AI Cut the Hours Spent Outlining Scans Before Rectal Cancer Radiotherapy

other · Journal of applied clinical medical physics · la publicación, 1 sep 2026 · gratis

La inteligencia artificial dibuja los órganos para radioterapia: más rápida y más parecida entre doctores

Un estudio de un solo hospital en China midió qué tan bien coincidían los contornos automáticos con los aprobados por los médicos. No midió la dosis de radiación ni los resultados del tratamiento.

Versión breve · la versión detallada sigue, unos 6 min

El estudio, de un vistazo
Quiénes: Personas con cáncer de recto que recibieron radioterapia antes de una cirugía
Cuántos: 150 pacientes, en tres grupos de 50
Dónde: Beijing, China
Cuándo: El sistema se instaló en mayo de 2018 y se actualizó en diciembre de 2024
Tipo de estudio: analysis of what people did
Quién lo hizo: Hospital del Cáncer del Centro Nacional del Cáncer de China
El límite que importa: No midió la dosis de radiación ni los resultados del tratamiento

Tiempo total de trabajo según el método de delineado

A mano41.15min

Con la segunda versión del sistema16.97min

Minutos de trabajo total en 21 casos redelineados por seis oncólogos; el estudio no midió la dosis de radiación ni los resultados del tratamiento.

Qué tanto se parecían los trazos automáticos a los trazos finales aprobados por los médicos

Antes de instalar el sistemafrente aDespués de la primera versión

Para el área del tumor subió de 0.87 a 0.88; para los órganos sanos subió de 0.80 a 0.88

Después de la primera versiónfrente aDespués de la actualización

Para el área del tumor subió de 0.88 a 0.93; para los órganos sanos subió de 0.88 a 0.95

Primera versión del sistemafrente aSegunda versión del sistema

Los trazos fallidos bajaron de 3.30% a 0.64%

Delineado a manofrente aDelineado con la segunda versión

El tiempo total bajó alrededor de 58.8%

Delineado con la primera versiónfrente aDelineado con la segunda versión

El tiempo total bajó 21.9%

Contornos automáticos sin editar de la primera versiónfrente aContornos automáticos sin editar de la segunda versión

95.2% de los de la segunda alcanzó calificación aceptable o mejor, frente a 28.6% de los de la primera

El estudio midió qué tan parecidos eran los trazos, no cuánta radiación recibió cada paciente.
Lectura de weeklyAI

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

Para tratar un cáncer de recto con radioterapia, antes hay que dibujar. Sobre cada tomografía, el oncólogo traza a mano el tumor, los ganglios de riesgo y cada órgano sano que debe quedar fuera del disparo: vejiga, intestino, colon, fémur, periné. El artículo llama a esa tarea laboriosa y lenta, y es fácil ver por qué: son muchas estructuras, capa por capa, y dos médicos pueden trazarlas distinto.

En un centro oncológico de Beijing, el Hospital del Cáncer del Centro Nacional del Cáncer de China, se instaló en mayo de 2018 un sistema de inteligencia artificial que hace ese trazado solo. En diciembre de 2024 fue reemplazado por una segunda versión. Para saber qué cambió, los investigadores compararon tres grupos de 50 pacientes cada uno: antes del sistema, después de la primera versión y después de la actualización. En total, 150 personas con cáncer de recto que recibieron radioterapia antes de una cirugía. El artículo apareció en una revista en 2026.

La medida era simple: qué tanto se parecían los trazos automáticos, todavía sin corregir, a los trazos finales que los médicos aprobaron para el tratamiento. Con la actualización, ese parecido subió de 0.88 a 0.93 para el área del tumor y de 0.88 a 0.95 para los órganos sanos. Los números van de 0 a 1: más cerca de 1 significa más coincidencia.

También bajaron los trazos fallidos: de 3.30% a 0.64%. Y en 21 casos adicionales, seis oncólogos volvieron a delinear usando tres métodos: a mano, con la primera versión y con la segunda. Con la segunda, el tiempo total bajó alrededor de 58.8% frente al trabajo completamente manual y 21.9% frente a la primera versión.

