🧠 Training Your Own Networks¶

The DL_Track_US package GUI includes the possibility to train your own neural networks.

Why train your own model?¶

To create models tailored to your own dataset.
To improve segmentation if the example models don't generalize well enough.
To learn more about deep learning for muscle ultrasound.

🚨 It’s highly recommended to have a working GPU setup; otherwise training can take much longer.
Check out our GitHub repository for setup instructions.

If you're new to neural networks, we recommend this introduction course.

📝 Note:
DL_Track_US UI allows training but not modifying network architectures!

The paired images and labeled masks needed for training are located in:
📁 DL_Track_US_example/model_training

Download DL_Track_US Examples & Models if you haven’t already.

In this tutorial, we will train a model for aponeurosis segmentation.
Training a fascicle segmentation model is identical — only the images and masks would differ.

1. Data Preparation and Image Labeling¶

Inside the DL_Track_US_example/model_training folder:

📁 apo_img_example → Original images
📁 apo_mask_example → Corresponding labeled masks

model training folder

⚡ IMPORTANT:
Image names and mask names must match exactly!

Example:

image mask

2. Specifying Relevant Directories¶

Open the UI.
Click the Advanced Methods button.
In the dropdown, select Train Model.

advanced methods select method train_model

Now specify the directories:

Select Image Directory¶

Click Images.
Select DL_Track_US_example/model_training/apo_img_example.

images_button

Select Mask Directory¶

Click Masks.
Select DL_Track_US_example/model_training/apo_mask_example.

masks_button

Select Output Directory¶

Click Output.
Choose a folder to save the trained model, loss plots, and CSV results.

output_button

3. Image Augmentation (Optional but Recommended)¶

Image augmentation artificially increases your dataset size by applying random transformations.

🚨 Especially recommended if you have fewer than 1500 images.

Click Augment Images.

augment_button

A messagebox will notify you once augmentation is complete.

4. Specifying Training Parameters¶

Keep the default settings for this tutorial.
NEVER use just 3 epochs for real training.
(3 epochs are okay only for testing.)

hyperparameters

Now click:

Start Training

training_button

Three messageboxes will guide you during the training process.

Once training is finished, you’ll find:

Trained model (Test_Apo.h5)
Training loss plot (Training_Results.tif)
Loss values per epoch (Test_apo.csv)

5. Using Your Own Networks¶

You can use your trained models like this:

Click Apo Model or Fasc Model in the GUI to load your trained model.

apo model fasc model

⚡ IMPORTANT:
Never use the same images for training and inference.
Always validate on unseen data to check your model's performance.

If unsure, feel free to ask in our DL_Track_US Discussion Forum!

6. Error Handling¶

Errors during training trigger a messagebox:

error handling

Follow the instructions shown.
Uncaught errors should be reported in the
DL_Track_US Discussion Forum.

See here for guidance on how to best report errors.

7. Labeling Your Own Images¶

To train your networks, you must label images correctly.

🛠 We provide a semi-automated script!

You need:

📁 Original images folder
📁 output_images folder
📁 fascicle_masks folder
📁 aponeurosis_masks folder

You’ll use ImageJ/Fiji and our script:
🗂 DL_Track_US/DL_Track_US/gui_helpers/gui_files/Image_Labeling_DL_Track_US.ijm

Drag the .ijm file into a running Fiji/ImageJ window to start.

fiji labelling

Labeling Steps¶

Set Directories:
a. Input images
b. Aponeurosis masks
c. Fascicle masks
d. Output images
Label Aponeuroses:
Use polygon tool to select superficial and then deep aponeurosis.
Label Fascicles:
Use segmented line tool for clearly visible fascicle parts only.
Save and Move to Next Image.

✅ You are now ready to create your own high-quality training datasets!