Caso Tobillo·DerechoThe Ankle·Case
● Experimento · lectura ciega por IA · modelos frontier● Experiment · blind AI reading · frontier models

¿Puede la IA de 2026 ver lo que vieron mis radiólogos?Can 2026 AI see what my radiologists saw?

Tres modelos frontier — Gemini 3.1 Pro, GPT-5.5 y Claude Fable 5.1 — leyeron la resonancia pre-quirúrgica de este caso (RM 31/05/2018) sin conocer los informes ni la historia clínica, y sus lecturas se compararon después contra el informe real del radiólogo y contra los hallazgos de la cirugía posterior. Resultado corto: ven la mayor parte del caso. Mejor lectura, 0.90 de sensibilidad; una lectura completamente a ciegas alcanza 0.80. Pero todos añaden hallazgos que el informe real niega — y ese detalle es la mitad de la historia. (Antes hubo una fase con dos modelos médicos open-source pequeños que no vieron el caso; queda resumida al final.) Nuevo: la fase 2 repite el protocolo con la RM de 2025 — y esta vez el resultado interesante no es lo que ven los modelos, sino que sus lecturas concuerdan mejor con el expediente 2018–2021 que el propio informe oficial de 8 líneas.

Three frontier models — Gemini 3.1 Pro, GPT-5.5 and Claude Fable 5.1 — read this case's pre-surgical MRI (May 31, 2018) with no access to the reports or the clinical history; their readings were then compared against the radiologist's actual report and against the findings of the subsequent surgery. Short answer: they see most of the case. Best reading, 0.90 sensitivity; one fully blind reading reaches 0.80. But all of them add findings the actual report denies — and that detail is half the story. (An earlier phase with two small open-source medical models, which couldn't see the case, is summarized at the end.) New: phase 2 repeats the protocol on the 2025 MRI — and this time the interesting result is not what the models see, but that their readings agree better with the 2018–2021 record than the official 8-line report itself.

3×2Modelos × brazosModels × arms
0.90Mejor sensibilidad (de 1.0)Best sensitivity (out of 1.0)
102Imágenes RM analizadasMRI images analyzed
~$8.5Coste de cómputo (API)Compute cost (API)

PlanteamientoRationale

La pregunta y las reglasThe question and the rules

Este caso tiene algo poco común: verdad de referencia quirúrgica. La RM del 31/05/2018 fue informada por un radiólogo y, cinco semanas después, una artroscopia confirmó por visión directa las lesiones. Eso permite hacer la pregunta al revés: si una IA de frontera de 2026 lee esa misma resonancia a ciegas, ¿encuentra lo mismo?

This case has something unusual: surgical ground truth. The May 31, 2018 MRI was reported by a radiologist and, five weeks later, an arthroscopy confirmed the lesions under direct vision. That allows asking the question in reverse: if a 2026 frontier AI reads that same MRI blind, does it find the same things?

Reglas fijadas antes de empezarRules set before starting

  • Solo imágenes anonimizadas. Los modelos reciben un set neutro de PNGs sin nombres, fechas ni centro; los DICOM originales nunca salen del equipo propio.
  • Dos brazos por modelo. Brazo A, ciego total: solo «serie X de una RM de tobillo derecho». Brazo B, contexto mínimo — lo que recibe un radiólogo real en el volante: «varón, 30s, dolor crónico de tobillo derecho, antecedente de esguince». Sin mencionar ligamentos ni cartílago.
  • Outputs generados antes de compararlos con el ground truth. Primero se guardan las lecturas; solo después se lee el informe real y se puntúa, con criterios definidos por adelantado.
  • Se publica salga lo que salga — incluido un resultado negativo.
  • Anonymized images only. The models receive a neutral set of PNGs with no names, dates or facility; the original DICOM never leaves my own machines.
  • Two arms per model. Arm A, fully blind: only "series X from a right ankle MRI study". Arm B, minimal context — what a real radiologist gets on the referral slip: "male patient in his 30s, chronic right ankle pain, remote history of an ankle sprain". No mention of ligaments or cartilage.
  • Outputs generated before comparing against the ground truth. The readings are saved first; only then is the actual report read and scored, with criteria defined in advance.
  • Published whatever the outcome — including a negative result.
MétodoMethod

Pipeline, modelos y protocoloPipeline, models and protocol

DatosData

DICOM original propio → NIfTI (dcm2niix) → PNG por serie: máximo 24 cortes por serie con muestreo uniforme, ventana de intensidad por percentiles 1–99, nombres neutros sin fechas ni centro, localizers excluidos. Para esta fase se usó únicamente el estudio pre-quirúrgico: 5 series, 102 imágenes, 11 MB. Verificación de datos personales en lo subido: cero.

My own original DICOM → NIfTI (dcm2niix) → PNGs per series: at most 24 slices per series with uniform sampling, intensity windowing at percentiles 1–99, neutral filenames with no dates or facility, localizers excluded. Only the pre-surgical study was used for this phase: 5 series, 102 images, 11 MB. Personal-data check on the uploaded set: zero hits.

ModelosModels

Gemini 3.1 ProGPT-5.5Claude Fable 5.1
origen / originGoogleOpenAIAnthropic
acceso / accessAPI vía OpenRouter, mismas llamadas y prompts para los tresAPI via OpenRouter, same calls and prompts for all three

EjecuciónExecution

Temperatura 0.1. Por estudio: una llamada por serie (todos sus cortes como imágenes) y una llamada final de síntesis, solo texto, con las observaciones propias del modelo. Los tres son modelos razonadores y su razonamiento interno consume el presupuesto de tokens: con el límite original de 1500, dos devolvían respuesta vacía y el tercero un texto truncado a mitad de frase — el primer intento (~$2.9) se invalidó y descartó íntegro, y se repitió con max_tokens = 16000. Coste total, primer intento incluido: ~$8.5.

Temperature 0.1. Per study: one call per series (all its slices as images) and one final text-only synthesis call fed with the model's own observations. All three are reasoning models and their internal reasoning consumes the token budget: with the original 1500-token cap, two returned empty responses and the third a mid-sentence truncation — the first attempt (~$2.9) was invalidated and discarded entirely, and the run was repeated with max_tokens = 16000. Total cost, first attempt included: ~$8.5.

Prompt de serie (idéntico para los tres modelos; el brazo B antepone el contexto clínico):Series prompt (identical for all three models; arm B prepends the clinical context):

You are an expert musculoskeletal radiologist. [context if arm B]
You are shown {n} slices (ordered, uniformly sampled) of the series
'{series}' from a right ankle MRI study. Report ALL abnormal findings
you can identify: bones, cartilage, ligaments, tendons, joint fluid,
soft tissues. For each finding give location, description and your
confidence (high/medium/low). If a structure looks normal, you may
say so briefly. Do not invent findings you cannot see.
Verdad de referenciaGround truth

Lo que había que encontrarWhat there was to find

Del informe real de la RM 31/05/2018 (PDF anonimizado), cinco hallazgos positivos y tres negativos explícitos:

From the actual May 31, 2018 MRI report (anonymized PDF), five positive findings and three explicit negatives:

IDHallazgo del radiólogoRadiologist's finding
G1Rotura crónica del ligamento peroneoastragalino anterior (LPAA)Chronic tear of the anterior talofibular ligament (ATFL)
G2Cambios degenerativos anteromediales tibioastragalinos con osteofitos marginalesAnteromedial tibiotalar degenerative changes with marginal osteophytes
G3Edema óseo en ambas vertientes articularesBone marrow edema on both articular surfaces
G4Ligera sinovitisMild synovitis
G5Sospecha de pinzamiento anteromedialSuspected anteromedial impingement
N1–N3Negativos: sin derrame; resto de ligamentos y tendones normales; Aquiles y fascia plantar normalesNegatives: no effusion; remaining ligaments and tendons normal; Achilles and plantar fascia normal

Y la confirmación quirúrgica (artroscopia, 05/07/2018): úlcera condral en el pilón tibial anterior + osteofitos astragalino anteromedial y tibial anterior, resecados. La rotura del LPAA no se abordó en esa cirugía.