El estudio midió qué tan parecidos eran los trazos, no cuánta radiación recibió cada paciente. Trazos mejor alineados no equivalen todavía a mejores resultados del tratamiento, y nadie debería esperar un desenlace distinto por este software. Todo vino de un solo hospital, con su propio protocolo de escaneo y sus propias guías; lo que pasó allí puede no repetirse en su clínica. Y cada contorno fue revisado y corregido por un oncólogo antes del tratamiento: es una ayuda para el personal, no un reemplazo del profesional que planifica su atención.

¿Qué podría hacer posible? En clínicas que adopten algo así, las horas de trazado que preceden a la radioterapia podrían acortarse, y los dibujos de distintos médicos podrían parecerse más entre sí.

Si alguien de su familia va a recibir radioterapia, vale preguntar en su centro de salud: ¿quién revisa y firma cada contorno antes de que empiece el tratamiento?

Qué significa para usted

Si en su clínica empiezan a usar una ayuda automática para dibujar los órganos antes de la radioterapia, lo que se midió fue el parecido de esos trazos y el tiempo de trabajo, no la dosis ni el resultado del tratamiento. Vale preguntar quién revisa y firma cada contorno antes de que comience.

Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768

Quién pagó: El artículo declara que los autores no tienen conflictos de interés y no menciona ningún financiador; señala que ambos modelos se entrenaron como prototipos de investigación con el conjunto de datos de la propia institución en un servidor interno.

No tome esto como consejo médico profesional.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1123 palabras · unos 6 minLeerla →Cerrar

La radioterapia empieza con un dibujo: un estudio midió cuánto ayuda la inteligencia artificial a hacerlo

En 150 pacientes con cáncer de recto en un solo hospital de China, el software que delinea estructuras en las tomografías acercó sus resultados a lo que los oncólogos aceptaron como tratamiento, y la segunda versión del sistema redujo el tiempo de trabajo.

Antes de que a una persona le apliquen radioterapia, alguien tiene que decidir con precisión qué zonas del cuerpo van a recibir dosis y cuáles no. Esa tarea, el delineado de los volúmenes que se van a tratar y de los órganos que hay que proteger, es crítica y consume mucho tiempo de trabajo especializado¹. Un grupo de investigadores del Centro Nacional del Cáncer de China, en Beijing, evaluó qué pasó en su propio servicio cuando dejaron que un programa de aprendizaje profundo hiciera ese delineado y luego lo actualizaron.

La tecnología funciona así: para cada paciente nuevo, las imágenes de tomografía de planificación salen del sistema donde se planifica el tratamiento hacia un servidor, allí el modelo traza los contornos automáticamente, y esos contornos vuelven al sistema de planificación para que el oncólogo los revise y los ajuste a mano². El estudio no comparó pacientes con y sin el software en cuanto a resultados de salud, sino que midió cuánto se parecían los trazos automáticos a los trazos finales que los médicos aprobaron para tratar.

Con la primera generación del sistema, el parecido entre el contorno automático y el contorno final del volumen que incluye el tumor y las zonas de riesgo fue de 0.87 antes de implantarlo y 0.88 después, una diferencia que los autores no consideraron concluyente; para los órganos que hay que proteger, en cambio, subió de 0.80 a 0.88³. Con la actualización a la segunda generación, ese parecido subió de 0.88 a 0.93 para el volumen del tumor y de 0.88 a 0.95 para los órganos protegidos⁴. En la misma línea, la proporción de casos en que el sistema se equivocaba lo suficiente como para considerarse un fallo bajó de 3.30% a 0.64%⁵.

En el tiempo de trabajo la diferencia fue la más fácil de imaginar: el método asistido por la segunda generación tardó unos 17 minutos en promedio, mientras que el delineado completamente manual tardó unos 41 minutos⁶. Medido de otra forma, el método de segunda generación redujo el tiempo total alrededor de 58.8% frente al manual y 21.9% frente al asistido por la primera generación⁷. Además, los contornos hechos con la segunda generación fueron los que más coincidieron entre distintos oncólogos⁸, y en una evaluación a ciegas por dos especialistas senior, 99.2% de los contornos finales revisados por los oncólogos recibió una calificación de aceptable o mejor, mientras que los contornos automáticos sin editar de la segunda generación fueron mejor valorados que los de la primera⁹. Entre los generados por la segunda generación, 95.2% alcanzó esa calificación, frente a 28.6% de los de la primera¹⁰.

