Pre-registration, result and change of course: a written rule against a U-Net and Segment Anything on 2,000 unopened dermoscopic images whose reference outlines were, in 99%, drawn by an automated method and accepted by an expert, 17 Nov 2026

New version (version 8): criterion of an exploratory analysis made after the results, deposited before it is run. Version 7 was published by 13:46:52 Brasilia time on 17 November 2026 (16:46:52 UTC, the "modified" field of its record in the Zenodo API). Question: how much the mean Dice of the written rule against the expert hand tracing would change if the rule's outline were widened or narrowed, by lesion size. Images: the 2,000 images of the first comparison and the 2,197 images, both with hand tracing by an expert (IMA++, tool T1), the same two banks as in the run sealed in version 4, without the images in which the tracing reaches the edge of the illuminated field. Margin: from -30% to +60% of the equivalent radius of the rule's own outline, in steps of 2.5%; size by the rule's own class (small up to 5% of the image, medium 5% to 20%, large above 20%). For each size, the margin with the highest mean Dice and the gain over no margin, with 95% bootstrap intervals; the margin chosen in one bank is then measured in the other bank, in both directions. Only the rule; no registered result changes, and the rule itself is not changed. This is exploratory: both banks had been opened and used in the development of the rule, so the numbers will be inflated, and any margin found here does not enter the rule; if it ever does, it will be a new version tested on unopened images. Declared as already seen: on 18 October 2026, an earlier version of the rule widened by 5% and by 10% of the radius tied with the network on the 2,000 images; on 19 October, a fixed margin of 8% of the radius was measured; on 28 October, a version that went to the last transition to skin was tested and not adopted. The author decided not to add a fixed margin to the rule: the edge is found in the photograph, not by the width of the reference. Before this deposit, a test on 20 images checked only the outline without margin against the stored values and the running time; no margin was looked at. The files are the criterion (with two addenda written before any margin was computed), the predictions (the Otsu's; the author gave none, in his words "it is mathematical"), the SHA-256 of the scripts that will run, and the timing of version 7. Nothing registered in versions 1 to 7 was changed; their files are kept here. Added in version 8: CRITERIO_MARGEM_17NOV.txt; PREVISOES_MARGEM_17NOV.txt; REGISTRO_HASHES_MARGEM_17NOV.txt; ZENODO_V7_PUBLICADO_17NOV.txt. New version (version 7): result of the two-judge test sealed in version 6. Version 6 was published by 13:05:13 Brasilia time on 17 November 2026 (16:05:13 UTC, the “modified” field of its record in the Zenodo API), after the seal (12:56:01); the run followed and the result was written at 13:17. Before counting, the run reproduced the per-image records of 15 November in all 164 images (largest difference 0.000000). As written before the run: this test is not conclusive, the sample is small, and the result enters only as a possible clue, never as a conclusion. Main count, the same as in the run sealed in version 4, on the 149 images without frame: with the doctor judge (hand tracing, stored Dice values), rule 0.8065 and U-Net 0.8113, rule minus U-Net -0.0048 (95% CI -0.0398 to +0.0277), inconclusive (known before the seal); with the machine judge (automated outline, Dice from the stored outlines), rule 0.7806 and U-Net 0.7699, rule minus U-Net +0.0107 (95% CI -0.0224 to +0.0466), non-inferior. The doctor judge computed from the outlines gave -0.0041, 0.0007 away from the stored values. Beside it, the change in the U-Net’s advantage when the judge changes, image by image: -0.0148 (95% CI -0.0360 to +0.0059), inconclusive. When the judge changed from the doctor to the machine, the U-Net lost more (-0.0398, 95% CI -0.0585 to -0.0218) than the rule (-0.0250, 95% CI -0.0461 to -0.0045). The reading changed with the judge, but in the direction opposite to the hypothesis, which expected the rule to come out inferior with the machine judge: in these images the clue goes against the hypothesis that the network has an affinity with machine-drawn outlines. In 133 of the 164 images the automated outline was accepted by a novice reviewer, not by an expert as in lot 1; in the 29 expert-accepted images without frame, the change was +0.0072 (95% CI -0.0528 to +0.0715). With a sample this small, the result neither confirms nor refutes the hypothesis; a large study with unopened images traced blind by independent specialists is the next step. The score of the predictions is in PLACAR_PREVISOES_JUIZES_17NOV.txt: the main predictions of the Fable and of the Otsu (that the network would gain with the machine judge) were refuted; the Otsu’s predicted width of the interval held (0.021); Fabio wrote “I do not know” and, before the run, that the sample was too small for a conclusive answer. The per-image