3DL_NuCount model

Model author: Fabrice Daian

Model description

3DL_NuCount model has been designed by fine tuning a pretrained Stardist3D model [1,2] using a home made dataset [3] in order to assess the number of cells present in a given 3D image stack acquired using an optical microscope. Training and Inference Notebooks are hosted on our Github repo [4].

Stardist Training parameters

• patch size: (48,96,96)
• batch size: 32
• epochs : 100
• data augmentation : flip/rotation/intensity
• image normalization: normalize channel independantly
• anisotropy: empirical
• rays : 96

Training dataset parameters

• tile size : (4,63,128,128)
• split : Train 0.8 / Val 0.2

Inference

• patch size : (784,784,:)
• image size : (2048,2048,:)
• model name : weights_best_1.h5
• config file : config.json
• threshold file : thresholds.json

References

• [1] Stardist Project: Github
• [2] Stardist Paper : ArXiv
• [3] NuCount Training Dataset : Zenodo
• [4] NuCount Github Project: Github