Runtime warning is raised when using Inverse ROI logic on Blob detection (OpenCV) generated ROIs

[StefanoPierini-Codra]

Summary

When creating ROIs running Blob detection (OpenCV) on an image, an Error is produced when one of this ROI is edited and "inverse ROI" is selected. In addition, the result of the inversion is not what expected.

[StefanoPierini-Codra]

Exploring this deeper, the warning skimage/exposure/exposure.py:605: RuntimeWarning: invalid value encountered in cast return (image * (omax - omin) + omin).astype(out_dtype) appears only when the ROI edited is at the same time modified (i.e. the coordinates of at least one point are changed.

This happens both using graphical edit and numerical edit.

Seems to be related to the recompute function, and the missing of the support of masked arrays from some scikit functions that are used to detect blobs.

An image of the warning is the following:

[StefanoPierini-Codra]

Two separate problems were identified:

1. Masked arrays not handled by detection algorithms: detection functions pass numpy.ma.MaskedArray directly to underlying libraries (scikit-image, OpenCV, scipy) that don't support masked arrays.
2. Wrong bounding box for inverse ROI: ImageObj.get_data(roi_index) crops the data to the bounding box of the ROI shape — but for an inverse ROI, the region of interest is the whole image minus the shape. The crop returns only the small rectangle around the shape, almost entirely masked, with useful pixels only in the corners between the shape and its bounding rectangle.

Root cause

Problem 1: Masked arrays in detection functions

ImageObj.get_data(roi_index) returns a MaskedArray when a ROI is defined. The detection functions in sigima/tools/image/detection.py pass this directly to algorithms that ignore or mishandle the mask:

• OpenCV (find_blobs_opencv): rescale_intensity casts NaN to uint8 → RuntimeWarning
• scikit-image (blob_dog/doh/log): ignores the mask, detects features in masked area
• scipy (get_2d_peaks_coords, get_hough_circle_peaks): same

Problem 2: Bounding box ignores inverse flag

get_data(roi_index) uses single_roi.get_bounding_box() to crop the data. This always returns the bounding box of the ROI shape itself. For an inverse ROI (where the exterior is the region of interest), this crops to the wrong area — the small rectangle inscribing the shape, where most pixels are masked.

The same issue affects compute_geometry_from_obj in sigima/proc/image/base.py, which applies a coordinate offset based on get_bounding_box() — incorrect for inverse ROIs where the data covers the full image.

Fix

1. sigima/tools/image/detection.py — Fill masked values before detection

All affected detection functions now convert masked arrays to regular arrays using data.filled(np.ma.median(data)), replacing masked pixels with the median of unmasked data. This provides a neutral background that minimizes detection artifacts at mask boundaries.

Affected functions: get_2d_peaks_coords, get_hough_circle_peaks, find_blobs_dog, find_blobs_doh, find_blobs_log, find_blobs_opencv.

Not modified: get_contour_shapes (already handles masked arrays correctly).

2. sigima/objects/image/object.py — Full-image bounding box for inverse ROI

In ImageObj.get_data(), when the ROI has inverse=True, the bounding box is now set to the entire image instead of the ROI shape:

if getattr(single_roi, "inverse", False):
    x0, y0, x1, y1 = 0, 0, self.data.shape[1], self.data.shape[0]
else:
    x0, y0, x1, y1 = self.physical_to_indices(single_roi.get_bounding_box(self))

Note: get_bounding_box() itself was not modified because other callers (ROI visualization, coordinate conversion) need the actual shape geometry.

3. sigima/proc/image/base.py — Post-filter and offset correction

Two changes in compute_geometry_from_obj:

• Post-filter: after detection returns coordinates (in pixel space), detections whose center falls on a masked pixel are discarded. This catches any remaining false positives from the fill value.
• Offset skip: for inverse ROIs, the ROI bounding box offset is skipped since the data already covers the entire image and coordinates are already in the full-image reference frame.

if getattr(single_roi, "inverse", False):
    pass  # No offset needed
else:
    x0, y0, _x1, _y1 = single_roi.get_bounding_box(obj)
    coords[:, colx] += x0 - obj.x0
    coords[:, coly] += y0 - obj.y0