An AI Reader Passed Moscow's Mammogram Tests. The Tests Used 100 Confirmed Cases.

US · CanadaAmérica Latina

Weekly AI Watch

Every week, what changed and what it means for you. No hurry, no alarm.

Week of 30 October 2026EN | ESPast editions15,230 papers found · 14,455 read · 4,301 belonged · 46 stories passed every check

Observatorio de la
Inteligencia Artificial

Cada semana, qué cambió y qué significa para usted. Sin prisa y sin alarma.

Semana del 30 de octubre de 2026EN | ESEdiciones anteriores15,230 estudios encontrados · 14,455 leídos · 4,301 correspondían · 46 notas pasaron todas las comprobaciones

weeklyAI · Week of 30 October 2026weeklyAI · Semana del 30 de octubre de 2026

← The good news← Las buenas noticias

other · Quantitative imaging in medicine and surgery · la publicación, 10 ago 2026 · gratis

Un lector automático de mastografías pasó tres años de pruebas antes de entrar al seguro público de Moscú

Con 100 mastografías de referencia para las pruebas finas y sin revisión fuera de la red de Moscú, el sistema acertó cada vez más; nada indica que dé los mismos resultados en otra parte.

Versión breve · la versión detallada sigue, unos 6 min

Pregunte al chat de weeklyAI

Pregúnteme sobre la IA en su vida: le señalo las notas que weeklyAI ha publicado sobre el tema, esta semana y antes.

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: Mujeres de 18 años o más a las que se les hizo una mastografía en Moscú
Cuántos: 404,502 mamografías
Dónde: Moscú, Rusia
Cuándo: 2 años y 5 meses de pruebas y seguimiento
Tipo de estudio: analysis of what people did
Quién lo hizo: Centro de Diagnóstico y Telemedicina del Departamento de Salud de Moscú
El límite que importa: La prueba fina usó solo 100 mamografías y nadie lo probó fuera de Moscú.

Cómo cambiaron las medidas del sistema entre el inicio y el final del estudio

Capacidad de distinguir un caso con cáncer de uno sin cáncer10.8%

Exactitud16.9%

Detección de casos con cáncer4.8%

Descarte correcto de casos sin cáncer30%

Fallas técnicas (bajaron)26.7%

Calificación clínica de los expertos47.8%

Son los cambios porcentuales que el estudio reporta entre el inicio y el final, medidos en la red de Moscú; no son resultados de otras ciudades ni de otros países.

Lo que el estudio comparó antes y después

Primeras pruebas de calibraciónfrente aMeta fijada de 0.80

Las dos primeras no llegaron a la meta; la tercera sí

Calificación clínica al iniciofrente aCalificación clínica al final

Subió de 54.38% a 80.36%

Fallas técnicas al iniciofrente aFallas técnicas al final

Bajaron de 3.0% a 0.8%

La máquina leyó 400 mil mamografías en Moscú y mejoró sus números. Lo que no puede decirle es si a usted le va a servir.
Lectura de weeklyAI

Ver · 0:26

El análisis automático de mastografías se sumó al programa de seguro médico obligatorio de Moscú seis meses después de que empezaran las pruebas. Antes de eso, un centro de investigación de la ciudad había sometido al programa Third Opinion Mammograms a una cadena de revisiones en orden.

Primero, una prueba de funciones: un radiólogo experto revisó cuatro mastografías de mujeres de 61 años en promedio. En una, el sistema no procesó bien todas las vistas. Los desarrolladores lo corrigieron.

Vino la prueba de calibración, con 100 mastografías, la mitad con el problema que se buscaba y la otra mitad sin él. Aquí se midió qué tan bien distingue el sistema un caso de cáncer de uno sano. Las dos primeras rondas no alcanzaron la meta fijada de 0.80. Tras los arreglos de los desarrolladores, la tercera pasó. Durante el seguimiento salieron seis actualizaciones del sistema. Las que cambiaron el funcionamiento del programa exigieron repetir la calibración.

