MedImageOSWorld: Benchmarking GUI Agents for Medical Image Consoles

Anonymous release accompanying a paper under double-blind review.

Code: anonymous review repository

MedImageOSWorld is a benchmark for evaluating whether general-purpose GUI agents can operate simulated CT, MR, and ultrasound acquisition consoles — and whether their actions actually acquire the intended anatomy.

Agents see only screenshots and act through mouse and keyboard. They configure protocols, plan acquisitions, inspect the resulting images, and make corrective adjustments. A hidden oracle then scores the final acquisition against anatomical ground truth in the digital subject's coordinate frame.

Why it is different: action-conditioned acquisition

In most computer-use and clinical-agent benchmarks, the data the agent looks at is fixed. In MedImageOSWorld, agent actions change the physical acquisition state, which changes the image evidence produced at later steps:

Console	What the agent plans	What changes in the next image
CT	scan range, isocentre, monitoring ROI, timing on the scout	coverage of the acquired stack
MR	plane position, orientation, FOV, slab/slices on localizers	resliced anatomy at the committed geometry
Ultrasound	probe position, orientation, contact, depth/gain/focus	live image rendered from the probe pose

Display-only actions (e.g. window/level) change the interface but not the acquisition. An agent can complete the workflow and still acquire the wrong anatomy; the benchmark measures both separately.

Environment

• Observation: a PNG screenshot of the console at a fixed viewport. No DOM, accessibility tree, simulator state, or scanner-control API.
• Action: constrained mouse events and an allowlisted set of keyboard events, plus explicit wait. Simulated time advances only through executed actions.
• Subjects: XCAT-based digital human anatomy phantoms with registered anatomical structures in a shared subject coordinate frame. No human or animal scans are acquired or used.
• Action receipts: every step logs the raw model output, any normalization/repair, whether it was executed or rejected, and feedback returned to the agent.

Tasks and capability ladder

The benchmark contains 49 task definitions — 8 CT, 34 MR, 7 ultrasound — designed by three medical imaging experts (assisted by a frontier AI model), cross-reviewed, and finalized against a shared capability ladder:

Level	Capability
L0	Operate / read a value
L1	Follow a workflow
L2	Judge view / adequacy
L3	Plan geometry (mm, °)
L4	Set parameters
L5	Control with feedback
L6	Detect & recover

Task families cover thoracic, renal, hepatic, and cardiac acquisition: CT lung coverage, contrast timing, dose-index constraints, and undercoverage recovery; MR cardiac cine planes, T1/T2 mapping, and bilateral renal coverage; ultrasound kidney search, optimization, and fault recovery. Each task condition targets 30 digital subjects × 3 seeds.

Evaluation

Each episode receives PASS, PARTIAL, FAIL, or an undetermined outcome; environment or scorer faults are INVALID and excluded from denominators. Three 0–100 metrics are reported side by side and never merged:

• SR — Success rate (primary): share of slots meeting the task's full success criterion.
• PR — Progress rate: mean fraction of workflow stages completed.
• AQ — Acquisition quality: continuous, normalized quality of the acquired result from hidden ground truth (geometry in mm/degrees, anatomical coverage fractions).

The overall score weights capability levels by w_i = i + 1:

$$\text{Score}=100\frac{\sum _{i=0}^6(i+1)\sum _{j=1}^{n_i}{s}_{ij}}{\sum _{i=0}^6(i+1)n_i}\backslash text\{Score\}\; =\; 100\backslash ,\backslash frac\{\backslash sum\_\{i=0\}^\{6\}(i+1)\backslash sum\_\{j=1\}^\{n\_i\}\; s\_\{ij\}\}\{\backslash sum\_\{i=0\}^\{6\}(i+1)\backslash ,n\_i\}$$

where s_ij is the SR proportion for task j at level i (weighted denominator 181).

Results

11 open-weight models (greedy decoding) and 2 API-served proprietary models, over 31,197 valid scored episodes (30,513 open-weight, 684 proprietary; proprietary agents use the first five assets per task).

