Maciek Szul
LightNews LightNews
banner
maciekszul.bsky.social
Maciek Szul
@maciekszul.bsky.social
660 followers 370 following 700 posts
📍 Tübingen, DE ∆ DevComPsy Lab ∆🔬👨‍💻 🧪∆ decisions, motor control, (laminar) MEG, MRI ∆ not interested in oscillations ∆🤘🤡🚴🏔️ ∆ he/him ∆
Posts Replies Media Videos
maciekszul.bsky.social
When you are a massive nerd, you get excited by flashing lights. This LIDAR will be used for silly stuff.
November 23, 2025 at 12:58 PM
maciekszul.bsky.social
Look what I found today. On one hand a capsule capturing mad 90's downhill suspension geometries and on the other, the fact that someone should really get a bigger bike. Not making fun of it, because fork tube extensions are death.
November 18, 2025 at 5:23 PM
maciekszul.bsky.social
Is meditation analysis just sitting and thinking the results into existence?
November 11, 2025 at 1:54 PM
maciekszul.bsky.social
Watching the first episode of the 4th season of Witcher I had only one person in mind. I'm not happy to say the least.
November 10, 2025 at 8:16 AM
maciekszul.bsky.social
Sourdough is an environmentally responsible tamagochi.
November 4, 2025 at 11:55 AM
maciekszul.bsky.social
What a nice score today. The record shop had Handsome Boy Modelling School too, but I decided to spend my budget on Rollins Band.
Two used CDs of "Come in and Burn" and "Insert Band Here: Live in Australia 1990" by Rollins Band.
September 23, 2025 at 3:49 PM
maciekszul.bsky.social
It's the right time to remind you of our results, that during development beta happens at lower frequency. Moving up until adulthood. Better, mean burst waveform shape is the same, just slower.

doi.org/10.1523/JNEU...
September 15, 2025 at 6:25 PM
maciekszul.bsky.social
August 21, 2025 at 9:31 AM
maciekszul.bsky.social
What differentiates this account from a shed load of other, frankly repetitive, PD beta papers is the use of burst waveform shape. Phase and amplitude contains much more information than time frequency features. PCA helps to find a waveform shape range that is sensitive to Levodopa medication.
July 16, 2025 at 12:13 PM
maciekszul.bsky.social
🚨
Levodopa increased burst waveform-specific connectivity between STN and the sensorimotor cortex, which correlated with medication related improvement.

Beta burst waveforms look like a promising biomarker candidate for both disease severity and treatment response.

www.medrxiv.org/content/10.1...
July 16, 2025 at 12:13 PM
maciekszul.bsky.social
When you read about academic/scientific excellence, that's what they mean.
Picture of Dr. Best toothbrush.
June 21, 2025 at 9:58 AM
maciekszul.bsky.social
WETCOW.
June 10, 2025 at 12:14 PM
maciekszul.bsky.social
My favourite journal right now.
June 6, 2025 at 3:02 PM
maciekszul.bsky.social
We also found that the forward model has to match the true cortical column orientation for it to accurately recover the location of the source. Lots of work ahead to get an accurate estimate empirically!
Impact of dipole orientation estimation on laminar source reconstruction accuracy. a) Example dipole orientation vectors at a subset of cortical locations across layers for four methods: link vectors (LV), original surface normals (OSN), cortical patch statistics (CPS), and downsampled surface normals (DSN). b) Relationship between orientation mismatch (angular difference in degrees) and laminar reconstruction error, over all simulation–reconstruction method pairs and cortical depths. c–e) Angular deviation (median across vertices) between orientation estimation methods at three cortical depths: superficial (c, surface layer 1), middle (d, surface layer 6), and deep (e, surface layer 11). f–h) Corresponding reconstruction error (in layers) for each simulation–reconstruction method pair at the same depths, averaged across 100 cortical locations. Errors are lowest along the diagonal, where simulation and reconstruction methods match, and largest when methods differ substantially (LV – link vectors; OSN – original surface normals; CPS – cortical patch statistics; DSN – downsampled surface normals. Suffix “f” indicates dipole orientations were fixed across layers).