Durante dos años y un mes, el sistema revisó todas las mastografías de los sitios participantes, sin excluir ninguna: 429,027 en el seguimiento técnico. En solo uno de 25 meses las fallas de procesamiento pasaron del 10%. La puntuación clínica, que un radiólogo experto asigna al comparar el resultado del sistema con el suyo, subió de 54.38% a 80.36% a lo largo de todo el estudio. Durante el seguimiento promedió 72.46%.

Ya dentro del seguro, sobre 547,909 mastografías, las discrepancias graves con los radiólogos promediaron 0.08%.

Las cifras grandes descansan en poco: la prueba de calibración usó 100 mastografías, y no hubo revisión con datos de otras regiones ni de otros países. Lo que el estudio describe es esa red de Moscú, con equipos de tres fabricantes.

Tampoco se midió si alguna mujer fue diagnosticada antes, si recibió tratamiento más pronto o si el costo cambió. El estudio midió qué tan bien clasifica el sistema, no qué pasó después con las pacientes.

Dónde mirar: cuando un programa público de mastografía anuncie que suma un lector automático, pregunte si publica sus propios resultados de antes y después, con mujeres de su población. Los números de Moscú no son los suyos.

Qué significa para usted

Cuando su clínica o su seguro anuncien que suman un lector automático a las mastografías, pregunte si muestran resultados de antes y después con mujeres de su propia zona. Los números de Moscú no son los suyos: el estudio no revisó datos de otras regiones ni de otros países. Para saber si a una mujer le detectaron el cáncer más temprano, la pregunta se la responde un profesional de salud.

Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864

Quién pagó: El trabajo fue financiado por el Departamento de Salud de Moscú; cuatro autores son empleados de Third Opinion Platform, la desarrolladora del sistema de IA evaluado, y el artículo indica que aportaron información sobre el sistema pero no participaron en la realización del estudio, el análisis de datos ni la interpretación de los resultados.

No tome esto como consejo médico profesional.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1229 palabras · unos 6 minLeerla →Cerrar

Una máquina leyó 400 mil mamografías en Moscú. Lo que no puede decirle es si a usted le va a servir.

Un sistema de inteligencia artificial entró al seguro médico público ruso después de casi dos años y medio de pruebas. Los números subieron. La pregunta que queda abierta es quién los revisa ahora.

El cáncer de mama es el tumor más frecuente en las mujeres y la primera causa de muerte por cáncer en ellas. Por eso muchos países tienen programas de tamizaje: mamografías periódicas para encontrar el tumor antes de que se palpe. El estudio que nos ocupa se hizo en Moscú, Rusia, con un sistema llamado Third Opinion Mammograms, y sus autores trabajan en el Centro de Diagnóstico y Telemedicina del Departamento de Salud de Moscú.

Piense en la sala de lectura de un hospital: una mamografía no la ve un solo médico, la ven dos, y a veces más. Eso se llama doble lectura, y consume tiempo de radiólogos que escasean y que trabajan de más¹. Además, el factor humano pesa: hay diagnósticos que llegan tarde y biopsias que se hacen sin necesidad². Ahí es donde entra la máquina.

El sistema hace dos cosas a la vez. Primero busca zonas sospechosas en la imagen; después clasifica lo que encontró. Solo analiza imágenes, no decide nada por su cuenta³. Se entrenó con unas 4,000 mamografías de varias clínicas, con pacientes de edades, densidades de seno y equipos distintos⁴.

El estudio siguió el sistema durante 2 años y 5 meses. En ese lapso se revisaron 404,502 mamografías de 206 organizaciones médicas y participaron 336 radiólogos⁵. Al final, la capacidad del sistema para distinguir un caso con cáncer de uno sin cáncer subió de 0.83 a 0.92; la exactitud, de 0.77 a 0.90; la sensibilidad, de 0.84 a 0.88; y la especificidad, de 0.70 a 0.91⁶. La tasa de fallas técnicas bajó de 3.0% a 0.8%, y la calificación clínica que le pusieron los expertos subió de 54.38% a 80.36%⁷.

Hubo tropiezos antes del final. Las dos primeras pruebas de calibración no alcanzaron el mínimo que los propios investigadores se habían fijado. La tercera sí lo logró⁸. Después vinieron actualizaciones del sistema, y con ellas nuevas pruebas.