Conviene retener los límites. Los autores no midieron la dosis de radiación que recibió ningún paciente¹¹, y compararon grupos distintos de pacientes en tres momentos del tiempo, lo que puede introducir diferencias que no vienen del software¹². También describieron por qué falla el sistema: confunde estructuras vecinas, se desorienta cuando el paciente no está en la posición habitual, y no incorpora información de resonancia ni el historial del paciente.

Este estudio pertenece a una familia de investigaciones que ya no preguntan si la inteligencia artificial puede hacer una tarea médica, sino qué pasa cuando entra en un servicio real. En mamografía, por ejemplo, la mayoría de los programas europeos de cribado dependen de que dos radiólogos lean cada estudio, un modelo que da beneficios pero que se ve presionado por la carga de trabajo y por los cánceres que aparecen entre una ronda y otra¹³; según el resumen de esa revisión, que solo pudimos leer en su resumen porque el texto completo está detrás de una suscripción, la inteligencia artificial se está evaluando para apoyar u optimizar esos circuitos ya establecidos¹⁴. Esa revisión reunió tres estudios grandes hechos dentro de programas nacionales de cribado en Europa¹⁵, y encontró que, sobre casi 600,000 exámenes, la integración de la inteligencia artificial podría producir un aumento pequeño en la detección de cáncer, del orden de un caso por cada mil personas examinadas, sin un aumento consistente en la cantidad de mujeres llamadas a repetir estudios¹⁶. Los autores de esa revisión concluyeron que la herramienta funciona mejor como un lector complementario dentro de esos circuitos, y que su implementación exige controles de calidad explícitos y seguimiento de los cánceres que aparecen entre rondas¹⁷.

La diferencia entre ese caso y el de radioterapia es instructiva. En mamografía, la inteligencia artificial se suma a una cadena de lectura ya existente. En radioterapia, el software no diagnostica: prepara el terreno para un tratamiento que igualmente va a aplicarse, y lo que se mide es cuánto trabajo humano ahorra y cuánto se parecen sus trazos a los que un especialista aceptaría. Por eso, los propios autores de la revisión europea piden vigilar los resultados a largo plazo¹⁸, algo que este estudio de recto tampoco puede ofrecer: los autores no midieron la dosis de radiación que recibió ningún paciente.

Así lo leemos nosotros. Cuando una tarea es repetitiva y exige precisión, tendemos a delegarla en un sistema que no se cansa y luego a revisar lo que produjo. Lo que este trabajo sugiere es que, en un servicio de radioterapia, el personal empieza a tratar el contorno automático como punto de partida y dedica menos tiempo a corregirlo, sobre todo cuando el sistema mejora. Esperaríamos ver eso mismo en más centros con el tiempo. Usted podría comprobarlo a la inversa: si los contornos finales se alejaran de los automáticos, si el tiempo de corrección no bajara o si los profesionales volvieran a delinear todo desde cero, nuestra lectura estaría equivocada.

Hay algo que conviene tener claro antes de entusiasmarse. Aquí nadie sustituyó al oncólogo: cada contorno pasó por revisión y ajuste humano antes del tratamiento, y el estudio no midió si eso cambió el resultado para el paciente. Es decir, la frontera entre quien produce y quien supervisa se vuelve borrosa, pero la responsabilidad sigue siendo de una persona. Si alguien de su familia recibe radioterapia, usted puede preguntar si el servicio usa delineación automática, quién revisa y firma cada contorno antes de aplicarlo, y si esa revisión queda registrada en la historia clínica. También puede preguntar si el centro ofrece una segunda opinión cuando el caso es complejo. Esas preguntas no dependen de que este software exista en su país: dependen de que alguien las haga. ¿Sabe usted, hoy, quién firma el plan de tratamiento de las personas que quiere?

