Claudin18.2 expression), aiding individualized therapy with agents such as trastuzumab or zolbetuximab."
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: "It may also integrate into China's national dual‐cancer screening programs, complementing HPV testing and cytology to enhance early detection of cervical adenocarcinomas, especially the underdiagnosed GAS subtype."

Do not take this as professional medical advice.