Plus the surgical confirmation (arthroscopy, July 5, 2018): chondral ulcer on the anterior tibial plafond + anteromedial talar and anterior tibial osteophytes, resected. The ATFL tear was not addressed in that surgery.

Matiz importante sobre la úlcera condral: no aparece en el informe de la RM — ni el radiólogo humano la vio en la imagen. Solo se descubrió por visión directa durante la artroscopia. Por eso no forma parte de los hallazgos puntuables (G1–G5): a los modelos solo se les exige lo que la propia radiología informó. Si alguno la describiera, contaría como mérito extra, no como obligación.

Important nuance about the chondral ulcer: it does not appear in the MRI report — the human radiologist did not see it on the images either. It was only discovered under direct vision during the arthroscopy. That is why it is not part of the scoreable findings (G1–G5): the models are only held to what radiology itself reported. Describing it would count as extra merit, not an obligation.

ResultadosResults

Qué vio cada lecturaWhat each reading saw

Puntuación: SÍ = 1, PARCIAL = 0.5 (patología correcta pero incompleta o mal localizada dentro de la misma articulación), NO = 0. Se puntúa solo la síntesis final de cada lectura; criterios fijados antes de puntuar.

Scoring: YES = 1, PARTIAL = 0.5 (correct pathology but incomplete or mislocalized within the same joint), NO = 0. Only each reading's final synthesis is scored; criteria were set before scoring.

HallazgoFindingGemini A
(ciego)(blind)
Gemini B
(contexto)(context)
GPT-5.5 A
(ciego)(blind)
GPT-5.5 B
(contexto)(context)
Fable A
(ciego)(blind)
Fable B
(contexto)(context)
G1 rotura LPAAATFL tearSÍ — «high-grade tear», sin cronicidadYES — "high-grade tear", no chronicitySÍ — «chronic tear and extensive scarring»YES — "chronic tear and extensive scarring"PARCIAL — «sprain/partial tear, possibly chronic»PARTIAL — "sprain/partial tear, possibly chronic"SÍ — «chronic sprain/tear/scarring»YES — "chronic sprain/tear/scarring"PARCIAL — esguince, cronicidad erradaPARTIAL — sprain, chronicity wrongSÍ — «chronic ATFL tear with scarring»YES — "chronic ATFL tear with scarring"
G2 osteofitos / degeneración anteromedialosteophytes / anteromedial degenerationNOSÍ — «kissing spurs»YES — "kissing spurs"SÍ — osteofitos plafond + cuello talar, EN CIEGOYES — plafond + talar-neck osteophytes, BLINDSÍ — «kissing anterior ankle spurs»YES — "kissing anterior ankle spurs"SÍ — osteofito tibial anteriorYES — anterior tibial osteophyteSÍ — «kissing osteophytes»YES — "kissing osteophytes"
G3 edema óseo ambas vertientesbilateral bone marrow edemaPARCIAL — solo talarPARTIAL — talar onlyPARCIAL — solo talarPARTIAL — talar onlyPARCIAL — solo talarPARTIAL — talar onlyPARCIAL — solo talarPARTIAL — talar onlyPARCIAL — solo talarPARTIAL — talar onlyPARCIAL — solo talarPARTIAL — talar only
G4 sinovitissynovitisNO — solo derrameNO — effusion onlyNO — scarring, no sinovitisNO — scarring, not synovitis
G5 pinzamiento anteromedialanteromedial impingementNOSÍ — impresión nº1YES — impression #1SÍ — «anterior impingement morphology»YES — "anterior impingement morphology"NO — dice pinzamiento POSTERIORNO — calls it POSTERIOR impingementSÍ — «early anterior impingement»YES — "early anterior impingement"
SensibilidadSensitivity0.300.700.800.900.600.90
Falsos positivosFalse positives465574

Falsos positivos, en detalleFalse positives, in detail

  • Gemini A: derrame; tendinosis aquílea «severa» con confianza alta — el peor falso positivo del experimento: el informe real declara el Aquiles normal, y también todos los informes posteriores (2018–2021); esguince sindesmal; tenosinovitis peroneal.
  • Gemini B: derrame; rotura crónica del CFL; esguince sindesmal remoto; tenosinovitis FHL; quiste intraóseo talar; osteofitos talonaviculares.
  • GPT-5.5 A: derrame; tenosinovitis FHL; fascitis plantar; edema/ganglión del seno del tarso; posible esguince deltoideo.
  • GPT-5.5 B: derrame; tenosinovitis peroneal y FHL; fibrosis del seno del tarso con posible coalición; posibles cuerpos libres; degeneración subastragalina.
  • Fable A: derrame; os trigonum con pinzamiento posterior; síndrome del seno del tarso; degeneración subastragalina con quistes; tenosinovitis múltiple; esguince deltoideo; edema subcutáneo difuso — la lectura más sobrellamada del experimento.
  • Fable B: derrame; tenosinovitis peroneal y FHL; esguince deltoideo crónico; sinovitis del seno del tarso.
  • Gemini A: effusion; "severe" Achilles tendinosis at high confidence — the worst false positive of the experiment: the actual report declares the Achilles normal, as do all later reports (2018–2021); syndesmotic sprain; peroneal tenosynovitis.
  • Gemini B: effusion; chronic CFL tear; remote syndesmotic sprain; FHL tenosynovitis; intraosseous talar cyst; talonavicular osteophytes.
  • GPT-5.5 A: effusion; FHL tenosynovitis; plantar fasciitis; sinus tarsi edema/ganglion; possible deltoid sprain.
  • GPT-5.5 B: effusion; peroneal and FHL tenosynovitis; sinus tarsi fibrosis with possible coalition; possible loose bodies; subtalar degeneration.
  • Fable A: effusion; os trigonum with posterior impingement; sinus tarsi syndrome; subtalar degeneration with cysts; multiple tenosynovitis; deltoid sprain; diffuse subcutaneous edema — the most overcalled reading of the experiment.
  • Fable B: effusion; peroneal and FHL tenosynovitis; chronic deltoid sprain; sinus tarsi synovitis.

Negativos explícitos del informe: N1 (sin derrame): 0/6 — las seis lecturas afirman derrame. N2 (resto de ligamentos y tendones normales): 0/6. N3 (Aquiles y fascia normales): 4/6.Explicit negatives from the report: N1 (no effusion): 0/6 — all six readings assert an effusion. N2 (remaining ligaments and tendons normal): 0/6. N3 (Achilles and plantar fascia normal): 4/6.