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

Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Accurate delineation of target volumes and organs‐at‐risk (OARs) is a critical yet labor‐intensive component of rectal cancer radiotherapy."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In clinical practice, for each new patient, planning CT images are transferred from the treatment planning system (TPS) to the server, where the model performs automatic contouring. The predicted contours are then transmitted back to the TPS for review and manual adjustment by the radiation oncologist."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The mean DSC values before and after implementing were 0.87 ± 0.04 and 0.88 ± 0.04 for CTV (P = 0.067) and 0.80 ± 0.06 and 0.88 ± 0.05 for OARs (P < 0.001), respectively."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The mean DSC values before and after updating were 0.88 ± 0.04 and 0.93 ± 0.04 for CTV (P < 0.001) and 0.88 ± 0.05 and 0.95 ± 0.02 for OARs (P < 0.001), respectively."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "After automatic contouring system update, the mean failure rate decreased from 3.30% to 0.64%."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The total time significantly differed between the three contouring methods (P < 0.05), with the Auto2‐assisted method being the fastest (16.97 ± 5.52 min) and manual contouring being the most time‐consuming (41.15 ± 10.43 min)."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Auto2‐assisted method decreased the total time by approximately 58.8% compared with the manual method, and 21.9% compared with the Auto1‐assisted method."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The Auto2‐assisted contouring group demonstrated significantly higher inter‐observer consistency than the other groups (all: P < 0.05)."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In the blinded clinical evaluation, 99.2% (125/126) of the oncologist‐revised final contours received a Likert score of ≥ 4, and Auto2‐generated unedited contours showed significantly higher clinical acceptability than Auto1 (4.02 ± 0.25 vs. 3.26 ± 0.49, P < 0.001)"
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In addition, 95.2% (20/21) of Auto2‐generated contours achieved a score of ≥ 4, compared with 28.6% (6/21) for Auto1."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "First, although the present study comprehensively evaluated contouring efficiency, geometric performance, inter‐observer consistency, and blinded clinical acceptability, dosimetric evaluation was not included in the current work."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Second, the current study compared different patient cohorts across three phases. This longitudinal design was necessary to accurately reflect the real‐world evolution of clinical workflows. However, this approach may introduce confounding effects due to individual patient differences across the distinct cohorts."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Most European population mammography screening programs rely on double reading with arbitration, a model that delivers mortality benefit but is increasingly challenged by radiologist workload, variable specificity, and interval cancers."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Artificial intelligence (AI) is being evaluated to support or optimize these established European screening pathways."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "We performed a prespecified, focused evidence synthesis of three large studies embedded within routine population screening programs operating under European-relevant workflows: MASAI (randomized AI-supported risk triage within a national program), ScreenTrustCAD (prospective paired-reader evaluation with AI as an independent reader in a double-reading framework), and PRAIM (nationwide decision-referral implementation)."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Across 597,419 examinations, the pooled CDR RD was +0.9 per 1000 (95% CI -0.0 to +1.8; I"
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "In European population screening programs characterized by double reading and arbitration, prospective program-embedded evidence suggests that AI integration may yield a small absolute increase in cancer detection (≈1/1000) without a consistent increase in recall, alongside improved PPV and efficiency signals."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "These findings suggestAI primarily as a complementary reader within European screening workflows, with implementation requiring explicit quality assurance and monitoring of interval cancers and stage distribution."

No tome esto como consejo médico profesional.

other · Journal of applied clinical medical physics · the paper, 1 Sep 2026 · free

AI Cut the Hours Spent Outlining Scans Before Rectal Cancer Radiotherapy

In one Chinese hospital, software that traces organs on planning scans got faster and more consistent. It did not change the treatment itself, and no one skipped the oncologist's review.

Short version · the longer version follows, about 6 min

The study at a glance
Who: People with rectal cancer receiving radiotherapy
How many: 150 patients, plus 21 more whose scans six oncologists redrew
Where: Beijing, China
When: Software in use May 2018; updated December 2024
Kind of study: analysis of what people did
Who did it: National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences
The limit that matters: It did not measure the radiation dose patients received.