El sistema entró al programa de seguro médico obligatorio de Moscú, y ahí se vigiló nueve meses. En ese periodo, las discrepancias graves entre lo que decía la máquina y lo que decían los radiólogos promediaron 0.08%⁹. Lo curioso es que un experto y los lectores de rutina apenas se dieron la razón la mitad de las veces, y un análisis más fino mostró que, en los casos de discrepancias importantes, esa coincidencia era en promedio casi nula.

Aquí están los límites, y son grandes. La prueba de calibración se hizo con 100 mamografías¹⁰. Cien. Con eso se afinan detalles, no se demuestra que un sistema sirva para toda una población. Tampoco hubo validación externa: nadie probó este sistema con datos de otra región u otro país¹⁰. Todo se agrupó a través de un solo centro de referencia en Moscú, con 238 instalaciones bajo su mando¹¹. Y el estudio no midió si a alguna paciente le detectaron el cáncer más temprano, ni si el trabajo se agilizó, ni cuánto costó¹².

Hay una razón para desconfiar de los números que se ven bonitos en el papel. Un análisis que resume diez años de estudios sobre inteligencia artificial en mamografía encontró que, cuando se prueba en datos independientes, la exactitud cae hasta 24% respecto a las pruebas internas¹³. Según el resumen de ese trabajo, que solo pudimos leer en su resumen porque el texto completo está detrás de una suscripción, la exactitud promedio es alta pero tan dispersa que no se puede trasladar a un programa nuevo sin validarlo localmente¹⁴. El mismo resumen concluye que la evidencia sostiene a la inteligencia artificial como segundo lector, no como reemplazo del radiólogo¹⁵.

Otros trabajos que solo pudimos leer en resumen apuntan en la misma dirección. Un análisis de ensayos aleatorizados encontró que la razón de detección agrupada fue de 1.13, con un rango que va de 0.83 a 1.55, es decir, sin una ganancia clara y consistente¹⁶. Y las ganancias más grandes vienen de estudios no aleatorizados, que es justo el patrón que produciría un sesgo¹⁷. Una revisión sobre imágenes de mama señala que la evidencia proviene sobre todo de países ricos con poblaciones poco diversas, y que faltan datos de mortalidad a largo plazo¹⁸.

Así lo leemos nosotros. Cuando una herramienta nueva entra a un servicio público, al principio responde a una necesidad real; con el tiempo tiende a volverse un trámite que se defiende a sí mismo. Por eso hay que revisarla una y otra vez, no solo al inicio. Es de esperar que, tras la entrada del sistema al seguro, las revisiones periódicas se espacien o se vuelvan puro formato, y que el puntaje alto logrado en la etapa de prueba se sostenga más por inercia que por vigilancia. Sabríamos que nos equivocamos si el programa publica revisiones periódicas con resultados que suben y bajan, si los errores detectados se corrigen y se vuelven a medir, y si las reglas de aceptación siguen aplicándose después de la entrada en vigor.

Hay otra cosa que conviene mirar. Quien produce una herramienta y quien la usa no siempre entienden lo mismo por "buen resultado". En este estudio, algunos de los autores trabajan para la empresa que desarrolló el sistema; la publicación aclara que dieron información sobre el aparato pero no participaron en el análisis ni en la interpretación. Aun así, los criterios de aceptación fueron propios de este estudio y de este sistema de salud, no umbrales internacionales obligatorios. Usted puede preguntar quién definió los criterios para aprobar el sistema y si esos criterios eran públicos antes de las pruebas. Si trabaja en un servicio de salud, puede pedir que se acuerden por escrito antes de contratar.

No sabemos si eso ocurrió aquí. Sí sabemos que usted puede preguntar qué papel tuvo el sistema en su estudio y pedir que le expliquen el resultado en palabras claras. Si algo no le cuadra, vale la pena insistir en una segunda lectura.

Los propios autores lo dicen: su diseño no permite separar el efecto de una actualización del sistema de los cambios que ocurrían al mismo tiempo en los datos y en el flujo de trabajo. Tampoco pudieron evaluar de forma rigurosa la variabilidad entre observadores, porque el grupo de expertos era pequeño.