Nota sobre el derrame (por qué cuenta como falso positivo, con asterisco)A note on the effusion (why it counts as a false positive, with an asterisk)

El informe de 2018 dice explícitamente «no se observa derrame articular», y la verdad de referencia se congeló antes de puntuar — así que el derrame computa como falso positivo en las seis lecturas. Pero merece registro: los estudios posteriores del mismo tobillo sí describen líquido articular (eco 01/2019, RM 02/2019, RM 09/2020, RM 03/2021, todos en pequeña cuantía). Los modelos leen la imagen de 2018, no las de después, y el criterio se mantiene; el apoyo evolutivo queda anotado. La tenosinovitis, en cambio, no tiene atenuante: todos los informes 2018–2021 dan los tendones normales.

The 2018 report explicitly states "no joint effusion is seen", and the ground truth was frozen before scoring — so the effusion counts as a false positive in all six readings. But it deserves a note: later studies of the same ankle do describe joint fluid (US 01/2019, MRI 02/2019, MRI 09/2020, MRI 03/2021, all small amounts). The models read the 2018 images, not the later ones, so the criterion stands; the evolutionary support is recorded. The tenosynovitis, by contrast, has no such mitigation: every report from 2018–2021 calls the tendons normal.

El caso frontera: la «lesión osteocondral» del domo talarThe borderline call: the talar dome "osteochondral lesion"

Cinco de las seis lecturas afirman o sugieren una lesión (osteo)condral del domo talar medial. La excepción es Fable A: la única que la niega explícitamente («no osteochondral lesion identified») — en línea con la RM de 2021, que descartó lesiones osteocondrales en la cúpula astragalina. Criterio aplicado: la señal subcondral talar computa dentro de G3 como PARCIAL, no como falso positivo. Fable B, por su parte, descartó activamente un os trigonum, una coalición y un artefacto de supresión grasa — trabajo de especificidad que el recuento de FPs no captura.

Five of the six readings assert or suggest a medial talar dome (osteo)chondral lesion. The exception is Fable A: the only one that explicitly denies it ("no osteochondral lesion identified") — in line with the 2021 MRI, which ruled out osteochondral lesions of the talar dome. Criterion applied: the talar subchondral signal counts within G3 as PARTIAL, not as a false positive. Fable B, for its part, actively ruled out an os trigonum, a coalition and a fat-suppression artifact — specificity work the FP count does not capture.

Las tres hipótesis, contrastadasThe three hypotheses, put to the test

  • H1 (¿alguien verá los osteofitos y/o la úlcera condral?): confirmada para los osteofitos — 5/6 lecturas los nombran; GPT-5.5 incluso en ciego total. La úlcera del pilón tibial sigue sin nombrarla nadie — coherente con que tampoco estaba en el informe de la RM (solo se vio en cirugía); GPT-5.5 B es quien más se acerca: «mild cartilage thinning… greatest anteriorly».
  • H2 (nadie detectará la rotura del LPAA): refutada — 4/6 la llaman rotura; Gemini A es la única lectura en ciego que llega a llamarla rotura.
  • H3 (el contexto mejora sin aumentar la alucinación): se sostiene — la sensibilidad sube en los tres (+0.40, +0.10, +0.30) y los falsos positivos no aumentan con contexto (Gemini es la excepción, 4→6; GPT 5→5; Fable 7→4).
  • H1 (will anyone see the osteophytes and/or the chondral ulcer?): confirmed for the osteophytes — 5/6 readings name them; GPT-5.5 even fully blind. The tibial plafond ulcer remains unnamed by everyone — consistent with it being absent from the MRI report too (it was only seen at surgery); GPT-5.5 B comes closest: "mild cartilage thinning… greatest anteriorly".
  • H2 (nobody will detect the ATFL tear): refuted — 4/6 call it a tear; Gemini A is the only blind reading that goes as far as calling it a tear.
  • H3 (context improves detection without more hallucination): holds — sensitivity rises in all three (+0.40, +0.10, +0.30) and false positives do not increase with context (Gemini is the exception, 4→6; GPT 5→5; Fable 7→4).

Las seis impresiones diagnósticas, literalesThe six diagnostic impressions, verbatim

Sección IMPRESSION de cada síntesis, sin editar (los FINDINGS completos y las observaciones por serie, en los ficheros enlazados).The IMPRESSION section of each synthesis, unedited (full FINDINGS and per-series observations in the linked files).

Gemini 3.1 Pro · brazo A (ciego)arm A (blind) · sens. 0.30 · 4 FP
1. Osteochondral lesion of the medial talar dome: Focal defect with subchondral cystic changes, marrow edema, and fluid undermining the osteochondral fragment, highly suggestive of an unstable lesion. (Confidence: High)
2. Lateral ankle sprain: Complete or high-grade partial tear of the anterior talofibular ligament (ATFL) with extensive associated anterolateral subcutaneous edema. (Confidence: High)
3. Syndesmotic sprain: Sprain of the anterior inferior tibiofibular ligament (AITFL). (Confidence: High)
4. Achilles tendinopathy: Severe, diffuse Achilles tendinosis without MRI evidence of an acute full-thickness tear. (Confidence: High)
5. Peroneal tenosynovitis: Moderate fluid within the common peroneal tendon sheath without evidence of high-grade tendon tear. (Confidence: High)
6. Joint effusion: Moderate to large complex tibiotalar joint effusion. (Confidence: High)

sintesis.md · JSON: T1 ax · PD ax · PD sag · T1 sag · MERGE cor

Gemini 3.1 Pro · brazo B (contexto)arm B (context) · sens. 0.70 · 6 FP
1. Anterior and Anterolateral Ankle Impingement: Prominent soft tissue scarring/thickening in the anterolateral gutter and anterior capsular recess ("meniscoid lesion"), combined with anterior tibiotalar "kissing" osteophytes. This is a classic sequela of prior trauma and a primary generator of chronic pain. (Confidence: High)
2. Chronic Lateral Ligamentous Injury: Markedly thickened and scarred anterior talofibular ligament (ATFL) and calcaneofibular ligament (CFL), consistent with chronic tears from the patient's remote sprain. Mild thickening of the AITFL suggests a remote anterior syndesmotic sprain. (Confidence: High)
3. Osteochondral Lesion (OCL): Focal OCL of the medial talar dome with a subchondral cyst, overlying cartilage defect, and reactive bone marrow edema. (Confidence: High)
4. Tenosynovitis and Effusion: Flexor hallucis longus (FHL) tenosynovitis and a small to moderate tibiotalar joint effusion. (Confidence: High)
5. Incidental Osseous Findings: Small intraosseous cyst/ganglion in the central talar body and mild dorsal talonavicular osteophytosis. (Confidence: High)

sintesis.md · JSON: T1 ax · PD ax · PD sag · T1 sag · MERGE cor

GPT-5.5 · brazo A (ciego)arm A (blind) · sens. 0.80 · 5 FP
1. Small medial talar dome/medial talar shoulder osteochondral lesion with focal subchondral/articular surface irregularity and suspected cartilage thinning/fissuring. No displaced fragment. Confidence: Medium.
2. Mild-to-moderate tibiotalar joint effusion/synovitis and small posterior subtalar effusion. Confidence: High.
3. ATFL abnormality suspicious for sprain/partial tear, possibly chronic/scarred; no definite complete tear. Mild possible chronic deltoid sprain/scarring. Confidence: Medium for ATFL; Low-to-medium for deltoid.
4. Mild sinus tarsi edema/fluid, possibly reflecting mild synovitis/scarring or small ganglion/synovial cyst component. Confidence: Medium.
5. Mild FHL tenosynovitis versus fluid communication from adjacent joint effusion. Trace peroneal tendon sheath fluid without tendon tear. Confidence: Medium.
6. Mild anterior tibiotalar degenerative/anterior impingement morphology and mild proximal plantar fascial enthesopathic/fasciitis-type change. Confidence: Medium.

