Thomas Yeo
@bttyeo.bsky.social
Interesting! Do large datasets nowadays use this recommended flip angle?
BTW, with regards to voxel size, another odd thing is that more (smaller) parcels decrease optimal scan time (a little). This is a little surprising to us as we thought smaller parcels => lower SNR => longer scan time
BTW, with regards to voxel size, another odd thing is that more (smaller) parcels decrease optimal scan time (a little). This is a little surprising to us as we thought smaller parcels => lower SNR => longer scan time
July 21, 2025 at 12:21 AM
Interesting! Do large datasets nowadays use this recommended flip angle?
BTW, with regards to voxel size, another odd thing is that more (smaller) parcels decrease optimal scan time (a little). This is a little surprising to us as we thought smaller parcels => lower SNR => longer scan time
BTW, with regards to voxel size, another odd thing is that more (smaller) parcels decrease optimal scan time (a little). This is a little surprising to us as we thought smaller parcels => lower SNR => longer scan time
& found that we need to add a *lot* of noise to the fMRI data in order to increase optimal scan time. This suggests that we need to improve fMRI SNR by a lot to shorten scan time. One reason might be that the multivariate predictive modeling we are using might be doing quite a bit of denoising.. 2/4
July 19, 2025 at 1:36 AM
& found that we need to add a *lot* of noise to the fMRI data in order to increase optimal scan time. This suggests that we need to improve fMRI SNR by a lot to shorten scan time. One reason might be that the multivariate predictive modeling we are using might be doing quite a bit of denoising.. 2/4
Thanks Dan, appreciate your thoughts! I agree that there might be other factors beyond # participants & scan duration. We did do some exploratory analyses, but were surprised that there's not an obvious impact of common scan parameters. We did some further simulations, ... 1/4
July 19, 2025 at 1:36 AM
Thanks Dan, appreciate your thoughts! I agree that there might be other factors beyond # participants & scan duration. We did do some exploratory analyses, but were surprised that there's not an obvious impact of common scan parameters. We did some further simulations, ... 1/4
Yeah! I know right? It's even crazier for some phenotypes. We think it has to do with some non-stationarity properties in fMRI-phenotypic relationship. We did a small analysis where we permute run order and that seems to remove the "odd" trajectories. But it would be interesting to study that more.
July 18, 2025 at 12:59 AM
Yeah! I know right? It's even crazier for some phenotypes. We think it has to do with some non-stationarity properties in fMRI-phenotypic relationship. We did a small analysis where we permute run order and that seems to remove the "odd" trajectories. But it would be interesting to study that more.
8/ In contrast to standard power calculations, our results suggest that jointly optimizing sample size and scan time can boost prediction accuracy while cutting costs. For more complex study design, you can check out our calculator: thomasyeolab.github.io/OptimalScanT...
July 17, 2025 at 1:36 AM
8/ In contrast to standard power calculations, our results suggest that jointly optimizing sample size and scan time can boost prediction accuracy while cutting costs. For more complex study design, you can check out our calculator: thomasyeolab.github.io/OptimalScanT...
7/ ... optimal scan time, so we recommend a scan time of at least 30 min. Compared with resting-state whole-brain prediction, the most cost-effective scan time is shorter for task-fMRI and longer for subcortical-to-whole-brain prediction.
July 17, 2025 at 1:36 AM
7/ ... optimal scan time, so we recommend a scan time of at least 30 min. Compared with resting-state whole-brain prediction, the most cost-effective scan time is shorter for task-fMRI and longer for subcortical-to-whole-brain prediction.
6/ Indeed, to achieve high prediction performance, 10-min scans are cost inefficient. In most scenarios, the optimal scan time is ≥20 min. On average, 30-min scans are the most cost effective, yielding 22% savings over 10 min scans. Overshooting is cheaper than undershooting ...
July 17, 2025 at 1:36 AM
6/ Indeed, to achieve high prediction performance, 10-min scans are cost inefficient. In most scenarios, the optimal scan time is ≥20 min. On average, 30-min scans are the most cost effective, yielding 22% savings over 10 min scans. Overshooting is cheaper than undershooting ...
3/11 Tom's model explains empirical prediction accuracies well across 76 phenotypes from 9 resting-fMRI & task-fMRI datasets (R2 = 0.89), spanning many scanners, acquisitions, racial groups, disorders & ages.
Does this mean that we should collect large datasets with short scans?
Does this mean that we should collect large datasets with short scans?
