---

name: deconvolution-intensity-scaling description: "KINTSUGI deconvolution: Output scaling must preserve original intensity relationships for quantitative analysis. Trigger: deconvolution saturates at 65535, channel intensity ratios inverted, hist_clip scaling issues." author: KINTSUGI Team date: 2026-01-15

Deconvolution Intensity Scaling - Preserving Quantitative Relationships

Experiment Overview

Context

After deconvolution, all channels were saturating at 65535 (max uint16) regardless of their original intensity ranges. This caused:

• Artificial clipping of bright regions
• Inverted relative intensity relationships between channels
• Broken quantitative marker expression analysis

Root Cause

The histogram clipping and rescaling logic in notebooks/Kdecon/main.py (lines 265-289) always stretched output to full 16-bit range:

Copy# BROKEN CODE (before fix)
if raw_max <= 65535:
    scale = 65535  # <-- Always stretched to full range!
else:
    scale = raw_max
result = result * scale

Impact Analysis

Critical: The CH1/CH2 intensity ratio was inverted from 14.28 to 0.71.

Solution

Change the scaling to preserve original intensity relationships:

Copy# FIXED CODE
# Scale to output range - preserve original intensity scale
# Using raw_max maintains relative intensity relationships between channels,
# which is critical for quantitative analysis of marker expression
scale = raw_max

result = result * scale

Fixed Results

Failed Attempts (Critical)

Diagnostic Checklist

When deconvolved images show intensity issues:

1. Check for saturation: np.sum(img >= 65535) should be near zero
2. Compare channel ratios: Before/after deconvolution should be similar
3. Check mean inflation: Deconvolved mean shouldn't be dramatically higher than input
4. Trace the scaling pipeline: Input → deconv → clip → normalize → scale → output

Key Parameters

Key Insights

• Histogram clipping normalizes to 0-1: The hist_clip parameter clips percentiles, then normalizes the result to 0-1 range. The final scaling determines the output range.
• Weak channels are most affected: Channels with low mean intensity show the largest inflation when stretched to full range.
• Relative relationships matter for quantitative analysis: Marker expression comparisons require consistent scaling across channels.
• Test with multiple channels: Single-channel testing won't reveal ratio inversion bugs.

Files Modified

• notebooks/Kdecon/main.py: Lines 283-288 (scaling logic)
• CLAUDE.md: Added documentation in Recent Enhancements section

References

• notebooks/Kdecon/main.py: KDecon class and decon() function
• notebooks/2_Cycle_Processing.ipynb: Deconvolution workflow
• Related skill: lightsheet-psf-deconvolution (PSF-related deconvolution issues)