sintesis.md · JSON: T1 ax · PD ax · PD sag · T1 sag · MERGE cor

GPT-5.5 · brazo B (contexto)arm B (context) · sens. 0.90 · 5 FP
1. Chronic osteochondral lesion of the medial/posteromedial talar dome/shoulder, with small subchondral cystic/low T1 change and mild articular surface irregularity. No displaced osteochondral fragment identified. Confidence: Medium.
2. Chronic lateral ankle ligament injury, greatest involving the ATFL, which is attenuated/irregular and poorly defined, with associated anterolateral gutter scarring/synovitis. Findings may contribute to chronic anterolateral ankle impingement symptoms. Confidence: Medium-high.
3. Small to moderate tibiotalar effusion with mild synovitis. Confidence: High.
4. Chronic anterior tibiotalar degenerative/impingement-type change with anterior distal tibial and talar neck osteophytes; possible small chronic anterior osteophyte fragments/intra-articular bodies. Confidence: High for osteophytes; Low-medium for small bodies/fragments.
5. Sinus tarsi scarring/fibrosis with mild subtalar degenerative/synovitis-type change. Narrowing/irregularity of the middle subtalar facet raises the possibility of fibrocartilaginous talocalcaneal coalition versus chronic subtalar degeneration, not definitive. Confidence: Medium.
6. Mild peroneal tenosynovitis and possible mild FHL tenosynovitis/communication with ankle joint fluid. Tendons remain intact. Confidence: Medium.

sintesis.md · JSON: T1 ax · PD ax · PD sag · T1 sag · MERGE cor

Claude Fable 5.1 · brazo A (ciego)arm A (blind) · sens. 0.60 · 7 FP
1. Lateral ankle ligament sprain (ATFL ± CFL) with anterolateral capsular and lateral periligamentous edema; discrete fiber discontinuity not confidently demonstrated – acute/subacute injury pattern. (medium–high confidence)
2. Moderate tibiotalar joint effusion with synovitis, most pronounced in the posterior recess, where nodular synovial thickening is present adjacent to a posterior talar process ossicle (os trigonum/Stieda fragment) showing abnormal synchondrosis signal – findings suggestive of posterior ankle impingement; small intra-articular bodies not excluded. (effusion high; os trigonum/impingement medium)
3. Sinus tarsi edema/synovitis with poorly defined interosseous ligament, in keeping with sinus tarsi syndrome pattern, likely related to the lateral ligament injury. (medium)
4. Posterior subtalar joint degenerative change with fluid and small subchondral cysts of the talus and calcaneus. (medium)
5. Medial flexor (FHL > PT/FDL) tenosynovitis with fluid extending along the deep posterior compartment, largely communicating with the ankle joint; mild peroneal tenosynovitis. No tendon tear. (medium; peroneal low–medium)
6. Possible low-grade deltoid sprain and small amount of syndesmotic fluid without definite tibiofibular ligament tear. (low)
7. Possible subtle medial talar body marrow edema/contusion; small anterior tibial osteophyte. No fracture or osteochondral lesion of the talar dome. (low)
8. Diffuse subcutaneous edema of the hindfoot/midfoot (lateral and plantar predominance) and plantar heel fat pad edema without plantar fascial thickening. (high for subcutaneous edema; medium for fat pad)
9. Achilles tendon normal.

sintesis.md · JSON: T1 ax · PD ax · PD sag · T1 sag · MERGE cor

Claude Fable 5.1 · brazo B (contexto)arm B (context) · sens. 0.90 · 4 FP
1. Chronic anterior talofibular ligament tear with scarring/attenuation and anterolateral gutter synovitis, in keeping with chronic lateral ankle instability from the remote sprain. CFL and PTFL intact. (High confidence – concordant on three sequences.)
2. Mild-to-moderate tibiotalar effusion with synovitis and small posterior subtalar effusion, with fluid/synovitis extending into the sinus tarsi and tarsal canal (interosseous ligament intact, fat largely preserved – early/mild sinus tarsi synovitis rather than established sinus tarsi syndrome). (High confidence for effusion; medium for sinus tarsi synovitis.)
3. Small medial talar dome chondral/early osteochondral lesion without fragment displacement or cystic change (no T1 marrow replacement). Equivocal subtle irregularity of the lateral dome, indeterminate. (Medium-low confidence.)
4. Early anterior tibiotalar impingement: small anterior tibial plafond osteophyte with anterior talar neck irregularity and anterior capsular/synovial thickening. (Medium-low confidence.)
5. Mild peroneal and flexor hallucis longus tenosynovitis; all tendons intact without tear, split or subluxation. FHL sheath fluid may partly reflect decompression of the joint effusion. (Medium confidence.)
6. Mild deep deltoid ligament thickening/heterogeneity, compatible with chronic sprain sequela; no discrete tear. (Low confidence.)
7. No fracture, marrow-replacing lesion, os trigonum, syndesmotic injury, Achilles or plantar fascia abnormality, or soft-tissue mass. Supramalleolar posterior compartment hyperintensity on axial PD FS is attributed to peripheral fat-suppression artifact (no T1 correlate). (Medium-high confidence.)

sintesis.md · JSON: T1 ax · PD ax · PD sag · T1 sag · MERGE cor

Fase 2 · RM 2025Phase 2 · 2025 MRI

Siete años después: la resonancia de 2025Seven years later: the 2025 MRI

Mismo pipeline y mismos tres modelos, ahora sobre la RM del 15/09/2025 (sin contraste; 6 series, 132 imágenes) — el estudio de la reactivación del caso. Novedad de diseño: tres brazos. A, ciego total; B, el mismo volante mínimo de la fase anterior; C, ese volante más una sola línea nueva: "previous ankle arthroscopy (2018, osteophyte resection)". Nueve lecturas, coste ~$9.8. Y una circunstancia que acabó siendo parte del experimento: al congelar los outputs, el informe oficial de esta RM no estaba en el archivo (pendiente de reclamar al hospital) — el análisis de consenso de abajo se redactó sin conocerlo, y el informe apareció en papel horas después.

Same pipeline and same three models, now on the Sept 15, 2025 MRI (no contrast; 6 series, 132 images) — the study from the case's reactivation. Design novelty: three arms. A, fully blind; B, the same minimal referral slip as the previous phase; C, that slip plus a single new line: "previous ankle arthroscopy (2018, osteophyte resection)". Nine readings, cost ~$9.8. Plus a circumstance that ended up being part of the experiment: when the outputs were frozen, the official report of this MRI was not in the archive (pending recovery from the hospital) — the consensus analysis below was written without knowing it, and the report surfaced on paper hours later.