file stays on the author’s computer. Nothing registered in versions 1 to 6 was changed; their files are kept here. Added in version 7: RESULTADO_JUIZES_17NOV.txt; RESULTADO_JUIZES_17NOV.txt.sha256; PLACAR_PREVISOES_JUIZES_17NOV.txt; ZENODO_V6_PUBLICADO_17NOV.txt. New version (version 6): seal of the two-judge test, deposited before the run. Among the 2,197 images, 164 also have in IMA++ an automated outline (tool T3) besides the expert hand tracing (T1 S1) used so far. They were chosen from metadata only (LISTA_164_JUIZES_17NOV.json); when an image has two automated outlines, one was drawn with seed 20261117. The automated outlines were opened only to check their md5 and size; no Dice against them has been computed. On these same images, the rule (package v2, unchanged) and the U-Net (outline stored on 5 November) will be scored against two judges: the hand tracing (the doctor) and the automated outline (the machine). The main count is the same as in the run sealed in version 4, done with each judge: rule minus U-Net, without the frame images (149 images), 95% interval by bootstrap (2,000 resamples, seed 20261200), read with the margin of 0.023; an error is a Dice below 0.05. With the doctor judge the stored Dice values are used. With the machine judge there is no stored value, so the Dice comes from the stored outlines, the same way for both methods, and the doctor judge computed the same way is shown beside. Also beside: how much the U-Net’s advantage changes when the judge changes, image by image. The question: does the reading change when the judge changes? Written before the run, and stated plainly: this test is not conclusive. The sample is small, and the bank is limited: the images with an unused expert hand tracing ran out, because they were used in the earlier comparisons. Whatever comes out enters only as a possible clue, pointing to a hypothesis still to be proved, never as a conclusion. Already known before the seal, from stored values: with the doctor judge on these 149 images, rule minus U-Net is -0.0048 (95% CI -0.0398 to +0.0277), which reads as inconclusive. In 133 of the 164 images the automated outline was accepted by a novice reviewer (S2), not by an expert as in lot 1 (99%); lines for expert-accepted only (29 images without frame) and novice-accepted only (120) are shown beside. The rule runs only to save its outline, and its run must reproduce the per-image records of 15 November in all 164 images, or no count is made. A check of the same path on 16 other images of the 2,197 passed (identical records). Deposited here: the criterion, the predictions of the three of us, the image list, the check result, the seal (LACRE_JUIZES_17NOV.txt, SHA-256 311ae71ae770b0a9ba83f63726d87c407ee6208ebd8d287b2d5daeaa21e55c5c) and the list of hashes. The scripts are not deposited; their hashes are in the seal. Nothing registered in versions 1 to 5 was changed; their files are kept here. Added in version 6: CRITERIO_TESTE_DOS_DOIS_JUIZES_17NOV.txt; PREVISOES_FABLE_E_FABIO_JUIZES_17NOV.txt; PREVISOES_OTSU_JUIZES_17NOV.txt; LISTA_164_JUIZES_17NOV.json; RESULTADO_TESTE_JUIZES_17NOV.txt; LACRE_JUIZES_17NOV.txt; REGISTRO_HASHES_JUIZES_17NOV.txt. New version (version 5): result of the descriptive run sealed in version 4. Version 4 was published on 17 November 2026 by 08:49:53 Brasilia time (11:49:53 UTC, the “modified” field of this record in the Zenodo API); the run wrote its first file at 08:53:03 and the result at 10:28. The seal, the three scripts and the result file were checked against their hashes. This is a description, with no registered verdict: the images had been seen and both banks were used to develop the rule, so the numbers are inflated, those of the 2,197 most of all. Against the expert hand tracing, without the frame images: on the 2,000 images of the first comparison (n = 1,839), the rule (package v2) had a mean Dice of 0.7740 and the U-Net 0.7548 (stored value), rule minus U-Net +0.0192 (95% CI +0.0085 to +0.0297); on the 2,197 (n = 2,017), 0.7739 and 0.7552, rule minus U-Net +0.0187 (95% CI +0.0082 to +0.0291). The two banks share no image. Read with the margin of 0.023 used in lot 1, both come out “superior”. With the frame images included, both come out “non-inferior” (rule minus U-Net -0.0039 and -0.0034); on the frame images alone the rule is far behind (0.56 against 0.83 on the 2,000; 0.56 against 0.81 on the 2,197). Images with a Dice below 0.05: rule 61, U-Net 76 on the 2,000; rule 70, U-Net 89 on the 2,197. By lesion size, the rule is ahead on small lesions and behind on medium and large ones. Lot 1 (reference drawn by a machine in 99% of images), sealed result shown only for comparison: rule 0.8157, U-Net 0.8619, rule minus U-Net -0.0462. The score of the three sets of predictions is in PLACAR_PREVISOES_MAOS_17NOV.txt: the Fable’s predictions for the 2,000 held; two of the Otsu’s items were refuted (the rule came out ahead by more than 0.017, and the swap T changed the mean Dice by 0.0028); Fabio predicted the direction and overestimated the values. The score sheet also declares an error in the Ot

