What participants receive¶

For all tasks, participants receive one folder per case containing three files:

The clinical-data file name is task-specific: prostate-biopsy-decision-clinical-data.json (Task 1), prostate-treatment-decision-clinical-data.json (Task 2), and prostate-time-to-recurrence-or-last-follow-up-clinical-data.json (Task 3).

Participants do not receive raw imaging (DICOM or WSI) at inference time; the agent reasons over the structured inputs, retrieved reports, and neural representations provided in these files. A full list of available variables and tool outputs is described on the Tools page.

Dataset¶

Training release: labels and reasoning¶

The three input files are provided for every training case. Ground-truth labels and reference reasoning are released only for a subset of training cases, and this subset is updated incrementally:

Label and reasoning files per task:

• Task 1: prostate-biopsy-decision.json ("yes" / "no") + prostate-biopsy-decision-reasoning.json
• Task 2: prostate-treatment-decision.json (treatment category) + prostate-treatment-decision-reasoning.json
• Task 3: prostate-time-to-recurrence-or-last-follow-up.json (months_to_recurrence + event); no reference reasoning provided

Per-task variables¶

The main structured variables in structured-prompt.json are listed below. Additional EHR-style context (vitals, medications, comorbidities, social history, IPSS, encounter metadata, and free-text note sections) is included where available.

Task 1 — MRI-only diagnostic decision¶

Ground truth¶

Task 1 ground truth is provided in two files:

• prostate-biopsy-decision.json: contains the biopsy decision as "yes" / "no".
• prostate-biopsy-decision-reasoning.json: contains the structured reference reasoning.

The ten variables included under variable_weights are age, fh (family history of prostate cancer), cspca, pirads, vol, psa, comorbidity (the recorded conditions listed under pmhx in structured-prompt.json), psad, dre, and bx.

Decision file: prostate-biopsy-decision.json

"yes"

Reasoning file: prostate-biopsy-decision-reasoning.json

{
  "confidence": "clear",
  "variable_weights": {
    "bx": "noted",
    "comorbidity": "noted",
    "cspca": "important",
    "psa": "important",
    "age": "important",
    "dre": "noted",
    "pirads": "decisive",
    "fh": "noted",
    "vol": "important",
    "psad": "important"
  },
  "reveal_sequence": [
    "family_history",
    "previous_notes",
    "laboratory_results",
    "psa_trend",
    "radiology_report"
  ],
  "free_text": "The PI-RADS 5 lesion, together with the markedly elevated and rapidly rising PSA, very high PSA density, and high csPCa probability, strongly supports proceeding to biopsy."
}

Task 2 — MRI + biopsy risk stratification¶

Task 2 shares the Task 1 variables, with biopsy pathology populated and the clinical-data file additionally containing a pathology_report.

Ground truth¶

Task 2 ground truth is provided in two files:

• prostate-treatment-decision.json: contains one treatment category (string): active_surveillance / continued_surveillance / watchful_waiting / active_treatment. The decision is evaluated as a 4-class target using weighted F1.
• prostate-treatment-decision-reasoning.json: contains the structured reference reasoning.

The reasoning file follows the same structure and uses the same possible values described for Task 1. The eleven variables included under variable_weights are bx_gl_prim, pirads, bx_isup, ct, fh, comorbidity, psa, bx_gl_sec, age, psad, and cspca.

Decision file: prostate-treatment-decision.json

"active_surveillance"

Reasoning file: prostate-treatment-decision-reasoning.json

{
  "confidence": "clear",
  "variable_weights": {
    "bx_gl_prim": "important",
    "pirads": "decisive",
    "bx_isup": "decisive",
    "ct": "important",
    "fh": "noted",
    "comorbidity": "noted",
    "psa": "important",
    "bx_gl_sec": "important",
    "age": "not_used",
    "psad": "important",
    "cspca": "not_used"
  },
  "reveal_sequence": [
    "family_history",
    "previous_notes",
    "laboratory_results",
    "psa_trend",
    "radiology_report",
    "pathology_report"
  ],
  "free_text": "Biopsy shows ISUP Grade Group 1 (Gleason 3+3) disease, with favorable clinical and imaging findings supporting active surveillance."
}

Task 3 — prostatectomy pathology prediction¶

Task 3 structured-prompt.json is minimal; most content is in the clinical-data file.

The clinical-data file provides radiology_report (volume, PSA density, PI-RADS, csPCa probability), pathology_report (biopsy Gleason/ISUP), surgical_pathology_report (prostatectomy Gleason/ISUP, pT stage, margins, extraprostatic extension, seminal vesicle invasion, LVI, lymph node status), previous_notes (e.g. 5-ARI use, Charlson Comorbidity Index), and family_history.

Ground truth and required output¶

Task 3 uses two files:

• prostate-time-to-recurrence-or-last-follow-up.json: contains months_to_recurrence (float), representing the time in months, and event (int), where 1 indicates observed BCR and 0 indicates a censored observation.
• prostate-time-to-recurrence-or-last-follow-up-reasoning.json: contains a free-text explanation supporting the predicted outcome. No reference reasoning is provided in the training release, but participants must generate this file as part of their output.

Outcome file: prostate-time-to-recurrence-or-last-follow-up.json

{
  "months_to_recurrence": 65.7,
  "event": 0
}

Neural representations¶

prostate-modality-level-neural-representations.json provides frozen foundation-model embeddings per modality:

Missing modalities are represented as an empty list []. For example, a Task 2 patient with two biopsy slide representations and no prostatectomy slides has the following structure:

{
  "MRI image": [
    [/* 1024-dimensional MRI vector */]
  ],
  "Biopsy slide": [
    [/* 960-dimensional biopsy slide vector 1 */],
    [/* 960-dimensional biopsy slide vector 2 */]
  ],
  "Prostatectomy slide": []
}

Underlying imaging and acquisition¶

The structured inputs provided to participants are derived from clinical imaging acquired as part of routine prostate cancer care. The table below describes the original imaging devices and protocols for reference.

Data origins and Domain shift¶

CHIMERA-agent builds on the CHIMERA challenge at MICCAI 2025, extending it with agent-based reasoning and sequential clinical decision-making. Data for CHIMERA-agent is collected independently.

All training and validation data originate from Radboud University Medical Center (Radboudumc), Nijmegen, The Netherlands and Canisius Wilhelmina Hospital (CWZ), Nijmegen, The Netherlands. The test set includes cases from Radboudumc, CWZ, and Karolinska Institute, Solna, Sweden. MRI examinations were interpreted by board-certified radiologists following PI-RADS v2.1. Biopsy procedures were performed by experienced urologists, and histopathology was assessed by board-certified pathologists specializing in genitourinary pathology.

Participants should expect domain shift between training and test cohorts due to differences in MRI acquisition parameters, H&E staining protocols, and scanner hardware.

External data and pretrained models¶

Participants who wish to train or fine-tune their models using external data and / or pretrained models are permitted provided that they are freely and publicly accessible under a permissive open-source license (e.g. Creative Commons, MIT, BSD, Open Database License) and described in the challenge manuscript.