Lo que las nueve lecturas dicen (consenso, redactado antes de leer el informe)What the nine readings say (consensus, written before reading the report)

HallazgoFindingLecturas que lo reportanReadings reporting it
Derrame tibioastragalino + sinovitisTibiotalar effusion + synovitis9/9
LPAA crónico roto/cicatricialChronic torn/scarred ATFL9/9
Tenosinovitis (peroneos y/o FHL y/o tibial posterior)Tenosynovitis (peroneal and/or FHL and/or tibialis posterior)9/9
Seno del tarso (edema/fibrosis)Sinus tarsi (edema/fibrosis)8/9
Lesión (osteo)condral / anomalía subcondral del domo talar(Osteo)chondral lesion / subchondral abnormality of the talar dome8/9
Pinzamiento anterior con osteofitosAnterior impingement with osteophytes7/9

Efecto del brazo. En ciego, cada modelo alucina una narrativa distinta e incompatible con las demás: Gemini A ve tendinosis aquílea y tibial posterior «severas» más un patrón de edema difuso que sugiere CRPS; GPT-5.5 A afirma una coalición talocalcánea «high confidence» que su propio brazo B degrada a «no confirmada» y su brazo C ni menciona; Fable A es el más conservador y degrada hallazgos dudosos a artefacto. Con contexto (B), los tres convergen en la misma historia. Y en el brazo C los tres releen los osteofitos anteriores como residuales/recurrentes tras la resección («recurrent», Gemini; «residual or recurrent», GPT; «post-arthroscopic… residual/recurrent», Fable) — pero ninguno detectó por la imagen que el tobillo estaba operado: ni A ni B mencionan cambios postquirúrgicos. El dato quirúrgico no añade hallazgos; reinterpreta los mismos.

Arm effect. Blind, each model hallucinates a different, mutually incompatible narrative: Gemini A sees "severe" Achilles and tibialis posterior tendinosis plus a diffuse edema pattern suggesting CRPS; GPT-5.5 A asserts a talocalcaneal coalition at "high confidence" that its own arm B downgrades to "not confirmed" and its arm C never mentions; Fable A is the most conservative, downgrading doubtful findings to artifact. With context (B), all three converge on the same story. And in arm C, all three re-read the anterior osteophytes as residual/recurrent after the resection ("recurrent", Gemini; "residual or recurrent", GPT; "post-arthroscopic… residual/recurrent", Fable) — yet none of them detected from the images that the ankle had been operated on: neither A nor B mentions post-surgical change. The surgical datum adds no findings; it reinterprets the same ones.

El ground truth: un informe de 8 líneasThe ground truth: an 8-line report

El informe oficial (PDF anonimizado, firmado el 22/09/2025) despacha las 6 series / 159 cortes en 8 líneas: todo explícitamente normal — articulaciones, maléolos, huesos del tarso, seno del tarso, tendones («sin hallazgos que sugieran rotura o tenosinovitis»), fascia plantar — salvo un único positivo: pequeño derrame articular en el receso tibioastragalino anterior y posterior, que es también toda su conclusión. Sin comparación con ninguno de los cinco estudios previos. Matiz de redacción que importa: dice «no signos de rotura ligamentosa aguda» — no se pronuncia sobre la rotura crónica del LPAA documentada desde 2018, así que las lecturas que dicen «chronic ATFL tear» no lo contradicen y no puntúan como fallo.

The official report (anonymized PDF, signed Sept 22, 2025) dispatches the 6 series / 159 slices in 8 lines: everything explicitly normal — joints, malleoli, tarsal bones, sinus tarsi, tendons ("no findings suggesting tear or tenosynovitis"), plantar fascia — except one single positive: small joint effusion in the anterior and posterior tibiotalar recess, which is also its entire conclusion. No comparison with any of the five prior studies. A wording nuance that matters: it says "no signs of acute ligament tear" — it takes no position on the chronic ATFL tear documented since 2018, so readings saying "chronic ATFL tear" do not contradict it and are not scored as errors.

Scoring contra ese informe. Sensibilidad: trivial — el derrame es el único positivo y las nueve lecturas lo reportan, 9/9 (aunque ocho lo gradúan «small-to-moderate» o «moderate» frente al «pequeño» del informe). Lo que discrimina es la especificidad: violaciones de los cuatro negativos explícitos comprobables (✗ = la lectura afirma patología donde el informe dice normal):

Scoring against that report. Sensitivity: trivial — the effusion is the only positive and all nine readings report it, 9/9 (though eight grade it "small-to-moderate" or "moderate" versus the report's "small"). What discriminates is specificity: violations of the four checkable explicit negatives (✗ = the reading asserts pathology where the report says normal):

LecturaReadingN1 articulacionesjointsN2 huesoboneN3 seno del tarsosinus tarsiN4 tendonestendonsViolacionesViolations
Gemini A3/4
Gemini B3/4
Gemini C4/4
GPT-5.5 A4/4
GPT-5.5 B4/4
GPT-5.5 C4/4
Fable A3/4
Fable B4/4
Fable C3/4

Media: 3.6 de 4 negativos violados por lectura. Los pocos aciertos de especificidad tienen nombre: Gemini B es la única lectura que da los tendones por normales; Fable A la única que respeta las articulaciones; Fable C la única que, con el dato quirúrgico en la mano, se niega a afirmar la lesión osteocondral («no definite osteochondral lesion, no marrow oedema»); y el ✓ de Gemini A en el seno del tarso es por omisión, no por negación.

Average: 3.6 of 4 negatives violated per reading. The few specificity wins have names: Gemini B is the only reading that calls the tendons normal; Fable A the only one that respects the joints; Fable C the only one that, holding the surgical datum, refuses to assert the osteochondral lesion ("no definite osteochondral lesion, no marrow oedema"); and Gemini A's sinus-tarsi ✓ is by omission, not denial.

El resultado de verdad no es ese número — es que este ground truth contradice el propio expediente. El informe de 2025 da por normales estructuras con patología documentada y nunca desmentida: la rotura completa crónica del LPAA (seis documentos, 2018–2021; una rotura crónica no se regenera) y la artrosis anteromedial con osteofitos (RM 2020 y 2021 «sin cambios»; osteofitos además vistos y resecados en la cirugía de 2018). Es decir: las lecturas B y C de los tres modelos concuerdan mejor con la historia documentada 2018–2021 del paciente que el propio informe oficial de 2025. Las dos interpretaciones posibles quedan registradas: (a) el informe 2025 es una infra-lectura — consistente con la tesis de este caso; (b) los modelos sobrellaman sistemáticamente — consistente con los falsos positivos de la fase anterior (derrame, tenosinovitis, seno del tarso: exactamente las mismas categorías). Probablemente ambas a la vez. La afirmación más expuesta de los modelos es la lesión osteocondral del domo talar (8/9 lecturas): la niegan explícitamente la RM de 2021 y la de 2025; a su favor solo tiene la úlcera condral vista en la cirugía de 2018 — que estaba en la tibia, no en el domo.The real result is not that number — it is that this ground truth contradicts the patient's own record. The 2025 report calls normal structures with documented, never-retracted pathology: the chronic complete ATFL tear (six documents, 2018–2021; a chronic tear does not regenerate) and the anteromedial arthrosis with osteophytes (2020 and 2021 MRI "unchanged"; osteophytes moreover seen and resected at the 2018 surgery). In other words: the B and C readings of all three models agree better with the patient's documented 2018–2021 history than the official 2025 report itself. Both possible interpretations are recorded: (a) the 2025 report is an under-read — consistent with this case's thesis; (b) the models systematically overcall — consistent with the previous phase's false positives (effusion, tenosynovitis, sinus tarsi: exactly the same categories). Probably both at once. The models' most exposed claim is the talar dome osteochondral lesion (8/9 readings): both the 2021 and the 2025 MRI explicitly deny it; its only support is the chondral ulcer seen at the 2018 surgery — which was on the tibia, not the dome.