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
This figure features simulation across ca. 330000 vertices (1000s hours of cluster time) to show that anatomical features like: cortical thickness, gyrification, and orientation of cortical columns constrain the precision of the results beyond co-registration error and SNR
Anatomical factors influencing laminar reconstruction error. a-b) Inflated‐surface maps of laminar reconstruction error (a) and lead‐field RMSE (deep vs. superficial; b) c-h) Inflated-surface maps and joint density and marginal histograms for each anatomical metric versus lead-field RMSE: and dipole orientation (c,d), scalp–cortex distance (e,f), and cortical thickness (g-h). Joint density plots are color‐coded by laminar reconstruction error (mm). Lower lead-field differences,tangential orientations, greater distance to the scalp, and thinner cortex are associated with increased laminar‐reconstruction error.
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
Our 11 layer source space evenly samples the cortical thickness, but the true cortical laminae are not uniform across the brain. We used the BigBrain cytoarchitectonic atlas and mapped our results onto different laminae thicknesses. It still works!
g) Mean laminar reconstruction error (in laminae) across SNR levels, computed as the absolute difference between the simulated and inferred lamina. Lines indicate the mean error for each lamina, with shaded bands representing 95% confidence intervals. Lighter shaded bands represent shuffled null distributions. Asterisks mark SNR levels where reconstruction error was significantly lower than chance. h) Mean laminar reconstruction error as a function of co-registration error. i) Relationship between laminar thickness and reconstruction error. Each line represents a different lamina (I–VI), showing how increased laminar thickness is associated with reduced reconstruction error.
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
Co-registration error (HEAD MOVEMENT!) matters a lot as well. True laminar precision falls apart gracefully at an error bigger than 2 mm. Headcasts greatly contribute to surpassing the requirements for reduction in head movement.
Laminar source‐reconstruction accuracy as a function of co‐registration error at fixed 0 dB SNR. a–d) Mean free energy (ΔF) matrices (row‐normalized to the worst model; S = superficial surface layer, D = deep surface layer) for simulated signals placed on 11 cortical surfaces at 0, 1, 2, and 5 mm registration error. Triangles mark the best‐fitting forward model for each simulated depth. e–h) Probability density of each surface model being the maximum‐evidence solution, aggregated across cortical locations. i, j) Laminar reconstruction error (in surface layers and millimeters) for simulated versus shuffled data at varying registration errors. k, l) Corresponding bias (in surface layers and millimeters). Shaded bands denote 95% confidence intervals, and asterisks mark significance at p < 0.05.
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
However, when recording in MEG there are many ways it can go not so well. You need SNR greater than -35dB to achieve accurate source recovery. Doable when you look at phenomena like e.g. ERFs or beta burst waveforms (www.sciencedirect.com/science/arti...)
Laminar source‐reconstruction accuracy as a function of sensor‐level SNR. a–d) Mean model evidence (free energy, ΔF) matrices (row‐normalized to the worst model; S = superficial surface layer, D = deep surface layer) for simulated signals placed on 11 cortical surfaces at four SNR levels (−∞ dB, −50 dB, −35 dB, −20 dB). Accurate laminar inference begins to emerge around –35 dB (c), with near-perfect identification at –20 dB (d). Triangular markers denote the forward model with peak free energy for each simulated depth. e–h) Probability density of each forward model being the maximum‐evidence solution, aggregated across cortical locations: a diagonal pattern appears from –35 dB, indicating accurate depth localization. i, j) Laminar reconstruction error (in layers and mm) falls below chance at –35 dB and improves at higher SNR; shuffled controls show ~1 mm error due to chance matches between simulated and evaluated layers. k, l) Laminar reconstruction bias (in surface layers and millimeters), again comparing simulated and shuffled data. Shaded bands represent 95% confidence intervals, and asterisks mark SNR levels at which observed error or bias differed significantly from chance (p < 0.05).