Agent	Access	Overall SR
GPT-6 Sol	API	34.5
Claude Opus 5.5	API	34.1
Qwen3.8-27B	open-weight	25.0
Muse-Glimmer-30B	open-weight	20.8
Gemma-4-31B	open-weight	14.6
UI-Mate-27B	open-weight	10.7
Qwen3-VL-30B-A3B-Thinking	open-weight	7.7
GUI-Owl-1.5-32B-Thinking	open-weight	6.5
GUI-Owl-1.5-32B-Instruct	open-weight	5.5
Qwen2.5-VL-72B-Instruct	open-weight	5.2
Qwen3-VL-30B-A3B-Instruct	open-weight	3.7
OpenCUA-32B	open-weight	3.3
UI-TARS-1.5-7B	open-weight	3.0

The API-served agents exceed Qwen3.8-27B in workflow progress by only 9–11 points (PR 85.5 and 83.4 vs. 74.5), but lead in acquisition quality by about 19 points (AQ 61.1 and 61.3 vs. 42.0).

Key findings:

• Agents complete workflows but rarely acquire the intended anatomy. Open-weight SR ranges 3.0–25.0 while PR reaches 21.5–74.5.
• Success collapses between perception and planning. The best open-weight agent reaches 79–90% at L0–L2, at most 12% at L3, and every open-weight agent scores 0% at L5.
• CT is nearly unsolved for open-weight agents (SR ≤ 1.5); proprietary agents reach 10.3 and 1.7.
• Agents observe but do not correct. Review gates are almost always passed, but agents change the responsible acquisition parameter in only 14% of renal MR coverage-correction episodes, and in CT recovery tasks no agent corrected a clipped range.
• Budget exhaustion dominates failure: 64.6% of scored open-weight episodes end with an exhausted action budget.

What is in this repository

tasks/<ct|mr|us>/<task_id>/
    task.json      agent-visible task setup: goal text, level, budgets, observation and action profile, initial state
    start.png      the step-0 console screenshot the agent receives (where available)
    success.png    final console state of a reference episode that passed the scorer (where available)
    success.gif    the same reference episode as a short animation (where available)
try-it/
    ct/torso-task-pack/      compiled CT task pack for the development phantom (7 CT tasks)
    ct/kub-task-pack/        compiled CT task pack for the low-dose KUB task
    mr/vfemale50-v1/         compiled MR subject bundle for the development phantom
    us/synthetic-abdomen-v1/ ultrasound scene generated from mathematical primitives only (no XCAT lineage)

All visualizations use a single development phantom (vfemale50). It is not among the 30 subjects used for scoring, so these images reveal nothing about evaluated episodes. Task cards contain only what the evaluated agent sees. Hidden oracles, ground-truth masks, reference geometry, scorer thresholds and verdicts are not included.

Success examples come from privileged reference solvers that read simulator state. They are not pixels-only model runs. Each one is shown only if the benchmark scorer graded the episode PASS.

Coverage	CT	MR	US
Task setup card	8 / 8	34 / 34	7 / 7
Step-0 screenshot	8 / 8	22 / 34	pending
Success example	8 / 8	15 / 34	pending

Visualizations are pending for the remaining tasks:

• some MR geometry and mapping tasks need private reference geometry that has not been built for the development phantom;
• the MR renal tasks need a renal bundle for the development phantom;
• the ultrasound kidney tasks need an XCAT kidney scene for the development phantom.

The ultrasound console can meanwhile be tried on the math-only synthetic scene.

Example

CT lung coverage, step 0	CT lung coverage, reference success

Data access and the XCAT license

The digital subjects are derived from the 4D extended cardiac-torso (XCAT) phantom (Segars et al., 2010), which is licensed by Duke University and is not openly redistributable. Because of these license terms, this repository shares only one compiled XCAT-derived model: the development phantom vfemale50, as answer-free runtime bundles under try-it/. These bundles are provided solely so that the consoles can be tried. They are not covered by the repository license and may not be redistributed. No raw phantom files (.nrb, .par, voxelized activity volumes), phantom software, or scoring cohort is included.

To evaluate on the full cohort, you can:

• Request access. The compiled cohort bundles used in the paper are available on request for non-commercial research, subject to the XCAT license terms. Please open a discussion on this repository.
• Use your own data. Anyone holding an XCAT license can generate equivalent subjects with their own phantom installation. Other digital phantoms or anatomical label volumes can be used too, once converted into the bundle format.