Las nueve impresiones diagnósticas, literalesThe nine diagnostic impressions, verbatim

Sección IMPRESSION de cada síntesis, sin editar (FINDINGS completos y observaciones por serie, en los ficheros enlazados).The IMPRESSION section of each synthesis, unedited (full FINDINGS and per-series observations in the linked files).

Gemini 3.1 Pro · brazo A (ciego)arm A (blind) · neg. 3/4
1. Lateral Ankle Injury: High-grade sprain or partial tear of the anterior talofibular ligament (ATFL) with extensive adjacent lateral soft tissue edema. (Confidence: High)
2. Medial Tendon and Ligament Pathology: Severe tendinosis and tenosynovitis of the posterior tibial tendon, accompanied by a sprain of the adjacent deltoid ligament complex. (Confidence: High)
3. Achilles Tendinopathy: Severe chronic Achilles tendinosis, characterized by marked fusiform thickening and intrasubstance signal alteration, without evidence of acute full-thickness tear. (Confidence: High)
4. Diffuse Tarsal Bone Marrow Edema: Widespread, patchy bone marrow edema involving multiple tarsal bones without a discrete acute fracture. Accompanied by extensive regional soft tissue and plantar intrinsic muscle edema, this pattern is non-specific but may represent a diffuse stress response, reactive edema from adjacent severe soft tissue inflammation, or complex regional pain syndrome (CRPS). Clinical correlation is recommended. (Confidence: High for finding; Medium for specific etiology)
5. Peroneal Tenosynovitis: Fluid within the common peroneal tendon sheath without evidence of underlying tendon tear. (Confidence: High)
6. Osteoarthritis and Joint Fluid: Medial compartment tibiotalar osteoarthritis with marginal osteophytosis. Small to moderate tibiotalar joint effusion and trace subtalar/talonavicular effusions. (Confidence: High)
7. Incidental Osseous Finding: Small os subfibulare or old, ununited fibular avulsion fracture. (Confidence: High)

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

Gemini 3.1 Pro · brazo B (contexto)arm B (context) · neg. 3/4
1. Chronic Lateral Ankle Injury and Anterolateral Impingement: Markedly thickened and irregular ATFL and CFL consistent with chronic tearing and scarring from the remote ankle sprain. This is associated with an old, ununited avulsion fracture at the lateral malleolar tip (os subfibulare) and adjacent soft tissue thickening in the anterolateral gutter, highly suggestive of anterolateral impingement syndrome. (Confidence: High)
2. Active Osteochondral Lesion: Focal osteochondral lesion of the talar dome demonstrating subchondral cystic change, overlying cartilage irregularity, and surrounding reactive bone marrow edema. (Confidence: High)
3. Anterior Bony Impingement: Prominent "kissing" osteophytes at the anterior tibial plafond and dorsal talar neck with adjacent reactive synovitis/capsular thickening in the anterior joint recess. (Confidence: High)
4. Sinus Tarsi Syndrome: Signal alteration and partial obliteration of the sinus tarsi fat, consistent with chronic scarring/inflammation of the interosseous ligaments. (Confidence: Medium to High)
5. Joint Effusion: Moderate tibiotalar joint effusion. (Confidence: High)

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

Gemini 3.1 Pro · brazo C (contexto + cirugía)arm C (context + surgery) · neg. 4/4
1. Recurrent Anterior Ankle Impingement Syndrome: Prominent recurrent/residual bony osteophytes at the anterior tibial plafond and dorsal talar neck with active "kissing" bone marrow edema. This is accompanied by anterior capsular scarring/synovial hypertrophy contributing to soft tissue impingement. (Confidence: High)
2. Osteochondral Lesion (OCL): Focal osteochondral lesion of the medial talar dome with a subchondral cyst, reactive bone marrow edema, and an overlying chondral defect. (Confidence: High)
3. Chronic Lateral Ligamentous Injury: Markedly thickened and scarred anterior talofibular ligament (ATFL) and calcaneofibular ligament (CFL), consistent with the patient's remote history of an ankle sprain. (Confidence: High)
4. Tenosynovitis: Marked tibialis posterior tenosynovitis and mild peroneal tenosynovitis. (Confidence: High)
5. Sinus Tarsi Syndrome: Edema within the sinus tarsi, likely a sequela of the patient's chronic lateral ligamentous instability/remote sprain. (Confidence: High)
6. Joint Effusions: Moderate tibiotalar and small-to-moderate subtalar joint effusions, with mild reactive subchondral edema at the posterior subtalar facet. (Confidence: High)
7. Deltoid Ligament: Mild chronic thickening of the deltoid ligament complex, likely representing chronic stress or low-grade sprain. (Confidence: Medium)

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

GPT-5.5 · brazo A (ciego)arm A (blind) · neg. 4/4
1. Talocalcaneal coalition at the middle subtalar facet/sustentacular region, likely osseous or fibrocartilaginous, with associated dorsal talar beaking, chronic hindfoot remodeling, and pes planus. High confidence.
2. Sinus tarsi edema/fluid with adjacent focal talar head/neck marrow edema, favored reactive/stress-related in the setting of subtalar coalition and altered hindfoot mechanics; no displaced fracture or collapse. High confidence for edema; medium confidence for etiology.
3. Small-to-moderate tibiotalar and posterior subtalar joint effusions/synovitis. High confidence.
4. FHL tenosynovitis, with mild additional posteromedial flexor tendon sheath fluid; no tendon rupture. High confidence.
5. Small central-to-medial/medial talar dome osteochondral lesion/subchondral abnormality without displaced fragment or collapse. Medium confidence.
6. Mild lateral ankle peri-ligamentous edema with indistinct ATFL region, suspicious for low-grade lateral ligament sprain/partial ATFL injury; no definite complete tear. Low-to-medium confidence.

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

GPT-5.5 · brazo B (contexto)arm B (context) · neg. 4/4
1. Chronic osteochondral lesion/subchondral cystic change of the medial talar dome/medial talar shoulder with surrounding marrow edema and mild articular surface irregularity. No displaced fragment or collapse. Confidence: high for talar osteochondral/subchondral abnormality; medium-high for precise articular extent.
2. Chronic lateral ankle ligament injury, greatest involving the ATFL, with nonvisualization/attenuation of the normal ATFL and anterolateral gutter scarring; mild chronic CFL/lateral collateral ligament scarring. Associated small chronic avulsion fragment/os subfibulare at the fibular tip. Confidence: medium-high.
3. Mild post-traumatic/degenerative tibiotalar arthrosis with anterior ankle impingement morphology, including anterior tibial plafond/dorsal talar neck spurring, and small-to-moderate tibiotalar effusion/synovitis. Confidence: high.
4. Sinus tarsi inflammation/scarring with mild subtalar recess fluid/synovitis, likely chronic/post-traumatic. Mild middle subtalar facet irregularity is suggested; talocalcaneal coalition is not definitively confirmed. Confidence: medium.
5. FHL tenosynovitis and mild peroneal tenosynovitis without definite tendon tear. Confidence: high for FHL; medium for peroneal tenosynovitis.

