📰 News¶
v0.3.0 Release Notes¶
- Released version 0.3.0 with major upgrades and bugfixes!
- New features: anual scaling tool, resize Video tool, crop video length tool & remove video parts tool.
- Faster model predictions & optional stacked (sequential) predictions.
- Improved user interface with visualization of model predictions and filtering/plotting of results.
- Automatic
settings.json
in GUI for easy switching of model parameters. - Filtering of fascicle length and pennation angle data using hampel sand savgol filters.
👉 See the Full Changelog on GitHub
Faster model predicitions on GPU & CPU¶
In the new version, we reduced processing time per frame by 40% from version 0.2.1 on GPU and CPU to 0.6s and ... , respectively.
Improved user interface¶
In version 0.3.0 we improved the user interface and included real time visualization of model predictions as well as a results terminal at the end of analyis. The analysis process is now more transparent and felxibel, since we included more analysis options in the settings.
New model with bi-directional short long term memory for video analysis¶
We further provide a new model with a new overall aproach for fascicle anylsis in videos. For the first time, we provide a model with memory and awareness of surrounding frames. The model is taken from Chanti et al. (2021) and is called IFSS_NET.
In our approach, we use a bi-directional short long term memory (BiLSTM) to capture the temporal context of the video. We excluded the siamese encoder from the orginal model. Furhtermore, we used a hybrid loss combination of the Dice loss and binary cross entropy loss, both weighted equally.
To reach this decision, we compared different models and their performance compared to a manual ground thruth and a kalman-filter based tracking apporach (UltraTimTrack) proposed by van der Zee et al. (2025).
Model Training results¶
We compard our previous vgg16unet model (Ritsche et al. (2024)) to SegFormer, uNet3+ and IFSS-Net architectures. The Results on a unseen test set of 120 images with examplary predictions can be seen below.
Moreover, we compared the models due to similar performance to the one of the validation videos from the original paper (Ritsche et al. (2024)). This video was recently used to compare the performance of different methods for fascicle tracking (van der Zee et al. (2025)). We demonstrate improvement in the results from DL_Track_US in terms of RMSD compared to manual annotation as displayed below. Of all networks, the IFSS-Net model performed best in a trade-off between pennation angle and fascicle length RMSD.
Note that, compared to v0.2.1, we introduced hampel-filtering of the fascicle values in each frame and additionally applied a savitzky-golay filter to the median fascicle data to furhter reduce root mean squared distance. The results for three different tasks are displayed below.
Calf Raise¶
VL fixed end maximal knee extentsion¶
🚨 More comparsions will follow in the upcoming publication.
🚨 We are currently working on implementing tracking of fascicles accounting for their curvature.