Nada de esto significa que la herramienta no sirva. Significa que sirvió en Moscú, con mujeres mayores de 18 años, con equipos de tres fabricantes concretos, dentro de un seguro público ruso. No hay nada en este estudio que muestre que funcionaría igual en una clínica de México, de Colombia o de Texas.

La pregunta que le queda a usted no es si la inteligencia artificial es buena o mala. Es una pregunta más concreta, y se la puede hacer en su próxima mamografía: ¿quién firma la lectura final? Si la respuesta es una persona, bien. Si es "el sistema", pregunte quién lo revisa y cada cuánto.

La máquina leyó 400 mil mamografías en Moscú y mejoró sus números. Lo que no puede decirle es si a usted le va a servir. Eso solo lo puede decir un programa que se mida donde usted vive, con pacientes como usted, y que lo publique.

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

Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "However, screening entails substantial resource demands, including double-reading by radiologists ( 4 ). Unfortunately, radiologists remain scarce facing widespread overwork and burnout as mammography volumes steadily increase ( 5 , 6 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Human factors contribute to diagnostic errors like delayed cancer diagnoses or unnecessary biopsies in malignancy-free women ( 7 , 8 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This solution simultaneously performs computer-aided detection (CADe) to localize target findings and computer-aided diagnosis (CADx) for classifying them. Its clinical task was detecting and classifying breast cancer signs on mammograms, analyzing images only."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The training dataset of approximately 4,000 mammograms was collected from several partner clinics of the Third Opinion Platform, ensuring broad population and ethnic diversity."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The test dataset comprised 404,502 mammograms from 206 medical organizations and three mammography equipment manufacturers. A total of 336 radiologists participated. The testing and monitoring period lasted 2 years and 5 months."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Over this time, AUC increased by 10.8% (from 0.83 to 0.92), accuracy by 16.9% (from 0.77 to 0.90), sensitivity by 4.8% (from 0.84 to 0.88), and specificity by 30.0% (from 0.70 to 0.91)."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The average technical defect rate decreased by 26.7% (from 3.0% to 0.8%), and the clinical assessment score rose by 47.8% (from 54.38% to 80.36%). The study culminated in the integration of the AI system into the regional CMI program."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The first two calibration tests failed to meet the predefined performance thresholds (0.80). The third test successfully satisfied these criteria ( Figure 2 , Table 2 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "CMI monitoring lasted 9 months. Mean critical BI-RADS discrepancies between AI and radiologists was 0.08% ( Figure 5A )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "A key limitation of this study is the relatively small retrospective calibration testing dataset (100 mammograms) and the lack of external validation on independent datasets from other regions or countries."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Second, although data were aggregated through the Moscow Reference Center for Diagnostic Imaging—which oversees 238 facilities and ensures population diversity—the study’s single-center design remains a critical limitation."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Finally, the study lacked a formal assessment of the AI system’s impact on patient-centered clinical outcomes, organizational workflows, or economic efficiency."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - el artículo del que trata esta nota — el artículo completo — el pasaje: "External validation on independent datasets occurs rarely and shows accuracy drops of up to 24% compared to internal tests ( 19 - 21 ). Many AI systems lack peer-reviewed efficacy evidence ( 22 )."
Ciurescu S, Buciu V, Ilaș DG, Pârvănescu R, Șerban D. (2026). Ten Years of Artificial Intelligence in Screening Mammography: A Systematic Review and Meta-Analysis of Diagnostic Accuracy and Clinical Implementation (Literature Published 2015–2025). Diagnostics. 10.3390/diagnostics16183045 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Pooled accuracy is high on average but so dispersed that it cannot be transferred to a new programme; local validation before deployment remains necessary."
Ciurescu S, Buciu V, Ilaș DG, Pârvănescu R, Șerban D. (2026). Ten Years of Artificial Intelligence in Screening Mammography: A Systematic Review and Meta-Analysis of Diagnostic Accuracy and Clinical Implementation (Literature Published 2015–2025). Diagnostics. 10.3390/diagnostics16183045 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "A decade of evidence supports AI as a second reader and triage tool in organised screening, not as an autonomous replacement for the radiologist."
Ciurescu S, Buciu V, Ilaș DG, Pârvănescu R, Șerban D. (2026). Ten Years of Artificial Intelligence in Screening Mammography: A Systematic Review and Meta-Analysis of Diagnostic Accuracy and Clinical Implementation (Literature Published 2015–2025). Diagnostics. 10.3390/diagnostics16183045 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Across randomised and paired prospective trials (k = 3), the pooled detection rate ratio was 1.13 (0.83-1.55), with the MASAI randomised trial alone reporting 1.29 (1.09-1.51) and a 44% reduction in screen reading."
Ciurescu S, Buciu V, Ilaș DG, Pârvănescu R, Șerban D. (2026). Ten Years of Artificial Intelligence in Screening Mammography: A Systematic Review and Meta-Analysis of Diagnostic Accuracy and Clinical Implementation (Literature Published 2015–2025). Diagnostics. 10.3390/diagnostics16183045 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The larger and more precise detection gains come from non-randomised designs, which is the pattern confounding would produce, so the randomised evidence remains the anchor."
Kumar A, Lip G. (2026). Artificial Intelligence in Breast Imaging and Screening: Current Evidence, Applications, and Future Directions. Indian Journal of Radiology and Imaging. 10.1055/s-0046-1827249 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "However, the evidence base remains predominantly derived from nondiverse, high-income country populations, and long-term outcome data including breast cancer mortality are not yet available."