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

GPT-5.5 · brazo C (contexto + cirugía)arm C (context + surgery) · neg. 4/4
1. Chronic post-traumatic tibiotalar arthropathy/impingement morphology with residual or recurrent anterior distal tibial and dorsal talar neck osteophytes, anterior capsular scarring, and medial/lateral gutter spurring. Small-to-moderate tibiotalar effusion with mild synovitis/scarring. Confidence: High.
2. Chronic lateral ligament injury, including chronic ATFL tear/attenuation with anterolateral gutter scar tissue; suspected chronic CFL sprain/partial tear. Findings may contribute to anterolateral ankle impingement symptoms. Confidence: Medium-high for ATFL; medium-low for CFL.
3. Prominent sinus tarsi fluid/edema with small lobulated ganglion/synovitis-type component and small subtalar effusion, compatible with sinus tarsi syndrome in the appropriate clinical setting. Suspected chronic degeneration/sprain of the sinus tarsi ligament complex. Confidence: High for sinus tarsi fluid/synovitis; Medium for ligament degeneration.
4. Suspected chronic medial-to-central talar dome osteochondral/chondral-subchondral lesion, without displaced fragment or convincing acute marrow edema. Confidence: Medium.
5. FHL tenosynovitis and/or communicating ganglion-like fluid, and mild peroneal tenosynovitis. No definite tendon tear. Confidence: Medium-high for FHL; Medium for peroneals.
6. Prominent posterior talar process/os trigonum region morphology with posterior recess low-signal tissue/fluid, which can be seen with posterior ankle impingement if clinically concordant. Confidence: Medium.
7. No acute osseous abnormality. Confidence: High.

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

Claude Fable 5.1 · brazo A (ciego)arm A (blind) · neg. 3/4
1. Small-to-moderate tibiotalar joint effusion with synovitis and fluid in the medial/lateral gutters; posterior subtalar joint fluid with sinus tarsi oedema/synovitis and poorly defined interosseous ligament (probable sinus tarsi syndrome pattern / interosseous ligament sprain). Confidence: high (effusion), medium (sinus tarsi).
2. Multicompartment tenosynovitis: peroneal tenosynovitis extending toward the cuboid tunnel; flexor hallucis longus tenosynovitis extending to the knot of Henry; mild tibialis posterior/FDL and trace extensor sheath fluid. No tendon tear or subluxation. Combined with (1), the pattern is of diffuse periarticular synovitis/tenosynovitis – post-traumatic (inversion injury) or inflammatory arthropathy should be considered clinically. Confidence: medium-high.
3. Probable lateral collateral ligament (ATFL/CFL) sprain with periligamentous oedema and anterolateral gutter fluid; possible mild deltoid sprain, deep fibres intact. Ligament fibres not optimally profiled; grade cannot be confidently assigned. Confidence: low-medium.
4. Mild subchondral marrow signal change in the talus – lateral talar dome shoulder (suspicious on coronal T1, faint on PD SPAIR) and inferior talar body adjacent to the posterior subtalar facet (STIR). A small osteochondral lesion or subchondral cyst is not excluded; no displaced fragment or defect. Confidence: low-medium. Consider follow-up or dedicated thin-section imaging if lateral ankle symptoms persist.
5. Crescentic plantar/posterolateral calcaneal high signal on coronal PD SPAIR only, not reproduced on STIR, axial fat-suppressed or T1 images and located at the margin of fat suppression – favoured to be technical artefact; a calcaneal contusion/stress reaction is considered unlikely. Confidence: low for true pathology. No fracture.
6. Minor/equivocal findings: prominent posterior talar process, possibly an os trigonum, without oedema; equivocal mild proximal plantar fascial thickening/peri-fascial oedema; plantar-lateral midfoot and periarticular subcutaneous soft-tissue oedema; tiny cyst-like focus at the calcaneocuboid joint (likely partial volume).
7. Normal: Achilles tendon, Kager fat pad and retrocalcaneal bursa; distal tibia and fibula (no fracture; distal tibial T1 signal represents physeal scar/haematopoietic marrow); syndesmosis; midfoot and forefoot bones and joints.

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

Claude Fable 5.1 · brazo B (contexto)arm B (context) · neg. 4/4
1. Chronic sequelae of prior lateral ankle sprain: thickened, scarred anterior talofibular ligament (possible associated chronic calcaneofibular ligament injury) with soft-tissue thickening in the anterolateral gutter – findings that can produce anterolateral soft-tissue impingement. No acute ligament tear. (Confidence: medium)
2. Small-to-moderate tibiotalar joint effusion with synovial thickening; posterior subtalar joint fluid. (Confidence: high)
3. Sinus tarsi edema/synovitis with indistinct interosseous/cervical ligaments and preserved fat on T1 – pattern of post-traumatic sinus tarsi syndrome. (Confidence: medium)
4. Probable small chronic osteochondral lesion of the medial talar dome (subchondral low-T1 signal with slight contour irregularity), without significant surrounding marrow edema, subchondral collapse or displaced fragment – favors a stable/quiescent lesion. (Confidence: medium; the lesion is inconspicuous on the fat-suppressed sequences. CT could further characterize if clinically warranted.)
5. Mild peroneal tenosynovitis and mild fluid in the medial flexor (FHL/FDL) tendon sheaths; tendons intact. (Confidence: low–medium; medial sheath fluid may be secondary to the joint effusion.)
6. Minor additional findings: small anterior tibial lip osteophyte (possible early anterior bony impingement), probable Stieda process/os trigonum without reactive change, and nonspecific mild signal at the medial malleolar deltoid attachment. (Confidence: low)
7. No fracture, no acute marrow edema, no tendon tear; Achilles tendon, plantar fascia and syndesmosis intact.

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

Claude Fable 5.1 · brazo C (contexto + cirugía)arm C (context + surgery) · neg. 3/4
1. Moderate tibiotalar joint effusion with anterior recess synovitis/capsular scarring and minor residual/recurrent anterior tibial lip and dorsal talar neck osseous prominence – combined soft-tissue and mild osseous anterior impingement pattern in this post-arthroscopic ankle; the most likely correlate for the patient's chronic anterior ankle pain. (Effusion: high confidence; synovitis/scar: medium; residual osseous impingement: medium.)
2. Anterolateral gutter soft-tissue thickening with chronically thickened/indistinct ATFL – post-sprain/post-arthroscopic capsuloligamentous scarring with anterolateral impingement configuration. No acute ligament tear. (Medium confidence.)
3. Sinus tarsi fluid/scarring with partial fat replacement and mild posterior subtalar effusion – sinus tarsi syndrome pattern; mild lateral talar process subchondral irregularity suggests early subtalar degenerative change. (Medium confidence.)
4. Peroneal tenosynovitis without definite tendon tear or subluxation; a subtle peroneus brevis split is questioned on one sequence but not confirmed. (Tenosynovitis: medium-high; split tear: low.)
5. Fluid in the FHL sheath extending distally toward the knot of Henry, most likely communicating with the joint effusion with possible mild FHL tenosynovitis. (Medium confidence.)
6. Medial gutter synovitis with chronic deltoid (superficial and deep) scarring; no acute deltoid tear. (Medium-low confidence.)
7. Questionable tiny (few-mm) subchondral focus in the central/medial talar dome, seen on sagittal STIR only and not confirmed on other sequences; no definite osteochondral lesion. If clinically relevant, could be re-evaluated with dedicated high-resolution cartilage-sensitive sequences or follow-up. (Low-medium confidence.)
8. Mild dorsal talonavicular/naviculocuneiform capsulitis with marginal osteophytes and minimal marginal marrow signal – early degenerative/impingement change. (Low-medium confidence.)
9. No fracture, no significant bone marrow oedema, no tendon rupture, no loose body. Syndesmosis, Achilles tendon (at most borderline distal tendinosis), plantar fascia and spring ligament are intact. (High confidence.)