July 17, 2025 at 1:36 AM
3/11 Tom's model explains empirical prediction accuracies well across 76 phenotypes from 9 resting-fMRI & task-fMRI datasets (R2 = 0.89), spanning many scanners, acquisitions, racial groups, disorders & ages.
Does this mean that we should collect large datasets with short scans?
Does this mean that we should collect large datasets with short scans?
2/11 We show that for functional MRI (fMRI), prediction accuracy increases with log of total scan duration (# participants x scan time per participant) for scans ≤20 min.
Ultimately sample size is more important, which @nichols.bsky.social explains with a theoretical model…
Ultimately sample size is more important, which @nichols.bsky.social explains with a theoretical model…
July 17, 2025 at 1:36 AM
2/11 We show that for functional MRI (fMRI), prediction accuracy increases with log of total scan duration (# participants x scan time per participant) for scans ≤20 min.
Ultimately sample size is more important, which @nichols.bsky.social explains with a theoretical model…
Ultimately sample size is more important, which @nichols.bsky.social explains with a theoretical model…
1/11 Excited to share our @Naturestudy led by @leonooi.bsky.social @csabaorban.bsky.social @shaoshiz.bsky.social
AI performance is known to scale with logarithm of sample size (Kaplan 2020), but in many domains, sample size can be # participants or # measurements...
doi.org/10.1038/s415...
AI performance is known to scale with logarithm of sample size (Kaplan 2020), but in many domains, sample size can be # participants or # measurements...
doi.org/10.1038/s415...
July 17, 2025 at 1:36 AM
1/11 Excited to share our @Naturestudy led by @leonooi.bsky.social @csabaorban.bsky.social @shaoshiz.bsky.social
AI performance is known to scale with logarithm of sample size (Kaplan 2020), but in many domains, sample size can be # participants or # measurements...
doi.org/10.1038/s415...
AI performance is known to scale with logarithm of sample size (Kaplan 2020), but in many domains, sample size can be # participants or # measurements...
doi.org/10.1038/s415...
... linking ICC of FC change with ICC of cross-sectional FC & correlation between FC at Year 0 & Year 2. Based on the formula, if the FCs are highly similar between the two timepoints (which has been previously shown), then FC change will actually have lower reliability!
June 11, 2025 at 1:48 AM
... linking ICC of FC change with ICC of cross-sectional FC & correlation between FC at Year 0 & Year 2. Based on the formula, if the FCs are highly similar between the two timepoints (which has been previously shown), then FC change will actually have lower reliability!
Many cool results from this paper, including a multivariate twist of Simpson's paradox! Stronger salience FC at baseline predicts better cognition at baseline. Would children with larger salience FC increase also enjoy greater cognitive gains over time? Or is the reverse true??
June 11, 2025 at 1:48 AM
Many cool results from this paper, including a multivariate twist of Simpson's paradox! Stronger salience FC at baseline predicts better cognition at baseline. Would children with larger salience FC increase also enjoy greater cognitive gains over time? Or is the reverse true??
Furthermore, DELSSOME can replicate neurodevelopmental insights from our previous study doi.org/10.1073/pnas... /end
Congrats to Tianchu and Tian Fang for leading this study and also co-authors @shaoshiz.bsky.social @bart-larsen.bsky.social @ted-satterthwaite.bsky.social @avramholmes.bsky.social
Congrats to Tianchu and Tian Fang for leading this study and also co-authors @shaoshiz.bsky.social @bart-larsen.bsky.social @ted-satterthwaite.bsky.social @avramholmes.bsky.social
April 11, 2025 at 1:35 AM
Furthermore, DELSSOME can replicate neurodevelopmental insights from our previous study doi.org/10.1073/pnas... /end
Congrats to Tianchu and Tian Fang for leading this study and also co-authors @shaoshiz.bsky.social @bart-larsen.bsky.social @ted-satterthwaite.bsky.social @avramholmes.bsky.social
Congrats to Tianchu and Tian Fang for leading this study and also co-authors @shaoshiz.bsky.social @bart-larsen.bsky.social @ted-satterthwaite.bsky.social @avramholmes.bsky.social
By embedding DELSSOME within an evolutionary optimization strategy, trained models generalize to new datasets without additional tuning, enabling a 50× speedup in mean field model estimation. The parameter estimates are equivalent to Euler integration 4/N
April 11, 2025 at 1:35 AM
By embedding DELSSOME within an evolutionary optimization strategy, trained models generalize to new datasets without additional tuning, enabling a 50× speedup in mean field model estimation. The parameter estimates are equivalent to Euler integration 4/N
Here we develop a deep learning approach (DELSSOME) to predict whether a biophysical model will produce realistic brain dynamics without numerical integration. When applied to a widely used mean field model, DELSSOME achieves a 2000× speedup over Euler integration 3/N
April 11, 2025 at 1:35 AM
Here we develop a deep learning approach (DELSSOME) to predict whether a biophysical model will produce realistic brain dynamics without numerical integration. When applied to a widely used mean field model, DELSSOME achieves a 2000× speedup over Euler integration 3/N
While the world burns, we cook up a new preprint! doi.org/10.1101/2025...