No tome esto como consejo médico profesional.

other · Quantitative imaging in medicine and surgery · the paper, 10 Aug 2026 · free

An AI Reader Passed Moscow's Mammogram Tests. The Tests Used 100 Confirmed Cases.

A government insurance program in the Russian capital took in an AI mammogram reader after two years of monitoring. The results describe that one city, and nobody checked whether any woman's cancer was found earlier.

Short version · the longer version follows, about 6 min

Ask the weeklyAI chat

Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.

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 aged 18 and over whose mammograms were read by an AI second reader
How many: 404,502 mammograms
Where: Moscow, Russia
When: 2 years and 5 months
Kind of study: analysis of what people did
Who did it: Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Department of Health
The limit that matters: The confirmed-cancer test set was 100 mammograms, and nothing was checked outside Moscow.

How the AI reader scored at the start and at the end of testing

Telling cancer from non-cancer, start0.830-to-1 scale

Telling cancer from non-cancer, end0.920-to-1 scale

Share of correct answers, start0.770-to-1 scale

Share of correct answers, end0.90-to-1 scale

Correct 'yes, this is cancer' calls, start0.840-to-1 scale

Correct 'yes, this is cancer' calls, end0.880-to-1 scale

Correct 'no' calls, start0.70-to-1 scale

Correct 'no' calls, end0.910-to-1 scale

These are the software's scores on a 0-to-1 scale before and after two years of testing and updates in one Moscow program; the confirmed-cancer test set behind them was only 100 mammograms, and nothing was checked outside Moscow.

What the study measured was how well the software sorted images.
weeklyAI's reading

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

A mammography AI system went through a series of tests in Moscow, and then the city's government insurance program let it read mammograms alongside doctors. Whether it helps women anywhere else is a different question. Nobody answered it here.

The Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Department of Health put the system, called Third Opinion Mammograms, through a staged process. First came a check of basic functions, on four mammograms. One of the four was processed incompletely, and the developers fixed it.

Then came the calibration tests: 100 mammograms, half of women with confirmed breast cancer and half without. The system was scored against the fixed bar the researchers had set for accuracy.

The first test scored 0.77 for accuracy against a bar of 0.80. The second scored 0.74. After the developers made changes, the third test passed at 0.81. During the monitoring period, the developers released six updates. Four of them changed the core model, and each of those went through the 100-mammogram test again.

Then the system ran on every mammogram from the sites that took part, no exceptions. Technical monitoring ran for two years and one month, and it covered 429,027 mammograms. In most months, fewer than one in ten images failed to process fully. One month went just above that line.