Share of unedited software drafts scored 4 or above out of 5 by senior oncologists

Second-generation system95.2%

First-generation system28.6%

These are shares of 21 unedited drafts per system, judged blind by two senior oncologists on a 5-point scale where 4 means good with minor edits needed.

Second-generation system compared with the first-generation system and with fully manual contouring

Second-generation systemagainstFirst-generation system

Total time fell by about 21.9 percent

Second-generation systemagainstFully manual contouring

Total time fell by about 58.8 percent

Second-generation systemagainstFirst-generation system

Senior oncologists scored the unedited drafts higher, 4.02 out of 5 against 3.26

Second-generation systemagainstFirst-generation system and manual contouring

Oncologists using it agreed with each other more closely than with either other method

And every outline was still reviewed and edited by an oncologist before treatment — this assists the staff who plan your care; it does not replace them.
weeklyAI's reading

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

For people with rectal cancer, radiotherapy often comes before surgery. Before any of it begins, a doctor must trace the tumor's target area and nearby healthy organs, layer by layer, on every planning scan. The study calls this work labor-intensive and time-consuming.

At a single cancer center in Beijing, that tracing has a longer history than most. A deep-learning contouring system went into use in May 2018, and in December 2024 it was replaced by a second-generation version. To see what changed, researchers compared three groups of 50 patients each — one group treated before any software existed, one after the first system arrived, and one after the update. That makes 150 patients in all, plus 21 more whose scans six oncologists redrew using each method.

The numbers describe how closely the software's outlines matched the outlines finally approved for treatment. After the update, that match rose from 0.88 to 0.93 for the tumor target and from 0.88 to 0.95 for the organs at risk — a scale where 1.0 would mean an exact overlap.

How often the software's outlines fell outside acceptable limits dropped from about 3 percent to under 1 percent. And when oncologists used the second-generation system, the total time spent contouring fell by about 58.8 percent compared with working entirely by hand.

The study also measured how long contouring took, how consistent different oncologists were, how accurate the outlines were against expert consensus, how often the software failed, and how acceptable the results were to senior oncologists. It did not measure the radiation dose patients received, so it cannot show that treatment results improved.

The data came from one hospital in China, with its own scanning protocol, and the outlines were drawn according to international guidelines. What happened there may not hold at your clinic. And every outline was still reviewed and edited by an oncologist before treatment — this assists the staff who plan your care; it does not replace them.

Still, in clinics that adopt such a system, the hours of outlining that precede radiotherapy could shrink, and the outlines different doctors produce could agree more closely.

If someone you know is preparing for radiotherapy, ask who reviews the planning scans — and whether software is helping draw them.

What this means for you

What this means for you, hearing only the alarm, is narrow and worth holding onto precisely: in one Beijing hospital, software that outlines scans before rectal cancer radiotherapy got faster and more consistent after an update, and no oncologist's review was skipped. Ask whether software helps draw the planning scans where you live, and who reviews them. It cannot show that treatment results improved.

Who paid: The article states the authors declare no conflicts of interest and lists no funder, and it says both models were trained as research prototypes on the institution's own dataset using an in-house server.

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 · 1167 words · about 6 minRead it →Close

An Updated Deep Learning Tool Raised Rectal Cancer Outline Agreement From 0.88 to 0.95 Against Final Contours

In one Beijing hospital, software drafted the outlines for rectal cancer radiotherapy. The oncologists still reviewed every one.

Before radiation can be aimed at a rectal tumor, someone has to draw the tumor and the healthy organs around it on the planning scans. That drawing is critical and it is slow — the kind of work that eats an afternoon and varies from one doctor to the next¹.

In a study at the National Cancer Center in Beijing, a deep learning system took over the first draft. For each new patient, the planning CT images go from the treatment planning computer to a server, the model draws the outlines, and the draft comes back for the radiation oncologist to review and adjust by hand². The authors compared 150 rectal cancer patients across three phases — before the software existed, after the first version arrived, and after the second version replaced it — and added a separate set of 21 patients that six oncologists redrew three times each: once by hand, once with the first system, once with the second.

