Image brightness indicates lumen direction in a patient-derived virtual colon: a learning-free approach to semi-autonomous colonoscopy

Background

Automated colonoscope navigation requires a way to determine where the lumen continues. One possible approach is to estimate the distance from the endoscope tip to the surrounding wall. However, adding a range sensor to a conventional colonoscope is technically difficult, and depth estimation from monocular endoscopic images remains challenging. We considered a simpler possibility. Because the light source of an endoscope is located close to the camera, distant regions of the lumen tend to appear darker than nearby walls.

Aim

To determine whether lumen direction can be estimated from image brightness alone in a patient-derived virtual colon, and to compare this estimate with one obtained from true wall distances.

Methods

A virtual colon was reconstructed from the author’s CT colonography data (0.5 mm isotropic resolution). Endoluminal images were rendered at 342 positions along 1368 mm of the colonic centreline. At each position, the image was divided into 24 × 24 directional cells. Lumen direction was estimated from the darkness-weighted centroid of the largest connected component among the darkest 10% of cells. For comparison, the true distance from the virtual endoscope to the wall was obtained by ray casting, and a distance-based direction was calculated using the same method. Two illumination models were tested: a directional light without distance attenuation and a point source at the endoscope tip with distance attenuation.

Results

With point-source illumination, the brightness-based direction closely matched the distance-based direction, with a median angular difference of 2.5°. Differences greater than 30° occurred in 2.3% of frames. Within each frame, brightness and true wall distance were negatively correlated in all 342 frames (median Spearman ρ = −0.770).

Using the direction toward the centreline point 40 mm ahead as a fixed geometric reference, the median angular error was 6.6° for the brightness-based estimator and 6.5° for the distance-based estimator. The median frame-by-frame difference in error was −0.09°. In contrast, the brightness cue failed under directional illumination without distance attenuation (median difference from the distance-based direction, 43.1°).

Conclusion

In this patient-derived virtual colon, image brightness provided nearly the same lumen-direction information as exact wall distance when a point-source illumination model with distance attenuation was used. Exact metric distance was therefore not necessary for this specific navigation task. These findings suggest a simple, learning-free approach to semi-autonomous colonoscope steering based on image brightness and deterministic calculation.