3D Segmentation of Cochlea Stacks

3D Segmentation of Cochlea Stacks#

The 3D Segmentation feature in VASCilia leverages deep learning to segment cochlear hair cell stacks, providing precise identification of cellular regions for further analysis. This process is automated using a pretrained model and is seamlessly integrated into the workflow.

—

Key Features#

### Deep Learning-Based Segmentation

  • Utilizes a model trained on Detectron2, with the Mask R-CNN R-50 FPN 3x architecture.

  • Trained with 45 manually annotated cochlear stacks, ensuring high accuracy and reliability for segmentation.

  • Automatically segments 3D stacks to identify and analyze specific cellular regions.

### Integrated Workflow

  • Ensures segmentation occurs after the Rotation step, aligning the cochlear stacks with the PCP (planar cell polarity).

  • Updates the analysis stage upon successful segmentation, ensuring smooth progression through the pipeline.

### Customizable Parameters

  • Allows user configuration for:

  • Number of training iterations.

  • Threshold values for segmentation.

  • Output model paths for storing segmentation results.

—

Usage Instructions#

### Step 1: Prepare for Segmentation

  • Ensure that the stack is aligned using the Rotate functionality.

  • Verify that the analysis stage is set correctly; segmentation requires that the analysis stage is 3 (Rotated).

### Step 2: Execute Segmentation

  1. Click the Segment with 3DBundleSeg button in the plugin.

  2. The system will:

    • Prepare the dataset and paths.

    • Run the segmentation model using the configured parameters.

### Step 3: Monitor Progress

  • A progress dialog will display during segmentation, showing real-time updates.

  • Upon completion, results are saved in the designated output directory.

—

Technical Details#

### Pretrained Model

  • The segmentation model is trained using Detectron2 with:

  • Mask R-CNN R-50 FPN 3x architecture.

  • Trained with 45 manually annotated cochlear stacks to detect cellular regions with precision.

### Segmentation Command The plugin constructs and executes a command to perform segmentation, for example:

wsl {path_to_executable} --train_predict 1 \
--folder_path {train_folder} \
--model_output_path {output_model_path} \
--iterations {train_iter} \
--rootfolder {current_folder} \
--model {model_path} \
--threshold 0.7

### Progress Monitoring - The progress bar updates in real time by parsing output messages from the segmentation process. - The dialog closes automatically once the process completes.

—

Why Segmentation is Important#

Cochlear hair cells are intricate structures requiring detailed 3D analysis. Proper segmentation allows researchers to:

  • Identify cellular regions for focused study.

  • Quantify cellular dimensions and properties.

  • Ensure accurate downstream processing for visualization and analysis.

—

Practical Considerations#

  1. Pre-Requirements:

    • Rotation of the stack is mandatory before segmentation.

    • Verify that the required wsl_executable path and model paths are correctly set in the config.json.

  2. Error Handling:

    • If segmentation is attempted at the wrong stage, the plugin will notify the user to rotate the stack first.

    • If the stack has already been segmented, the plugin will prevent redundant processing.

  3. Output Files:

    • The segmented results are stored in the output folder configured in config.json.

—

Extending the Functionality#

To add or modify functionality, edit the following files:

  • segment_cochlea_action.py: Handles the segmentation process.

To fine-tune our model with your images, use the Training Sction

—

Segment Action Preprocessing Example