sintesis.md · JSON: STIR sag · T1 sag · T1 · DP ax · T1 aTSE · DP cor

AntecedenteBackground

La fase previa: dos modelos médicos open-source que no vieron el casoThe earlier phase: two open-source medical models that couldn't see the case

Antes de los frontier, el mismo protocolo se ejecutó con dos modelos médicos open-source pequeños, autoalojados en una RTX 4090 alquilada (vast.ai, coste total $1.65) — la vía que no exige enviar imágenes a nadie. El resultado fue categórico: MedGemma 1.5 4B (Google) puntuó 0.10 en ambos brazos, con 3 falsos positivos en ciego y una síntesis que degeneró en un bucle repitiendo 17 veces un bloque que contradecía sus propios hallazgos; Lingshu-7B (Alibaba DAMO) puntuó 0.00 en ciego — su informe completo fue, literalmente, «no abnormality» — y 0.20 con contexto, con 3 falsos positivos inducidos por el propio contexto clínico. Esa fase motivó la pregunta que responde esta página: ¿fallan porque son IA, o porque son pequeños? Sus cuatro informes crudos siguen disponibles: MedGemma A · B, Lingshu A · B.

Before the frontier models, the same protocol ran on two small open-source medical models, self-hosted on a rented RTX 4090 (vast.ai, total cost $1.65) — the route that requires sending images to no one. The result was categorical: MedGemma 1.5 4B (Google) scored 0.10 in both arms, with 3 blind false positives and a synthesis that degenerated into a loop repeating 17 times a block contradicting its own findings; Lingshu-7B (Alibaba DAMO) scored 0.00 blind — its entire report was, literally, "no abnormality" — and 0.20 with context, with 3 false positives induced by the clinical context itself. That phase motivated the question this page answers: do they fail because they are AI, or because they are small? Its four raw reports remain available: MedGemma A · B, Lingshu A · B.

Lectura globalOverall reading

Conclusiones y limitacionesConclusions and limitations

  • Los open-source pequeños no ven el caso. Mejor sensibilidad, 0.20. La «segunda opinión por IA» autoalojable de 2026, en RM musculoesquelética real, no sustituye nada.
  • Los frontier: salto de clase. De 0.20 a una mediana de 0.75; dos lecturas con contexto alcanzan 0.90 y reproducen cuatro de los cinco hallazgos del informe real, fallando solo la bilateralidad del edema óseo. Y el ciego ya no es un muro: GPT-5.5 alcanza 0.80 sin ningún dato clínico.
  • La sobrellamada persiste a cualquier escala: 4–7 falsos positivos por lectura frontier, patrón «informe completista» (derrame, tenosinovitis, seno del tarso), y ninguna de las seis lecturas respeta los negativos explícitos del informe.
  • El contexto clínico sigue siendo un factor enorme — pero en frontier ya no es imprescindible para ver la articulación, y (salvo Gemini) no dispara la alucinación.
  • Fase 2 invierte la pregunta: sobre la RM de 2025, las lecturas con contexto de los tres modelos concuerdan mejor con la historia documentada 2018–2021 del paciente que el informe oficial de 8 líneas — infra-lectura del informe, sobrellamada de los modelos, o (probablemente) ambas. El detalle, en la sección de fase 2.
  • Conclusión práctica: la IA de 2026 todavía no sustituye al radiólogo — un informe con cinco falsos positivos genera ruido clínico real — pero en un solo salto de clase pasó de «no ve el caso» a «ve casi todo y añade cosas de más». Este documento existe precisamente porque la opinión que importa sigue siendo humana.
  • The small open-source models can't see the case. Best sensitivity, 0.20. Self-hostable "AI second opinion" in 2026, on real musculoskeletal MRI, replaces nothing.
  • The frontier models: a class jump. From 0.20 to a median of 0.75; two context readings reach 0.90 and reproduce four of the five findings of the actual report, missing only the bilaterality of the bone marrow edema. And blindness is no longer a wall: GPT-5.5 reaches 0.80 with zero clinical data.
  • Overcalling persists at every scale: 4–7 false positives per frontier reading, a "completist report" pattern (effusion, tenosynovitis, sinus tarsi), and none of the six readings respects the report's explicit negatives.
  • Clinical context remains a huge factor — but for frontier models it is no longer required to see the joint, and (Gemini aside) it does not trigger hallucination.
  • Phase 2 inverts the question: on the 2025 MRI, the context readings of all three models agree better with the patient's documented 2018–2021 history than the official 8-line report — an under-read report, overcalling models, or (probably) both. Details in the phase 2 section.
  • Practical takeaway: 2026 AI still does not replace the radiologist — a report with five false positives generates real clinical noise — but in a single class jump it went from "can't see the case" to "sees almost everything and adds extras". This document exists precisely because the opinion that matters is still human.

Limitaciones (importantes)Limitations (important)

n = 1 paciente, 1 estudio. PNGs re-muestreados (máx. 24 cortes/serie), no el DICOM completo. Evaluador no ciego a la identidad de los modelos, y los criterios «PARCIAL» implican juicio. Las imágenes anonimizadas se enviaron a APIs comerciales vía OpenRouter — a diferencia de la fase previa autoalojada, y como desviación deliberada de la regla inicial del experimento. Además, esta fase se orquestó con asistencia de Claude (Anthropic), cuyo modelo Fable 5.1 participa como evaluado; los criterios de puntuación estaban fijados de antemano, pero el conflicto se declara. Esto es un experimento documental sobre un caso propio, no una validación clínica, y no sustituye a un radiólogo.

n = 1 patient, 1 study. Re-sampled PNGs (max 24 slices/series), not the full DICOM. The evaluator was not blinded to model identity, and the "PARTIAL" criteria involve judgment. The anonymized images were sent to commercial APIs via OpenRouter — unlike the earlier self-hosted phase, and as a deliberate deviation from the experiment's initial rule. This phase was also orchestrated with the assistance of Claude (Anthropic), whose Fable 5.1 model is among those evaluated; the scoring criteria were fixed in advance, but the conflict is disclosed. This is a documentary experiment on my own case, not a clinical validation, and it does not replace a radiologist.

Fase 2 (hecha): la RM de 2025, con un tercer brazo que añade el antecedente quirúrgico — y un informe oficial de 8 líneas que contradice el propio expediente. Ver la sección completa ↑Phase 2 (done): the 2025 MRI, with a third arm adding the surgical history — and an official 8-line report that contradicts the patient's own record. See the full section ↑