cochlea-synapses-segmentation-unet2d

Cochlear Synapse ROI Segmentation

MONAI/PyTorch implementation of a two-stage 2D U-Net pipeline for ROI instance segmentation in single-channel cochlear microscopy TIFF images. This model was trained for the segmentation of fluorescently-labelled putative cochlear synaptic boutons.

The model uses:

• a binary foreground U-Net
• a center heatmap U-Net whose peaks are used as watershed markers

The final ROI labels are produced by marker-controlled watershed using the heatmap-derived markers inside the binary foreground prediction.

Model

• Input: single-channel 2D TIFF image
• Output: instance-label ROI mask (0 = background, 1..N = predicted ROIs)
• Optional outputs: binary foreground mask, center heatmap, marker image
• Architecture: MONAI U-Net, defined by the two YAML config files
• Binary config: config/binary_model_v5_synaptic.yml
• Heatmap config: config/heatmap_model_v5_synaptic.yml
• Weights: best_model_binary_heatmap_v5_synaptic.pth

Use

Run inference on a TIFF image:

python infer_rois.py input_Avg.tif --output input_rois.tif

The script loads:

best_model_binary_heatmap_v5_synaptic.pth
config/binary_model_v5_synaptic.yml
config/heatmap_model_v5_synaptic.yml

If --output is omitted, the script writes <input>_rois.tif.

To also save the intermediate binary mask, heatmap, and marker image:

python infer_rois.py input_Avg.tif --output input_rois.tif --save-intermediates

Useful inference parameters can be overridden from the command line:

python infer_rois.py input_Avg.tif \
  --output input_rois.tif \
  --binary-threshold 0.5 \
  --heatmap-threshold 0.3 \
  --min-size 20 \
  --min-distance 5

Training Data

The model was trained on manually annotated cochlear microscopy TIFF images with ROI masks and annotation-filtered center heatmaps generated from those masks.

The training pipeline and dataset preparation code are available in the GitHub repository: https://github.com/fedeceri85/cochlea-synapses-binary-segmentation-unet2d

Evaluation

The repository contains notebooks for checking binary predictions, heatmap predictions, and the combined ROI output:

binary_check_predictions.ipynb
heatmap_check_predictions.ipynb
binary_heatmap_check_predictions.ipynb

The combined ROI check notebook reports binary Dice for the foreground prediction and panoptic-quality components RQ/SQ for final ROI predictions.

Limitations

• Manual labels contain subjective decisions, so boundary and ROI-count disagreements can reflect annotation ambiguity.
• Research use only.

Code

Training and inference code: https://github.com/fedeceri85/cochlea-synapses-binary-segmentation-unet2d