The update is where the numbers move. After the second version arrived, the agreement between the software's draft and the final approved outline rose from 0.88 to 0.93 for the tumor target and from 0.88 to 0.95 for the organs at risk³. The share of drafts that fell outside acceptable limits dropped from 3.30 percent to 0.64 percent⁴. Total working time fell from about 41 minutes by hand to about 17 minutes with the newer system⁵ — roughly 59 percent less than manual work and about 22 percent less than the first system⁶.

Two senior oncologists, kept blind to which contours came from where, scored the unedited drafts. The newer system averaged 4.02 out of 5 against 3.26 for the older one⁷. Nearly all of the newer drafts — 20 of 21 — scored 4 or above, against 6 of 21 for the older system⁸. And the six oncologists working with the newer system agreed with each other more closely than with either other method⁹.

What the study did not do is measure radiation dose. The authors say so themselves: dosimetric evaluation was not included¹⁰. Better-matched outlines are not yet shown to mean a better or safer treatment plan. The work was also retrospective, comparing three different groups of patients rather than the same patients followed forward, which the authors acknowledge may allow other differences to creep in¹¹. All of it happened at one hospital in China, under one scanning protocol and one set of contouring guidelines.

This belongs to a family of stories about software that drafts and a professional who finishes. Mammography screening has its own version: most European programs use two readers who check each other's work, and researchers are testing whether AI can serve as one of those readers¹²¹³. A review of three large studies embedded in national screening programs — MASAI, ScreenTrustCAD and PRAIM — looked at nearly 600,000 examinations and found a small absolute increase in cancers detected, on the order of one per thousand, without a consistent rise in women called back for extra tests¹⁴. We could read only the summary of that review; the full paper is behind a subscription¹⁵. The review's own conclusion is the same shape as the rectal cancer study's: AI belongs alongside the existing readers, and using it that way requires explicit quality checks and monitoring¹⁶.

The mechanism in both cases is the same and worth picturing. The software is not deciding anything. It produces a first draft — a set of lines on a scan — and a trained person decides whether those lines are right. In Beijing, every contour, regardless of which system drafted it, went through junior oncologists, then senior oncologists, then a departmental review before treatment². Nothing reached a patient unreviewed.

The study also catalogued when the software got things wrong. It mistook the levator ani muscle for an involved lymph node, because it only ever saw CT images and not the MRI that would have clarified the difference. It pulled loops of bowel next to the tumor into the target. It confused small intestine with colon. It read an enlarged prostate as bladder. It misread pelvic bone as femoral head. And it stumbled when a patient who could not lie face-down was scanned on their back instead¹¹. These are the failures that a team would need to know about before trusting any automatic draft.

Here is how we read it. When a drafting tool enters skilled work, the people using it tend to start from its version rather than from a blank page — so the tool's habits quietly become the department's habits. That means some of the measured improvement after the first system arrived may reflect changed human behavior rather than a better machine. We would expect the same drift in any clinic that adopts a tool like this: the final outlines slowly converge on whatever the software offers first. You would know we are wrong if hand-drawn outlines, made with no software in the room, stayed just as close to the software's drafts as the edited ones did. So when a hospital announces that its software improved its results, the question worth asking is whether the comparison was against the same people working the old way, or against people who had already started working around the software.

Here is a second thing we would watch. People tend to trust a newer version of anything partly because it is newer and because the people around them are using it. In this study the judges were kept blind to which system produced which draft, which is the right way to do it — but that is one hospital, one team, one set of eyes. Where you live, if a clinic tells you its contouring software has been upgraded, the useful question is not which version they bought. It is whether the people judging its output knew which version they were looking at. A blind comparison is the thing to ask for, and it costs nothing to request.

