Cochlear Inner Hair Cell Binary Segmentation
MONAI/PyTorch implementation of a 2D U-Net for binary semantic segmentation of fluorescent cochlear hair cells in single-channel two-photon microscopy TIFF images. This model segments 2-photon fluorescence microscopy images containing fluorescently-labelled hair cells in the cochlea. The model is trained to segment fluorescently-labelled inner hair cells (mostly GCaMP6) in the neonatal cochlea from in-vivo 2-photon imaging experiments (Paper). The original segmentation was performed with Cellpose and manually revised in napari.
As a result, the algorithm performs better than Cellpose in ignoring cells different from inner hair cells, such as outer hair cells or supporting cells, and in "blurry" in-vivo experiments.
Model
- Input: single-channel 2D TIFF image
- Output: binary mask (0 = background, 1 = hair cell), with optional watershed instance labels (0 = background, 1..N = individual cells)
- Architecture: MONAI model defined in
config.yml - Config:
config.yml - Weights:
best_model_binary_v1.pth
Use
Run inference on a TIFF image:
python infer_binary.py input_Avg.tif --output prediction_mask.tif
The script loads config.yml and the model path configured in model.best_model_path.
To output watershed-separated instance labels instead of a binary mask:
python infer_binary.py input_Avg.tif --output-type labels --output prediction_labels.tif
To save both outputs:
python infer_binary.py input_Avg.tif --output-type both --output prediction_mask.tif --labels-output prediction_labels.tif
Training Data
The model was trained on a microscopy dataset acquired as part of the paper In vivo spontaneous Ca2+ activity in the pre-hearing mammalian cochlea by De Faveri F., Ceriani F. and Marcotti W. The training dataset consisted of two-photon fluorescence images of GCaMP-labelled cochlear hair cells together with manually curated binary segmentation masks.
Evaluation
Validation Dice:0.787
Limitations
- Trained on mouse cochlear two-photon microscopy.
- Expected to perform best on images acquired under similar imaging conditions.
- Not evaluated on other microscopy modalities or species.
- Research use only.
Code
Training and inference code: https://github.com/fedeceri85/cochlea-hair-cell-binary-segmentation-unet2d