5 open-source tools for microscopy, one production sports platform, and a transformer that reads 50 frames at once. Code and trained models for all of them.
Temporal Context Neural Networks for Microscopy
A U-Net with multi-head attention that localizes emitters from 50 frames of context instead of one. Validated on the SMLM Fight Club dataset.
To my knowledge the first application of transformer attention to SMLM data. The network links emitter positions across 50 frames, so 76.6% of its fits land below the CRLB.
Efficiency scores on CRLB dataset
Three microscopy tools and one sports platform.
Compressed Sensing + Deep Learning
Combines compressed sensing with convolutional neural networks for confocal dSTORM. Localizes emitters where the PSF is irregular and a Gaussian model does not fit. Published in BMC Bioinformatics.
Sports Analytics Platform
Web platform for endurance athletes. Imports training data from Strava and Garmin, computes training load, and predicts performance for triathlon and cycling. Live at endurexai.de.
Biological Structure Analysis
Detects line-shaped structures such as filaments in microscopy data and computes position, orientation and morphology with sub-pixel precision. Documented on ReadTheDocs.
3D Point Cloud Renderer
OpenGL point cloud renderer with a Python interface. Renders and filters 3D SMLM datasets with millions of points in real time.
How AttentionAI uses 50 frames of context
Convolutional layers extract spatial features from individual frames before temporal processing
Transformer attention learns temporal correlations across 50-frame sequences
Probabilistic localization with uncertainty quantification for each emitter
| Method | Efficiency Score | % Below CRLB | Context Frames |
|---|---|---|---|
| AttentionUNet | 94 | 76.6% | 50 |
| DECODE | 91 | 39.9% | 3 |
| ThunderSTORM | 23 | 17.5% | 1 |
* CRLB = Cramer-Rao Lower Bound (theoretical minimum uncertainty). Higher percentages below CRLB indicate better precision.
What I actually use.
Python
PyTorch
TensorFlow
CUDA
OpenCV
OpenGL
Django
PostgreSQL
Docker
FastAPI
Git
GitHub
Code, models and simulators for every paper.
All research projects include complete implementations, trained models, and simulation engines. This ensures other researchers can reproduce, extend, and build upon the work.
If one of these tools fits your problem, or almost fits, write to me.