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24267497
Primary Topic
Cutaneous Melanoma Detection and Management
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Pre-registration, result and change of course: a written rule against a U-Net and Segment Anything on 2,000 unopened dermoscopic images whose reference outlines were, in 99%, drawn by an automated method and accepted by an expert, 17 Nov 2026

Fabio da Camara Oliveira Ferreira
4 links
Zenodo (CERN European Organization for Nuclear Research)
Cutaneous Melanoma Detection and Management
article

Pre-registration, result and change of course: a written rule against a U-Net and Segment Anything on 2,000 unopened dermoscopic images whose reference outlines were, in 99%, drawn by an automated method and accepted by an expert, 17 Nov 2026

Fabio da Camara Oliveira Ferreira
article en

Abstract

New version (version 8): criterion of an exploratory analysis made after the results, deposited before it is run. Version 7 was published by 13:46:52 Brasilia time on 17 November 2026 (16:46:52 UTC, the "modified" field of its record in the Zenodo API). Question: how much the mean Dice of the written rule against the expert hand tracing would change if the rule's outline were widened or narrowed, by lesion size. Images: the 2,000 images of the first comparison and the 2,197 images, both with hand tracing by an expert (IMA++, tool T1), the same two banks as in the run sealed in version 4, without the images in which the tracing reaches the edge of the illuminated field. Margin: from -30% to +60% of the equivalent radius of the rule's own outline, in steps of 2.5%; size by the rule's own class (small up to 5% of the image, medium 5% to 20%, large above 20%). For each size, the margin with the highest mean Dice and the gain over no margin, with 95% bootstrap intervals; the margin chosen in one bank is then measured in the other bank, in both directions. Only the rule; no registered result changes, and the rule itself is not changed. This is exploratory: both banks had been opened and used in the development of the rule, so the numbers will be inflated, and any margin found here does not enter the rule; if it ever does, it will be a new version tested on unopened images. Declared as already seen: on 18 October 2026, an earlier version of the rule widened by 5% and by 10% of the radius tied with the network on the 2,000 images; on 19 October, a fixed margin of 8% of the radius was measured; on 28 October, a version that went to the last transition to skin was tested and not adopted. The author decided not to add a fixed margin to the rule: the edge is found in the photograph, not by the width of the reference. Before this deposit, a test on 20 images checked only the outline without margin against the stored values and the running time; no margin was looked at. The files are the criterion (with two addenda written before any margin was computed), the predictions (the Otsu's; the author gave none, in his words "it is mathematical"), the SHA-256 of the scripts that will run, and the timing of version 7. Nothing registered in versions 1 to 7 was changed; their files are kept here. Added in version 8: CRITERIO_MARGEM_17NOV.txt; PREVISOES_MARGEM_17NOV.txt; REGISTRO_HASHES_MARGEM_17NOV.txt; ZENODO_V7_PUBLICADO_17NOV.txt. New version (version 7): result of the two-judge test sealed in version 6. Version 6 was published by 13:05:13 Brasilia time on 17 November 2026 (16:05:13 UTC, the “modified” field of its record in the Zenodo API), after the seal (12:56:01); the run followed and the result was written at 13:17. Before counting, the run reproduced the per-image records of 15 November in all 164 images (largest difference 0.000000). As written before the run: this test is not conclusive, the sample is small, and the result enters only as a possible clue, never as a conclusion. Main count, the same as in the run sealed in version 4, on the 149 images without frame: with the doctor judge (hand tracing, stored Dice values), rule 0.8065 and U-Net 0.8113, rule minus U-Net -0.0048 (95% CI -0.0398 to +0.0277), inconclusive (known before the seal); with the machine judge (automated outline, Dice from the stored outlines), rule 0.7806 and U-Net 0.7699, rule minus U-Net +0.0107 (95% CI -0.0224 to +0.0466), non-inferior. The doctor judge computed from the outlines gave -0.0041, 0.0007 away from the stored values. Beside it, the change in the U-Net’s advantage when the judge changes, image by image: -0.0148 (95% CI -0.0360 to +0.0059), inconclusive. When the judge