Manual Correction

Manual Correction provides an interactive editing interface for coded images:

1. Visual overlay — The coded image is displayed with color-coded labels overlaid on the original fluorescence image, making it easy to identify where detection boundaries don't match visible nuclear boundaries.
2. Correction tools — Interactive tools for: Split (draw a line to divide one object into two), Merge (click two adjacent objects to combine them), Add (draw a new object boundary), Delete (remove a selected object).
3. Label management — New objects receive new unique labels. Merged objects inherit one label. Deleted objects' labels are retired. The coded image maintains unique, non-overlapping labels throughout.
4. Propagation — After correction, all dependent engines (measurements, classification, spatial analysis) can be recalculated on the corrected coded image.

Why automated analysis needs human review: Pawley's Confocal Handbook chapter on automated 3D analysis emphasizes that validation requires multiple independent assessments: "A single observer is insufficient" for establishing ground truth. Automated detection, like a single observer, has systematic biases — it consistently mishandles the same types of edge cases (touching cells, dim nuclei, unusual morphologies). Manual correction doesn't just fix individual errors; it compensates for the systematic blind spots of the algorithm.

The efficiency argument: A typical tissue section might contain 50,000 cells. Manual segmentation of each would take days. Automated detection processes them in seconds but may make 500 errors (1% error rate). Manual correction of those 500 errors takes minutes. The combination achieves near-perfect accuracy at a fraction of the time — 99%+ of the work is automated, and human expertise handles only the exceptions.

Error types in segmentation: Roysam et al. categorize segmentation errors as: false positives (detecting non-cells), false negatives (missing real cells), separation errors (merging or splitting), and boundary errors (incorrect object shape). Each type has different downstream consequences. False positives introduce noise into population statistics. False negatives bias measurements by excluding certain cell types. Separation errors distort per-cell measurements. Manual correction can address all four types, but separation errors are the most common and have the largest impact on measurement accuracy.

When is correction necessary? Not always. For large-scale studies where population-level statistics (mean, median, distribution) are more important than individual cell accuracy, a 1-2% detection error rate may be acceptable — the errors average out across thousands of cells. Manual correction becomes critical when: (1) individual cell identity matters (tracking specific cells across serial sections), (2) rare populations are being quantified (a few false positives could double the apparent count), or (3) the results will inform clinical decisions.

Preparing a tissue section for a clinical trial biomarker analysis:

1. Automated Nuclei Detection finds 32,000 cells
2. Review reveals ~300 merged doublets (two cells counted as one) in dense tumor regions
3. Manual split corrections separate each doublet → 32,300 individual cells
4. ~50 false positives (debris at tissue edges) are deleted → 32,250 validated cells
5. ~20 missed lymphocytes in dim regions are manually added → 32,270 final cells

Total correction time: approximately 15 minutes. The corrected detection ensures accurate phenotyping of the tumor microenvironment, where undercounting immune cells by 1% could affect the spatial interaction analysis that guides therapy selection.