Projection Images

Suite2p computes several reference images during registration and detection. These images are used for quality assessment and as inputs to Cellpose anatomical segmentation.

Processing Pipeline

Images stored in ops.npy are computed at different stages:

Step	Description	Disable
Temporal Binning	Average consecutive frames to reduce noise	ops["nbinned"] = nframes
Temporal High-Pass	Remove slow baseline drift from movie	ops["high_pass"] = 0*

* The MBO fork allows disabling high_pass; standard Suite2p always applies it.

What’s stored:

• meanImg - mean of binned movie (computed before HP filter)

• meanImgE - enhanced mean from registration (spatial high-pass of meanImg); GUI display only, not a Cellpose input

• max_proj - max of binned + HP filtered movie

• Vcorr - correlation map or cellpose input image

Cellpose Input Images

Cellpose segments one static image, chosen by cellpose_settings.img. Suite2p 1.1.0 accepts three values:

img	Image	Description
"max_proj / meanImg"	log(max_proj / mean_img)	default; log ratio emphasizes active regions
"meanImg"	mean_img	temporal mean
"max_proj"	max_proj	maximum projection of HP-filtered movie

The enhanced mean image (meanImgE) is no longer a Cellpose input; the old anatomical_only=3 was removed upstream.

Selecting the image (recommended). Enable Cellpose with algorithm="cellpose" and pick img:

ops = {"algorithm": "cellpose", "img": "max_proj / meanImg"}  # default
ops = {"algorithm": "cellpose", "img": "meanImg"}
ops = {"algorithm": "cellpose", "img": "max_proj"}

Legacy. The integer anatomical_only still works and maps to img:

ops = {"anatomical_only": 1}   # img = "max_proj / meanImg" (default)
ops = {"anatomical_only": 2}   # img = "meanImg"
ops = {"anatomical_only": 4}   # img = "max_proj"

anatomical_only=0 (or algorithm="sparsery"/"sourcery") uses functional detection. anatomical_only=3 warns and falls back to 1.

The three Cellpose input images. meanImg: temporal mean. max_proj: maximum projection of the HP-filtered movie. max_proj / meanImg (default): log ratio highlights active regions.

The image is built exactly as in suite2p/detection/anatomical.py::select_rois:

if img == "max_proj / meanImg":
    out = np.log(np.maximum(1e-3, max_proj / np.maximum(1e-3, mean_img)))
elif img == "meanImg":
    out = mean_img
elif img == "max_proj":
    out = max_proj

Spatial High-Pass Filter

spatial_hp_cp (upstream cellpose_settings.highpass_spatial; integer, default 0) sharpens the chosen img before Cellpose. It is separate from the temporal high_pass used during detection, and does not modify the stored meanImg or max_proj.

Effect of spatial_hp_cp values (0, 0.5, 1, 3) on the max projection. Higher values sharpen cell boundaries but may amplify noise.

When spatial_hp_cp > 0, the image is normalized to [0, 1] and a Gaussian-blurred copy (sigma = diameter × spatial_hp_cp) is subtracted twice:

from scipy.ndimage import gaussian_filter
from cellpose.transforms import normalize99

def apply_hp_filter(img, diameter, spatial_hp_cp):
    img = np.clip(normalize99(img), 0, 1)
    sigma = diameter * spatial_hp_cp
    img = img - gaussian_filter(img, sigma)
    img = img - gaussian_filter(img, sigma)
    return img

Recommendations

ops = {
    "algorithm": "cellpose",         # enable Cellpose
    "img": "max_proj / meanImg",     # detection image (default)
    "spatial_hp_cp": 0,              # increase to sharpen the input image
    "diameter": 6,                   # cell size in pixels
    "cellprob_threshold": 0.0,
    "flow_threshold": 0.4,
}

If detection is poor:

• Try img="max_proj" or img="meanImg"

• Increase spatial_hp_cp to 1-3 to sharpen cell boundaries

• Adjust diameter to match cell sizes

• img="max_proj / meanImg" (default) suits data with strong baseline fluorescence

See Also

• User Guide - Complete parameter reference

• Cellpose Documentation - Cellpose model details