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
We pick a vertex and simulate a signal located at one layer and perform a source reconstruction (with MSP) separately for each layer. We then compare the model evidence across depths. If all goes well, the depth where the source was located has the highest model evidence.
e) Each colored line represents the model evidence (relative to the worst model, ΔF) over source layer models, for a signal simulated at a particular layer (the simulated layer is indicated by the line color). The source layer model with the maximal ΔF is indicated by “˄”. f) Result matrix summarizing ΔF across simulated source locations, with peak relative model evidence marked with “˄”. g) Error is calculated from the result matrix as the absolute distance in mm or layers from the simulated source (*) to the peak ΔF (˄). h) Bias is calculated as the relative position of a peak ΔF(˄) to a simulated source (*) in layers or mm.
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
First you need a suitable source space. We have used an 11 layer mesh that evenly covers the thickness of the cortex between white matter and pial boundaries. Then we constrain the dipole orientation to vectors linking corresponding vertices.
Overview of the simulation strategy and analysis. a) Pial and white matter boundaries surfaces are extracted from anatomical MRI volumes. b) Intermediate equidistant surfaces are generated between the pial and white matter surfaces (labeled as superficial (S) and deep (D) respectively). c) Surfaces are downsampled together, maintaining vertex correspondence across layers. Dipole orientations are constrained using vectors linking corresponding vertices (link vectors). d) The thickness of cortical laminae varies across the cortical depth (70–72), which is evenly sampled by the equidistant source surface layers.
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
🚨🚨🚨PREPRINT ALERT🚨🚨🚨
Neural dynamics across cortical layers are key to brain computations - but non-invasively, we’ve been limited to rough "deep vs. superficial" distinctions. What if we told you that it is possible to achieve full (TRUE!) laminar (I, II, III, IV, V, VI) precision with MEG!
Overview of the simulation strategy and analysis. a) Pial and white matter boundaries surfaces are extracted from anatomical MRI volumes. b) Intermediate equidistant surfaces are generated between the pial and white matter surfaces (labeled as superficial (S) and deep (D) respectively). c) Surfaces are downsampled together, maintaining vertex correspondence across layers. Dipole orientations are constrained using vectors linking corresponding vertices (link vectors). d) The thickness of cortical laminae varies across the cortical depth (70–72), which is evenly sampled by the equidistant source surface layers. e) Each colored line represents the model evidence (relative to the worst model, ΔF) over source layer models, for a signal simulated at a particular layer (the simulated layer is indicated by the line color). The source layer model with the maximal ΔF is indicated by “˄”. f) Result matrix summarizing ΔF across simulated source locations, with peak relative model evidence marked with “˄”. g) Error is calculated from the result matrix as the absolute distance in mm or layers from the simulated source (*) to the peak ΔF (˄). h) Bias is calculated as the relative position of a peak ΔF(˄) to a simulated source (*) in layers or mm.
June 2, 2025 at 11:55 AM
maciekszul.bsky.social
French already embraced the meme (unintentionally).
Z.E. ELECTRIC RENAULT means Zero Emission Electric Renault, but also (in thick french accent) "ZIS IS ZE ELECTRIC RENAULT".
May 28, 2025 at 12:40 PM
maciekszul.bsky.social
2 bidons with 40 g of carbs each + salt and a 3/4 pack of watermelon Haribos later I got back. With that amount of climbing I was worried that I would bonk terribly, but it was alright. For the next round, I'm taking the train and eating all the kebabs.
May 18, 2025 at 4:19 PM
maciekszul.bsky.social
On the way back, I used a cycling highway between Sindelfingen and Böblingen. Pretty awesome thing. Extension is in the works.
May 18, 2025 at 4:19 PM
maciekszul.bsky.social
I got to Stuttgart and voted (very uneventful). The parts of the city I passed through are not bad. Seems livable. Was so tempted by all the kebab places, but caving in would mean a train home.
May 18, 2025 at 4:19 PM