Manual Correction

Manual Correction is an interactive editor for coded images:

1. Visual overlay. The coded image appears as color-coded labels over the original fluorescence image, so you can see where a detection boundary and the visible nuclear boundary disagree.
2. Correction tools. Split divides an object in two, and Merge joins two objects into one. Add creates a new object from a shape you draw, and Remove takes a false detection out of the set. For Split, Merge and Remove you draw a shape over the objects concerned, and every object the shape touches undergoes that action.
3. Corrected output. Detection and correction are treated as a single process, so the detection's output becomes the corrected set of objects. The original, uncorrected set stays available as the Original Detection Results output, and each correction category also produces a mask image of its drawn shapes.
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: An algorithm is like a single observer with systematic biases: it mishandles the same kinds of edge cases every time, such as touching cells, dim nuclei and unusual morphologies. Pawley's Confocal Handbook, on automated 3D analysis, notes that "a single observer is insufficient" for establishing ground truth, so validation needs several independent assessments. Manual correction compensates for the algorithm's systematic blind spots, not only its individual errors.

The efficiency argument: A typical tissue section might hold 50,000 cells. Segmenting each by hand would take days; automated detection takes seconds but may make 500 errors (a 1% error rate), which take minutes to correct by hand. With more than 99% of the work automated and human expertise reserved for the exceptions, the combination reaches near-perfect accuracy in a fraction of the time.

Error types in segmentation: Roysam and colleagues sort segmentation errors into four kinds: false positives (detecting non-cells), false negatives (missing real cells), separation errors (merging or splitting) and boundary errors (the wrong object shape). False positives add noise to population statistics, false negatives bias measurements by leaving out certain cell types, and separation errors distort per-cell measurements. Manual correction can address all four. Separation errors are the most common and have the largest effect on measurement accuracy.

When correction is necessary: not always. When a large study depends on population statistics (mean, median, distribution) more than on individual cells, a 1-2% detection error rate may be acceptable, because the errors average out over thousands of cells. Manual correction becomes necessary in three cases: when individual cell identity matters, as in tracking specific cells across serial sections; when rare populations are being counted, since a few false positives could double the apparent count; and when a biomarker study's conclusions will rest on the results.

Here is a tissue section being prepared for a biomarker study:

1. Automated Nuclei Detection finds 32,000 cells.
2. Review shows about 300 merged doublets (two cells counted as one) in dense tumor regions.
3. Split corrections separate each doublet, giving 32,300 individual cells.
4. About 50 false positives, debris at the tissue edges, are removed, leaving 32,250 validated cells.
5. About 20 lymphocytes missed in dim regions are added by hand, for 32,270 final cells.

The whole correction takes about 15 minutes. The corrected detection ensures accurate phenotyping of the tumor microenvironment, where undercounting immune cells by 1% could change the result of the spatial interaction analysis.