Biophysical modeling is a key tool to derive mechanistic insights into the brain. These models are governed by biologically meaningful parameters (unlike artificial neural networks), but the dirty secret ... 1/N
Biophysical modeling is a key tool to derive mechanistic insights into the brain. These models are governed by biologically meaningful parameters (unlike artificial neural networks), but the dirty secret ... 1/N
April 11, 2025 at 1:35 AM
While the world burns, we cook up a new preprint! doi.org/10.1101/2025...
Biophysical modeling is a key tool to derive mechanistic insights into the brain. These models are governed by biologically meaningful parameters (unlike artificial neural networks), but the dirty secret ... 1/N
Biophysical modeling is a key tool to derive mechanistic insights into the brain. These models are governed by biologically meaningful parameters (unlike artificial neural networks), but the dirty secret ... 1/N
Have you said thank you yet?
April 7, 2025 at 12:18 AM
Have you said thank you yet?
This is exactly what you get when you give terrorists everything they ask for without even getting the hostages back.
March 15, 2025 at 9:21 AM
This is exactly what you get when you give terrorists everything they ask for without even getting the hostages back.
2/2 noise, a *huge* amount of noise is needed to noticeably shift the optimal scan time (plot below). Sigma of 1 => 2x the noise in the original data, which only has a small effect! So we should still go for high SNR acquisition, but we probably need a quantum leap in SNR to see a big difference.
January 24, 2025 at 1:17 AM
2/2 noise, a *huge* amount of noise is needed to noticeably shift the optimal scan time (plot below). Sigma of 1 => 2x the noise in the original data, which only has a small effect! So we should still go for high SNR acquisition, but we probably need a quantum leap in SNR to see a big difference.
1/2 Thanks for the great Q! We are finalizing a new preprint version but at least among the datasets we considered, there was not a strong relationship with TR. Scan parameters obviously matter! But there are probably other complex factors beyond what we considered. In fact, if we purposely add...
January 24, 2025 at 1:17 AM
1/2 Thanks for the great Q! We are finalizing a new preprint version but at least among the datasets we considered, there was not a strong relationship with TR. Scan parameters obviously matter! But there are probably other complex factors beyond what we considered. In fact, if we purposely add...
Users can previously set a budget (e.g., $1M), and then find the sample size N & scan time T to maximize prediction accuracy within the budget.
Now users can also set a target accuracy (e.g., 80% of max accuracy), and then find N & T with the lowest cost that achieves the accuracy target.
Now users can also set a target accuracy (e.g., 80% of max accuracy), and then find N & T with the lowest cost that achieves the accuracy target.
January 24, 2025 at 12:02 AM
Users can previously set a budget (e.g., $1M), and then find the sample size N & scan time T to maximize prediction accuracy within the budget.
Now users can also set a target accuracy (e.g., 80% of max accuracy), and then find N & T with the lowest cost that achieves the accuracy target.
Now users can also set a target accuracy (e.g., 80% of max accuracy), and then find N & T with the lowest cost that achieves the accuracy target.
Thanks to everyone's feedback, we have updated our calculator to optimize sample size N & scan time T for fMRI studies: leonoqr.github.io/ORSP_Calcula...
The first new feature is that users can explore how different N & T leads to different accuracy, e.g., N=1000 & T=30min => 81% max accuracy. 🧵
The first new feature is that users can explore how different N & T leads to different accuracy, e.g., N=1000 & T=30min => 81% max accuracy. 🧵
January 24, 2025 at 12:02 AM
Thanks to everyone's feedback, we have updated our calculator to optimize sample size N & scan time T for fMRI studies: leonoqr.github.io/ORSP_Calcula...
The first new feature is that users can explore how different N & T leads to different accuracy, e.g., N=1000 & T=30min => 81% max accuracy. 🧵
The first new feature is that users can explore how different N & T leads to different accuracy, e.g., N=1000 & T=30min => 81% max accuracy. 🧵
OHBM has confirmed that Dec 17 is a hard deadline for abstracts.
December 12, 2024 at 10:54 PM
OHBM has confirmed that Dec 17 is a hard deadline for abstracts.