In June 2023, six months into testing, it entered Moscow's compulsory medical insurance program. Across 547,909 mammograms there, the system and the radiologists reading with it disagreed on a serious finding 0.08 percent of the time, which is fewer than one mammogram in a thousand.

The 100 mammograms are a small set to carry those headline gains. And the study never tested the system outside this one Moscow network. If a screening clinic in your city adopts an AI reader, none of these numbers tell you how it will do there. The study cannot show that the AI caused the improvement either, because the updates came alongside other changes.

What the study measured was how well the software sorted images. It did not follow the women. It cannot say that a cancer was caught earlier, treated sooner, or that care cost less.

Where you live, there is a question to keep for any screening program that adds an AI reader: show us its own before-and-after results, on our population. Ask for a number from here.

What this means for you

For now, there is nothing to do with this result in your own city. The study describes one Moscow network, and it did not follow the women to see whether any cancer was found earlier. If a clinic near you adds an AI reader, ask the people who run it for their own before-and-after numbers, on your population, and ask a doctor what those numbers mean for your screening.

Who paid: The work was funded by the Moscow Health Care Department; four authors are employees of Third Opinion Platform, the developer of the AI system evaluated, and the article states they provided information about the system but were not involved in conducting the study, data analysis or interpretation of results.

Do not take this as professional medical advice.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1195 words · about 6 minRead it →Close

A Mammography AI Earned Its Place in Moscow's Public Insurance. The Test It Passed Was Small.

Over two years and five months, one Russian screening program watched an AI second reader improve, then let it into a government insurance scheme. It was never compared against a group that went without.

A mammogram is two black-and-white pictures of each breast, taken from above and from the side, and in a public screening program a radiologist has to read all four. In Moscow, a software system was given that job as a second reader. The system is called Third Opinion Mammograms, and it does two things at once: it circles the spots on the image where something suspicious sits, and it sorts those findings into harmless and worth-a-look¹. It reads images only. It does not talk to the patient, order a biopsy or decide anything.

The study ran for two years and five months and covered 404,502 mammograms from 206 medical organizations, read alongside 336 radiologists². Over that time, the system's ability to tell a cancer from a non-cancer rose from 0.83 to 0.92 on a 0-to-1 scale where 1 is perfect; its overall share of correct answers rose from 0.77 to 0.90³. Its share of correct "yes, this is cancer" calls rose from 0.84 to 0.88, and its share of correct "no" calls rose from 0.70 to 0.91³. The share of images the software failed to process properly fell from 3.0% to 0.8%, and the score expert radiologists gave its clinical output rose from 54.38% to 80.36%⁴. In the end, it was let into Moscow's regional compulsory medical insurance program⁴.

That entry was not automatic. The first two rounds of testing missed the bar the researchers had set for themselves; the third passed⁵. Six updates followed during monitoring, and the four that changed the model's inner workings had to be retested from scratch before use.

The limits are as plain as the gains. The test that checks the system against confirmed cancer cases rested on 100 mammograms⁶. Nothing was checked against independent data from another region or country⁶. The whole thing was aggregated through one Moscow reference center, which the authors call a critical limitation⁷. And the study did not measure whether any woman's cancer was found earlier, whether care got better, or what it cost⁸.

Consider what the clinic visit covers. Breast cancer is the most common cancer in women and the leading cause of cancer death among them, which is why so many countries run screening programs⁹. Those programs are expensive in a specific way: they require a radiologist to read every image twice¹⁰. Radiologists are scarce, and the article notes widespread overwork and burnout as mammography volumes rise¹⁰. Human readers also make mistakes that matter — a cancer read late, or a biopsy taken from a woman who turned out to have nothing¹¹.

The technology in the Moscow study is not a chatbot. It is a pair of image-reading models trained on about 4,000 mammograms gathered from several partner clinics, chosen for diversity of age, breast density, equipment and image quality¹². One model hunts for clusters of tiny calcium specks; the other looks for masses, asymmetry and architectural distortion. The training answers came from a consensus reading by three expert radiologists, and the dataset included both biopsy-proven cases and cases diagnosed on imaging alone.