What this makes possible for you is narrower and more solid than it first sounds. If you or someone you love ever needs radiotherapy, the outlining step is one of the slowest parts of getting from scan to treatment, and it is where two doctors can reasonably disagree. Software that drafts that step faster and more consistently could mean less waiting and less variation between one clinician's plan and another's. It does not mean a different outcome from the treatment itself — nobody has shown that yet. So the sentence to keep is this one: ask who reviewed the software's draft before it became your plan, and ask what happens when the software gets it wrong. That is the question this study was built to answer, and it is the question you are entitled to ask anywhere.

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

Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "Accurate delineation of target volumes and organs‐at‐risk (OARs) is a critical yet labor‐intensive component of rectal cancer radiotherapy."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "In clinical practice, for each new patient, planning CT images are transferred from the treatment planning system (TPS) to the server, where the model performs automatic contouring. The predicted contours are then transmitted back to the TPS for review and manual adjustment by the radiation oncologist."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "The mean DSC values before and after updating were 0.88 ± 0.04 and 0.93 ± 0.04 for CTV (P < 0.001) and 0.88 ± 0.05 and 0.95 ± 0.02 for OARs (P < 0.001), respectively."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "After automatic contouring system update, the mean failure rate decreased from 3.30% to 0.64%."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "The total time significantly differed between the three contouring methods (P < 0.05), with the Auto2‐assisted method being the fastest (16.97 ± 5.52 min) and manual contouring being the most time‐consuming (41.15 ± 10.43 min)."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "Auto2‐assisted method decreased the total time by approximately 58.8% compared with the manual method, and 21.9% compared with the Auto1‐assisted method."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "In the blinded clinical evaluation, 99.2% (125/126) of the oncologist‐revised final contours received a Likert score of ≥ 4, and Auto2‐generated unedited contours showed significantly higher clinical acceptability than Auto1 (4.02 ± 0.25 vs. 3.26 ± 0.49, P < 0.001)"
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "In addition, 95.2% (20/21) of Auto2‐generated contours achieved a score of ≥ 4, compared with 28.6% (6/21) for Auto1."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "The Auto2‐assisted contouring group demonstrated significantly higher inter‐observer consistency than the other groups (all: P < 0.05)."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "First, although the present study comprehensively evaluated contouring efficiency, geometric performance, inter‐observer consistency, and blinded clinical acceptability, dosimetric evaluation was not included in the current work."
Wang, N., Xu, T., Kang, Y. et al. (2026). Real‐world clinical impact of implementing and updating a deep learning‐based automatic contouring system in rectal cancer radiotherapy. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70768 - the article this story is about — the whole article — the passage: "Second, the current study compared different patient cohorts across three phases. This longitudinal design was necessary to accurately reflect the real‐world evolution of clinical workflows. However, this approach may introduce confounding effects due to individual patient differences across the distinct cohorts."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — only the abstract - the full paper is behind a subscription — the passage: "Most European population mammography screening programs rely on double reading with arbitration, a model that delivers mortality benefit but is increasingly challenged by radiologist workload, variable specificity, and interval cancers."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — only the abstract - the full paper is behind a subscription — the passage: "Artificial intelligence (AI) is being evaluated to support or optimize these established European screening pathways."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — only the abstract - the full paper is behind a subscription — the passage: "In European population screening programs characterized by double reading and arbitration, prospective program-embedded evidence suggests that AI integration may yield a small absolute increase in cancer detection (≈1/1000) without a consistent increase in recall, alongside improved PPV and efficiency signals."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — only the abstract - the full paper is behind a subscription — the passage: "We performed a prespecified, focused evidence synthesis of three large studies embedded within routine population screening programs operating under European-relevant workflows: MASAI (randomized AI-supported risk triage within a national program), ScreenTrustCAD (prospective paired-reader evaluation with AI as an independent reader in a double-reading framework), and PRAIM (nationwide decision-referral implementation)."
Ferre R, Benefield T, Kuzmiak CM. (2026). Artificial intelligence–supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs. Clinical Imaging. 10.1016/j.clinimag.2026.110923 — only the abstract - the full paper is behind a subscription — the passage: "These findings suggestAI primarily as a complementary reader within European screening workflows, with implementation requiring explicit quality assurance and monitoring of interval cancers and stage distribution."

Do not take this as professional medical advice.