The First Scribble Does Most of the Work

TL;DR for operators

When an automated system can be corrected repeatedly, the product question is not simply whether correction helps. It is how many corrections are worth asking a user to make.

Sherif and colleagues test that question in interactive multitracer PET/CT lesion segmentation.¹ Mean Dice rises from 0.5539 with no correction to 0.7223 after one scribble and 0.7512 after five. Lesion-level F1 follows the same pattern: 0.5281, then 0.7036, then 0.7326. Roughly 85% of the five-round gain arrives after the first correction.

For an imaging-product team, that makes the first correction disproportionately valuable. A workflow built around automated first-pass segmentation followed by one targeted intervention may capture much of the measured benefit without requiring a long correction sequence.

The mechanism is also deployable in an incremental sense: the authors add foreground and background correction channels to a pretrained CT/PET network while zero-initialising the new weights, so the expanded model initially behaves exactly like the pretrained model before learning how to respond to scribbles.

The boundary is equally important. The corrections are generated automatically from ground-truth errors, not by clinicians; 571 lesion-free scans do not enter the headline per-case Dice and F1 averages; and the reported cross-validation results are for fold-specific models rather than held-out test performance of the submitted five-model ensemble.

The interaction curve is front-loaded

Imagine a clinician reviewing an automatically segmented whole-body PET/CT study. The software marks suspected lesions. The reader corrects one error. The model updates. Then the workflow offers another correction, and another.

It would be reasonable to expect each extra round to keep buying meaningful accuracy. The reported results do not follow that pattern.

The first correction adds 0.1684 Dice points. The next four combined add only 0.0289. For lesion-level F1, the first correction adds 0.1755; rounds two through five together add 0.0290.

This changes the workflow question. The relevant variable is no longer just whether interaction improves segmentation. It is the marginal value of the next interaction.

The result is also consistent across the five validation folds for Dice, which increases at every round. Performance differences among the fold-specific models contract substantially as corrections accumulate: the Dice spread falls from 0.1729 at unaided inference to 0.0344 after five rounds, while the F1 spread falls from 0.1748 to 0.0371.

That suggests correction can partly compensate for variation in baseline model quality under this benchmark. It does not establish that five corrections are required to obtain that stabilisation in clinical use.

New correction inputs without breaking the pretrained model

Interactive capability creates an engineering problem before it creates a user-interface problem.

The starting network already consumes CT and PET. The interactive version must also accept two spatial inputs: one marking places that should be lesion foreground and another marking places that should be background.

Simply expanding the input layer risks disturbing a pretrained model before the new correction signal has been learned.

The authors avoid that by extending the model from two input channels to four while setting the weights connected to the new scribble channels to zero. At initialization, those channels contribute nothing. The model therefore reproduces the pretrained CT/PET mapping exactly. Fine-tuning can then teach the new pathways how a correction should alter the segmentation.

For teams extending an existing model, this is a useful architectural pattern: introduce a new control surface without forcing the deployed representation to begin from a newly perturbed function.

The system also has to cope with different PET tracers. The dataset combines 1,014 FDG and 597 PSMA examinations. PET intensity is normalized relative to a scan-specific blood-pool reference, primarily estimated from the aorta, before transformation and standardisation. That preprocessing is intended to reduce dependence on tracer-specific intensity scales without requiring tracer-specific lesion labels.

Most of the gain comes from finding lesions that were missed

The aggregate score increase does not come from all error types equally.

Across five correction rounds, 829 false negatives are converted into true positives. False positives, meanwhile, rise slightly from 6,869 to 6,998.

That pattern indicates that the measured improvement is driven primarily by recovering missed lesions rather than cleaning up erroneous detections.

It also complicates how the two correction channels should be interpreted. Background scribbles account for 69% of the 5,426 generated corrections, yet the paper argues that they contribute relatively little to the improvement.

The proposed explanation is a semantic mismatch. During training, background scribbles resemble thin regions near true lesions. During evaluation, background scribbles instead identify regions where the model has produced false-positive predictions. The channel is therefore asked at evaluation time to represent a meaning that differs from the signal it predominantly encountered during training.

This is not a controlled ablation proving the isolated causal contribution of foreground versus background prompts. It is an error-analysis-supported explanation. But it raises a broader product-development issue: an interaction signal is only useful if its training semantics resemble what users will actually communicate through the interface.

A one-correction workflow is a hypothesis worth testing

The paper directly shows that simulated sparse correction substantially improves segmentation under its five-fold autoPET/CT V protocol, and that most of the measured improvement arrives immediately.

Cognaptus’ inference is narrower: teams building interactive imaging systems should test whether the economically relevant unit of interaction is the first targeted correction, rather than assuming that an interface should be optimized for repeated correction cycles.

That affects several design decisions. A product may need to make the first error easy to identify and correct, return an updated segmentation quickly, and measure whether subsequent corrections still justify clinician attention. An interface designed around five rounds because a benchmark permits five rounds could impose workflow cost without proportional measured benefit.

The strongest product experiment would therefore compare not only model accuracy but also reader time, correction frequency, latency, residual false positives and false negatives, and the probability that a second or third intervention changes a clinical decision.

The benchmark does not yet establish a clinical stopping rule

Three boundaries prevent the interaction curve from becoming a deployment rule.

First, the challenge simulator generates corrections from ground truth. Every scribble therefore points to a genuine model error and is cleaner than ordinary human interaction is likely to be.

Second, the headline per-case Dice and lesion-level F1 averages cover 1,040 tumour-bearing studies. The remaining 571 lesion-free studies are excluded because the evaluator treats those per-case metrics as undefined when no annotated lesion exists. False positives on lesion-free examinations therefore do not influence those macro averages.

Third, the cross-validation analysis evaluates each held-out fold using its corresponding fold-specific model. It does not report held-out challenge test performance for the five-model ensemble submitted to the competition.

Those limitations do not erase the front-loaded pattern. They define what it currently supports: a strong reason to test whether one well-designed correction captures most of the value in real clinical workflows, not evidence that one correction is already the optimal clinical policy.

The paper’s most transferable result is therefore less about scribbles than about interaction economics. More opportunities for human correction do not automatically mean proportionally more value. Measure where the improvement actually arrives, then design the workflow around that curve.

Cognaptus: Automate the Present, Incubate the Future.