Back to News
InnovationAI Understanding briefing

Longer fMRI scans can cut medical AI research costs by up to 22%, study finds

A 2025 Nature study modeled that 30‑minute fMRI scans save an average of 22% versus 10‑minute scans when aiming for comparable prediction performance, challenging the assumption that more participants always lower total cost.

5 min readRead the linked source
Source-provided image accompanying Longer fMRI scans can cut medical AI research costs by up to 22%, study finds
Source referenceSource recorded
Publisher
koreatechdesk.com
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

What happened

A 2025 study published in Nature examined how participant numbers and fMRI scan duration affect the cost‑effectiveness of brain‑wide association studies used for medical AI. Modeling 108 cost‑performance scenarios across 76 phenotypes, the researchers found that in 85% of cases the cheapest design used scans of at least 20 minutes, and that 30‑minute scans yielded an average modeled cost saving of 22% compared with 10‑minute scans for comparable predictive accuracy. The analysis incorporated hypothetical scanner fees of $500–$1,000 per hour and participant‑related overhead of $500–$1,000. The authors also released an “Optimal Scan Time Calculator” to let investigators explore trade‑offs between sample size, scan length, and budget. The article contextualizes these findings for Korean medical‑AI startups, noting Korea’s 2026 medical‑AI data‑voucher program (KRW 800 million for ten SMEs) and emphasizing that data‑collection decisions should be tied to a clearly defined development milestone.

The Nature paper, titled “Longer scans boost prediction and cut costs in brain‑wide association studies,” was co‑authored by Dr. Leon Ooi (CEO of Singapore‑based B1Neuro and NUS research fellow) and colleagues. It analyzed nine resting‑state and task‑based fMRI datasets, covering 76 phenotypic outcomes ranging from cognitive scores to clinical measures. Using a mathematical model that explained 89% of observed prediction variance (R² = 0.89), the authors simulated how varying participant numbers and scan lengths would affect both predictive accuracy and total study cost.

Cost assumptions included two scanner‑hour rates ($500 and $1,000) and two participant‑overhead rates ($500 and $1,000). Across 108 simulated scenarios, the cheapest configuration in 85% of cases involved scans of at least 20 minutes per participant. When comparing a standard 10‑minute protocol to a 30‑minute protocol under identical performance targets, the longer scans saved an average of 22% in total cost, primarily because the additional scan time extracted more predictive signal per participant, reducing the need for extra recruitment.

The authors caution that the 22% figure is model‑based and contingent on the specific cost inputs and prediction goals used. Different clinical questions, scanner pricing structures, or recruitment environments could shift the optimal balance. Nonetheless, the analysis provides a quantitative framework for evaluating the trade‑off between measurement depth and cohort size.

To help other researchers apply the framework, the team released an online “Optimal Scan Time Calculator.” Users can input their own cost parameters, desired prediction performance, and target phenotype to explore cost‑optimal combinations of sample size and scan duration. The tool is tailored to fMRI‑based brain‑wide association studies and may not directly translate to other imaging modalities or AI tasks.

The article also references Korea’s 2026 medical‑AI data‑voucher program, which allocates KRW 6.4 billion in support for SMEs. While the vouchers fund data‑utilization broadly rather than specific scan time, the program underscores the relevance of cost‑efficient study design for startups seeking public funding.

Source details: koreatechdesk.com

Why it matters

The study overturns a common budgeting rule that recruiting more participants is always the most economical way to improve model performance. By showing that longer scans can reduce total cost under realistic overhead assumptions, it gives medical‑AI founders a concrete tool to allocate scarce capital more efficiently. This matters because fMRI studies are among the most expensive data sources for AI‑driven diagnostics and therapeutics; mis‑allocating funds to recruit additional subjects can quickly drain startup runway without delivering proportional predictive gains. The findings also highlight that the optimal design depends on the rarity of the target phenotype—rare conditions still demand larger cohorts, so longer scans cannot replace recruitment in those cases. For investors and policymakers, the work provides a quantitative basis to evaluate grant proposals and voucher applications, ensuring that public funds support study designs that maximize scientific return per dollar. The publicly available calculator further democratizes this insight, allowing research teams worldwide to model their own cost‑performance trade‑offs before committing resources.

Cost efficiency directly impacts the viability of early‑stage medical‑AI companies, many of which operate on limited venture capital or grant funding. By demonstrating that longer scans can lower overall expenditure, the study offers a concrete lever to stretch budgets without sacrificing model quality.

The insight reshapes how investors evaluate research proposals. Rather than assuming that a larger participant pool is inherently better, due diligence can now assess whether a proposed design aligns with the optimal scan‑time trade‑off, potentially reducing funding risk.

Policymakers designing grant or voucher programs can incorporate these findings to set guidelines that encourage cost‑effective data collection, ensuring public money supports studies with the highest expected scientific return.

For the broader AI research community, the work provides a template for integrating economic modeling into experimental design, a practice that has been under‑explored in many AI‑driven biomedical fields.

Finally, the publicly available calculator democratizes access to this analysis, allowing smaller labs and startups without extensive economics expertise to make data‑driven budgeting decisions.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Model Parameter Size:8B Parameters
VRAM Required5.5 GBGPU memory footprint
Target HardwareMacBook / Single GPUDeployment tier
Privacy100% Air-GappedLocal device capability
Core takeaway: Small, quantized models (3B–8B) now run directly inside smartphones and laptops with complete data privacy, while mammoth 400B+ models remain the domain of datacenter clusters.
Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

What to watch next

Future medical‑AI projects will likely cite the optimal‑scan model when justifying budgets, and we may see more startups adopting longer fMRI protocols or using the calculator to tailor study designs. Watch for follow‑up studies that validate the modeled savings with real‑world cost data, and for policy updates in Korea or elsewhere that incorporate these cost‑efficiency insights into funding criteria for medical‑AI research.

Adoption of the Optimal Scan Time Calculator by academic labs and industry R&D teams, potentially leading to published follow‑up studies that compare modeled savings with actual expenditures.

Changes in grant application guidelines from Korean agencies or other national bodies that reference scan‑time optimization as a criterion for funding eligibility.

Emergence of new startup business models that prioritize deeper imaging per participant over large‑scale recruitment, especially in markets where scanner time is relatively inexpensive.

Potential revisions to the cost assumptions (e.g., scanner pricing, recruitment overhead) as technology evolves, which could shift the optimal balance and updated modeling.

Monitoring of any regulatory or ethical discussions about longer scan durations, such as participant burden or safety considerations, which could affect the feasibility of implementing longer protocols.

Related guides & quizzes

Found this useful?