This family of studies has a history, and it is not all good news. The article describes a study in which external checks on independent data happen rarely, and when they do, accuracy can drop by up to 24% compared with the developer's own tests¹³. The article describes a study in which many AI systems on the market have no peer-reviewed evidence at all¹³. The article describes a study in which most testing frameworks are one-time assessments that ignore how a model changes when it is retrained, and that leave radiologists out of the loop until it is too late to shape the tool¹⁴.

A pooled analysis of 14 studies covering more than 1.2 million examinations found an average score of 0.890 — close to Moscow's final figure — but with results spread so widely that the summary's own conclusion is that a high average cannot be carried to a new program, and local testing before launch remains necessary¹⁵¹⁶. A separate review, again read only in summary, reports that AI-supported screening raised cancer detection by 10 to 29% and cut radiologist workload by 31 to 64% in double-reading settings, but that the evidence comes mostly from wealthy countries and long-term outcome data are not yet available¹⁷¹⁸.

Here is how we read it. A tool judged against the decisions people already made inherits their habits, including their disagreements with each other. In Moscow, the expert and the everyday radiologists agreed with one another only about half the time — which tells you something about how hard these images are, and something about how carefully any "match the radiologist" number should be read. When you hear that an AI agrees with doctors at some rate, the fair question is: which doctors, and did those doctors agree with each other?

There is a second pattern worth carrying home. A screening test that finds more early-stage disease can make a program look better without anyone living longer, because some of what it finds would never have harmed the woman. The Moscow study did not measure lives or late-stage cancers, so it cannot settle that. If your clinic announces that its new AI caught more cancers early, ask whether late-stage cancers or deaths in the screened group went down too — that is the number that belongs to patients rather than to programs.

And a third, smaller thing. People will accept a machine's opinion more readily when they can argue with it. The Moscow authors built radiologist feedback into every stage, and the clinical score climbed — though whether that came from the feedback or the retraining cannot be separated, as the authors themselves note. We also read other work in this field whose finding is that AI belongs beside the radiologist as a second reader, not in place of one. That is not a doubt about what happened in Moscow. It is about what carries over.

For a reader in Latin America, the United States or Canada, the honest position is this: the study shows a system improving inside one public insurance program in one city, on machines from three manufacturers, in women aged 18 and over. It says nothing about how the same software would behave in your clinic, on your equipment, with your radiologists. It also says nothing about cost.

So what can you do with it? When you next sit in a screening room, you can ask two questions that cost nothing. Who reads my final result — a person, a machine, or both? And if the machine and the person disagree, who decides? Those are the questions this study makes askable. They are also the ones a program using AI should be able to answer without looking anything up.

The four pictures are still taken the same way. What changed, in one city, is who looks at them second.

