Cerebrovascular treatment research review
Multimodal treatment recommendation with auditable clinical reasoning.
The complete 83-case cohort combines preoperative clinical text with AP and lateral DSA views. Every case has an out-of-fold Top-1 and Top-2 recommendation, a model-generated rationale, and an explicit uncertainty statement.
Review methodology
Case evidence and review methodology
The technical pipeline covers all 83 cases. Blinded physician review remains the next validation step.
Data audit
83 de-identified cases were linked to 166 PNG images. Each case has one AP and one lateral view.
Leakage-controlled evaluation
Each case appeared in a held-out test fold exactly once across five folds.
Multi-reference scoring
Top-1 and Top-2 predictions were compared with all supplied expert references, preserving disagreement.
Reasoning generation
Qwen3.6-27B generated 83 case rationales from model outputs without receiving the reference labels.
Interpretation
Strong candidate coverage, limited exact agreement
Top-2 coverage is high when any expert reference is accepted. Exact agreement with the technique actually used remains modest, reflecting both model error and real treatment-choice variability.
Important limits
Not yet a clinical validation result
- No patient-to-case mapping was supplied; evaluation is case-level.
- No independent locked external cohort has been tested.
- Class scores have not been calibrated as probabilities of clinical correctness.
- Generated rationales still require blinded physician review.
Cohort organization
One primary grouping, three complementary views
Cases are grouped by the model's Top-1 technique and can be filtered by performance, score separation, and expert-reference agreement.
Complete cohort
All 83 de-identified cases
Use the filters to inspect recommendations, failure modes, uncertainty, and expert disagreement without changing the underlying case set.
Blinded qualitative physician review
16 selected de-identified cases
Selected to cover confident, ambiguous, adjunctive-device, and retreatment scenarios. Outcome groups and expert-reference agreement are intentionally hidden in this section.
Review case 1 of 16
Case 61
Assigned out-of-fold recommendation
Coil embolization alone
Alternative: Flow diversion with coil embolization


View A is AP; View B is lateral, as confirmed by the data provider.
Case summary
A 75-year-old male presents with an unruptured right ACA A1 aneurysm measuring 5.3 mm x 4 mm x 5.4 mm with a 3.1 mm neck and a daughter sac.
Image observation
On the supplied views, a saccular outpouching appears on the proximal anterior cerebral artery with a visible neck connecting to the parent vessel.
Model-generated clinical rationale
The documented dimensions and neck width suggest a morphology that may be amenable to endovascular coiling. However, the presence of a daughter sac and the specific neck-to-dome ratio are critical factors that are not fully quantified in the narrative, potentially complicating coil stability.
Model uncertainty
The OOF classifier showed a relatively clear preference for Coil embolization alone over Flow diversion with coil embolization, but its scores are not probabilities of clinical correctness.
Clinical uncertainty / missing factor
The precise neck-to-dome ratio and the exact relationship of the daughter sac to the aneurysm neck are missing, which are essential for determining the feasibility of coil-only treatment versus adjunctive techniques.
Explicit evidence used
Unruptured right ACA A1 aneurysm | Dimensions 5.3 mm x 4 mm x 5.4 mm with 3.1 mm neck | Presence of a daughter sac