The console source code and runtime instructions are in the anonymous review repository.

Tasks

Modality	Level	Task	Setup	Step 0	Success example
CT	L3	ccta_heart_coverage_isocenter	card	start	png gif
CT	L3	ct_liver_coverage_isocenter	card	start	png gif
CT	L3	ct_lung_coverage_isocenter	card	start	png gif
CT	L3	ct_renal_bilateral_coverage	card	start	png gif
CT	L3	ct_ui_smoke_routine	card	start	png gif
CT	L4	ct_kub_low_dose_constraint	card	start	png gif
CT	L6	ct_renal_undercoverage_recovery	card	start	png gif
CT	L6	ct_undercoverage_recovery	card	start	png gif
MR	L0	L0_physio_trigger	card	start	png gif
MR	L0	mr_drag_coronal_to_viewport2_v1	card	start	png gif
MR	L0	mr_rename_sequence_4ch_v1	card	start	png gif
MR	L0	mr_select_transverse_viewport_v1	card	start	png gif
MR	L0	mr_show_shim_box_v1	card	start	png gif
MR	L1	L2a_first_acquisition	card	start	png gif
MR	L1	L2a_graphics_display_mode	card	start	png gif
MR	L1	L2a_queue_mutation	card	start	png gif
MR	L1	L2a_skip_row	card	start	png gif
MR	L1	mr_repeat_and_open_v1	card	start	png gif
MR	L2	L1_go_availability	card	start	png gif
MR	L2	L1_idle_or_scanning	card	start	png gif
MR	L2	mr_renal_acquisition_adequacy_v2	card	—	pending
MR	L2	mr_select_sequence_v1	card	start	png gif
MR	L2	mr_select_transverse_image_only_v1	card	start	png gif
MR	L3	L3_coarse_2ch	card	start	pending
MR	L3	L3_coarse_4ch	card	—	pending
MR	L3	L4_2ch	card	start	pending
MR	L3	L4_4ch	card	—	pending
MR	L3	L4_fine_2ch	card	—	pending
MR	L3	L4_fine_4ch	card	—	pending
MR	L3	L4_long_axis_pair	card	start	pending
MR	L3	L5_fine_lax_from_scratch	card	start	pending
MR	L3	L5_sax_stack	card	—	pending
MR	L3	mr_renal_bilateral_geometry_planning_v2	card	—	pending
MR	L3	mr_t1t2_mapping_geometry_staged	card	—	pending
MR	L4	L4_heart_shim_box	card	start	pending
MR	L4	mr_renal_numeric_coverage_correction_v2	card	—	pending
MR	L4	mr_t1t2_mapping_parameters	card	—	pending
MR	L5	mr_t1t2_mapping_from_boot	card	start	pending
MR	L5	mr_t1t2_mapping_from_empty	card	start	pending
MR	L5	mr_t1t2_mapping_staged	card	—	pending
MR	L6	L6_stopped_row_retry	card	start	png gif
MR	L6	mr_renal_bilateral_coverage_recovery_v2	card	—	pending
US	L0	us_kidney_set_acquisition_v1	card	—	pending
US	L0	us_kidney_set_display_v1	card	—	pending
US	L1	us_kidney_measure_annotate_save_v1	card	—	pending
US	L1	us_kidney_preset_acquire_save_v1	card	—	pending
US	L3	us_kidney_long_axis_search_v1	card	—	pending
US	L4	us_kidney_acquisition_optimization_v1	card	—	pending
US	L6	us_kidney_contact_vs_gain_v1	card	—	pending

Scope and limitations

• Simulated consoles and task-specific physical abstractions — not a scanner, not a medical device, and not evidence of scanner equivalence or clinical validation.
• Geometry and coverage measurements characterize acquired anatomy under these models; clinical image quality and diagnostic utility require separate evaluation.
• The capability ladder organizes task requirements; it is not a psychometrically validated difficulty scale.
• Results depend on the subject cohort, task versions, and agent interface.
• XCAT phantoms are licensed separately and are not redistributed under this repository's license.

Citation

@misc{medimageosworld2026,
  title  = {MedImageOSWorld: Benchmarking GUI Agents for Medical Image Consoles},
  author = {Anonymous},
  year   = {2026},
  note   = {Under review}
}