aliceh0ng/cellpose-biosensor

Cellpose-SAM fine-tuned for bacterial segmentation in mouse colon tissue

Fine-tuned weights for Cellpose-SAM (cpsam) to segment individual bacteria in Zeiss Airyscan confocal images of mouse colon tissue sections.

Model description

The base model (cpsam) was fine-tuned on manually annotated BFP-channel patches from 16-bit TIFF stacks acquired on a Zeiss LSM 900 Airyscan at 100× magnification. Images contain bacteria expressing a BFP constitutive reporter, embedded in colon tissue with fecal autofluorescence.

Key imaging parameters:

• Pixel size: 0.035 µm/px
• Expected bacterial diameter: 43 px (1.5 µm)
• Input channel: BFP (constitutive, used for segmentation only)
• Stack size: (4, 6323, 6344) uint16

Available model runs

Run	Train patches	Val patches	Cells annotated	AP@0.5 base	AP@0.5 fine-tuned
4stacks_5x5_norm	75	25	3,913	0.641	0.724
5stacks	35	8	4,641	0.885	0.842
1stack	—	—	—	—	—

4stacks_5x5_norm is the recommended model (largest annotated patch set, best AP improvement over base).

How to use

from cellpose import models
import numpy as np

# Load fine-tuned model
model = models.CellposeModel(
    pretrained_model="path/to/models/finetuned/4stacks_5x5_norm/models/cpsam_4stacks_5x5_norm",
    gpu=True
)

# bfp_norm: float32 [0, 1] normalised BFP channel (2D array)
masks, flows, styles = model.predict(
    [bfp_norm],
    diameter=43,        # round(1.5 / 0.035)
    channels=[0, 0],
    normalize=True
)

Important: always normalise the BFP channel before inference (normalize=True or manual p1–p99 stretch). Measure fluorescence intensities from the raw uint16 stack, not the normalised image.

Training details

• Base model: cpsam (Cellpose-SAM)
• Epochs: 100
• Learning rate: 1e-5
• Weight decay: 0.1
• Batch size: 1
• Framework: Cellpose 4.x (train.train_seg)
• Hardware: NVIDIA GPU (CUDA 12.6)