Dual- vs Single-Suggestion AI Support for Radiographic Interpretation in Residents: Randomized Multireader Study
A · Strong evidence Registered, prospective, randomized three-arm multireader trial (ChiCTR2600129243) with 123 residents at three hospitals and pre-specified outcomes; strong method, moderate sample and a purposively selected case set.
Summary based on the full text·Original: about 40 minutes to read
In short
A randomized multireader study from Nanchang University hospitals and the Hong Kong University of Science and Technology (first authors Lin Wu and Zhe Xu), posted to arXiv on 7 October 2026. Junior doctors read radiographs first alone, then with either one AI suggestion (GPT-5.4) or two (GPT-5.4 plus Kimi-K2.6, or plus Gemini-3.6 Flash). A second opinion roughly doubled accuracy on cases where the shared GPT-5.4 suggestion was wrong. Overall accuracy gains were larger with two suggestions for radiology residents, but not for residents from other specialties.
Key numbers
- 40.1% / 40.4% vs 20.0%radiology residents' accuracy on cases where GPT-5.4 was wrong, with two AI suggestions vs one
- 31.3% / 31.0% vs 12.1%same comparison for non-radiology residents
- 46.3% to 61.5%mean diagnostic accuracy of all 123 residents before and after AI assistance
- 65.0%accuracy of each of the three AI models on the 60-case set
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- 154 residents assessed, 132 randomized 1:1:1, 123 analyzed (mean age 24.1, 65 women; 60 radiology, 63 non-radiology), all with under 3 years of experience.
- Each read the same 60 radiographs (30 chest, 30 abdominal). All three models were right on 39 of 60 cases (65.0%); a senior-resident baseline scored 54.3% unaided.
- Across all residents, accuracy rose from 46.3% to 61.5% with AI (15.2 percentage points; 95% CI 13.7-16.8). Gains: 13.9 pp single, 15.0 and 16.8 pp dual.
- Radiology residents gained 6.69 pp (B vs A) and 7.87 pp (C vs A) more with dual support (Holm-adjusted P = .030). No difference for non-radiology residents (P = .20).
- On the 21 cases where GPT-5.4 was wrong: radiology residents scored 40.1% and 40.4% with two suggestions vs 20.0% with one; non-radiology 31.3% and 31.0% vs 12.1% (all P < .001).
- The specialty interaction for the dual-suggestion effect was 10.44 pp (95% CI 4.36-16.52).
- Interpretation time and confidence change did not differ across support conditions.
How they got it
Prospective three-arm trial at three hospitals in China, July to September 2026, registered as ChiCTR2600129243. Cases came from 218 examinations at one centre (2023-2024) with confirmed diagnoses; 60 were selected so each model had identical overall accuracy. Model outputs were generated once and shown with rationale; residents recorded an unaided and then an AI-assisted diagnosis on a smartphone or tablet.
Limits
- The 60 cases were purposively selected to balance disease mix and AI accuracy, so they do not reflect routine case mix.
- The benefit depends on the second model disagreeing when the first is wrong; other model pairs may behave differently.
- Mechanisms (attention, how residents weighted each suggestion) were not measured.
Why it matters for companies
Most enterprise AI tools show one answer from one model, and the trial shows how strongly a single wrong answer pulls junior staff along: with one suggestion, accuracy on those cases fell to 12-20%. Showing a second independent model's answer roughly doubled accuracy on those cases, with no added reading time. For review-heavy work such as claims, contracts or screening, a disagreement signal between two models is a cheap design choice. The effect was larger for people with domain training, so the second opinion helps experts more than novices.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org