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

Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "This solution simultaneously performs computer-aided detection (CADe) to localize target findings and computer-aided diagnosis (CADx) for classifying them. Its clinical task was detecting and classifying breast cancer signs on mammograms, analyzing images only."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "The test dataset comprised 404,502 mammograms from 206 medical organizations and three mammography equipment manufacturers. A total of 336 radiologists participated. The testing and monitoring period lasted 2 years and 5 months."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "Over this time, AUC increased by 10.8% (from 0.83 to 0.92), accuracy by 16.9% (from 0.77 to 0.90), sensitivity by 4.8% (from 0.84 to 0.88), and specificity by 30.0% (from 0.70 to 0.91)."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "The average technical defect rate decreased by 26.7% (from 3.0% to 0.8%), and the clinical assessment score rose by 47.8% (from 54.38% to 80.36%). The study culminated in the integration of the AI system into the regional CMI program."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "The first two calibration tests failed to meet the predefined performance thresholds (0.80). The third test successfully satisfied these criteria ( Figure 2 , Table 2 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "A key limitation of this study is the relatively small retrospective calibration testing dataset (100 mammograms) and the lack of external validation on independent datasets from other regions or countries."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "Second, although data were aggregated through the Moscow Reference Center for Diagnostic Imaging—which oversees 238 facilities and ensures population diversity—the study’s single-center design remains a critical limitation."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "Finally, the study lacked a formal assessment of the AI system’s impact on patient-centered clinical outcomes, organizational workflows, or economic efficiency."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "Breast cancer is the most prevalent malignancy among women and the leading cause of cancer-related mortality in this population ( 2 ). Consequently, numerous countries implement mammography screening programs, facilitating early detection and timely intervention to reduce mortality ( 3 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "However, screening entails substantial resource demands, including double-reading by radiologists ( 4 ). Unfortunately, radiologists remain scarce facing widespread overwork and burnout as mammography volumes steadily increase ( 5 , 6 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "Human factors contribute to diagnostic errors like delayed cancer diagnoses or unnecessary biopsies in malignancy-free women ( 7 , 8 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "The training dataset of approximately 4,000 mammograms was collected from several partner clinics of the Third Opinion Platform, ensuring broad population and ethnic diversity."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "External validation on independent datasets occurs rarely and shows accuracy drops of up to 24% compared to internal tests ( 19 - 21 ). Many AI systems lack peer-reviewed efficacy evidence ( 22 )."
Vasilev, Y., Rumyantsev, D., Vladzymyrskyy, A. et al. (2026). Implementation of an artificial intelligence-based system for mammography in the compulsory medical insurance program: results of a 3-year study. Quantitative Imaging in Medicine and Surgery. https://doi.org/10.21037/qims-2026-0864 - the article this story is about — the whole article — the passage: "First, most frameworks overlook AI’s dynamic evolution through retraining, requiring continuous post-update assessment ( 32 ). Second, they neglect radiologist involvement throughout development and testing, which risks creating misaligned solutions ( 33 , 34 )."
Ciurescu S, Buciu V, Ilaș DG, Pârvănescu R, Șerban D. (2026). Ten Years of Artificial Intelligence in Screening Mammography: A Systematic Review and Meta-Analysis of Diagnostic Accuracy and Clinical Implementation (Literature Published 2015–2025). Diagnostics. 10.3390/diagnostics16183045 — only the abstract - the full text could not be fetched — the passage: "Pooled standalone AUC across 14 studies and 1,214,885 examinations was 0.890 (95% CI 0.858-0.915), with I 2 = 97.4% and a 95% prediction interval of 0.731-0.960."
Ciurescu S, Buciu V, Ilaș DG, Pârvănescu R, Șerban D. (2026). Ten Years of Artificial Intelligence in Screening Mammography: A Systematic Review and Meta-Analysis of Diagnostic Accuracy and Clinical Implementation (Literature Published 2015–2025). Diagnostics. 10.3390/diagnostics16183045 — only the abstract - the full text could not be fetched — the passage: "Pooled accuracy is high on average but so dispersed that it cannot be transferred to a new programme; local validation before deployment remains necessary."
Kumar A, Lip G. (2026). Artificial Intelligence in Breast Imaging and Screening: Current Evidence, Applications, and Future Directions. Indian Journal of Radiology and Imaging. 10.1055/s-0046-1827249 — only the abstract - the full text could not be fetched — the passage: "Current data demonstrate that AI-supported mammography screening can increase cancer detection rates by 10 to 29% and reduce interval cancer rates by up to 12% while reducing radiologist workload by 31 to 64% in double-reading screening settings, without increasing false-positive rates."
Kumar A, Lip G. (2026). Artificial Intelligence in Breast Imaging and Screening: Current Evidence, Applications, and Future Directions. Indian Journal of Radiology and Imaging. 10.1055/s-0046-1827249 — only the abstract - the full text could not be fetched — the passage: "However, the evidence base remains predominantly derived from nondiverse, high-income country populations, and long-term outcome data including breast cancer mortality are not yet available."

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 free weekly drip by email

One email a week: the free story of every section, with its picture, and the headlines of the rest.

No payment, nothing to cancel. Unsubscribe from any issue.

El goteo semanal gratis por correo

Un correo por semana: la nota gratuita de cada sección, con su imagen, y los titulares de las demás.

Sin pago y sin nada que cancelar. Puede darse de baja desde cualquier envío.