changed from the doctor to the machine, the U-Net lost more (-0.0398, 95% CI -0.0585 to -0.0218) than the rule (-0.0250, 95% CI -0.0461 to -0.0045). The reading changed with the judge, but in the direction opposite to the hypothesis, which expected the rule to come out inferior with the machine judge: in these images the clue goes against the hypothesis that the network has an affinity with machine-drawn outlines. In 133 of the 164 images the automated outline was accepted by a novice reviewer, not by an expert as in lot 1; in the 29 expert-accepted images without frame, the change was +0.0072 (95% CI -0.0528 to +0.0715). With a sample this small, the result neither confirms nor refutes the hypothesis; a large study with unopened images traced blind by independent specialists is the next step. The score of the predictions is in PLACAR_PREVISOES_JUIZES_17NOV.txt: the main predictions of the Fable and of the Otsu (that the network would gain with the machine judge) were refuted; the Otsu’s predicted width of the interval held (0.021); Fabio wrote “I do not know” and, before the run, that the sample was too small for a conclusive answer. The per-image file stays on the author’s computer. Nothing registered in versions 1 to 6 was changed; their files are kept here. Added in version 7: RESULTADO_JUIZES_17NOV.txt; RESULTADO_JUIZES_17NOV.txt.sha256; PLACAR_PREVISOES_JUIZES_17NOV.txt; ZENODO_V6_PUBLICADO_17NOV.txt. New version (version 6): seal of the two-judge test, deposited before the run. Among the 2,197 images, 164 also have in IMA++ an automated outline (tool T3) besides the expert hand tracing (T1 S1) used so far. They were chosen from metadata only (LISTA_164_JUIZES_17NOV.json); when an image has two automated outlines, one was drawn with seed 20261117. The automated outlines were opened only to check their md5 and size; no Dice against them has been computed. On these same images, the rule (package v2, unchanged) and the U-Net (outline stored on 5 November) will be scored against two judges: the hand tracing (the doctor) and the automated outline (the machine). The main count is the same as in the run sealed in version 4, done with each judge: rule minus U-Net, without the frame images (149 images), 95% interval by bootstrap (2,000 resamples, seed 20261200), read with the margin of 0.023; an error is a Dice below 0.05. With the doctor judge the stored Dice values are used. With the machine judge there is no stored value, so the Dice comes from the stored outlines, the same way for both methods, and the doctor judge computed the same way is shown beside. Also beside: how much the U-Net’s advantage changes when the judge changes, image by image. The question: does the reading change when the judge changes? Written before the run, and stated plainly: this test is not conclusive. The sample is small, and the bank is limited: the images with an unused expert hand tracing ran out, because they were used in the earlier comparisons. Whatever comes out enters only as a possible clue, pointing to a hypothesis still to be proved, never as a conclusion. Already known before the seal, from stored values: with the doctor judge on these 149 images, rule minus U-Net is -0.0048 (95% CI -0.0398 to +0.0277), which reads as inconclusive. In 133 of the 164 images the automated outline was accepted by a novice reviewer (S2), not by an expert as in lot 1 (99%); lines for expert-accepted only (29 images without frame) and novice-accepted only (120) are shown beside. The rule runs only to save its outline, and its run must reproduce the per-image records of 15 November in all 164 images, or no count is made. A check of the same path on 16 other images of the 2,197 passed (identical records). Deposited here: the criterion, the predictions of the three of us, the image list, the check result, the seal (LACRE_JUIZES_17NOV.txt, SHA-256 311ae71ae770b0a9ba83f63726d87c407ee6208ebd8d287b2d5daeaa21e55c5c) and the list of hashes. The scripts are not deposited; their hashes are in the seal. Nothing registered in versions 1 to 5 was changed; their files are kept here. Added in version 6: CRITERIO_TESTE_DOS_DOIS_JUIZES_17NOV.txt; PREVISOES_FABLE_E_FABIO_JUIZES_17NOV.txt; PREVISOES_OTSU_JUIZES_17NOV.txt; LISTA_164_JUIZES_17NOV.json; RESULTADO_TESTE_JUIZES_17NOV.txt; LACRE_JUIZES_17NOV.txt; REGISTRO_HASHES_JUIZES_17NOV.txt. New version (version 5): result of the descriptive run sealed in version 4. Version 4 was published on 17 November 2026 by 08:49:53 Brasilia time (11:49:53 UTC, the “modified” field of this record in the Zenodo API); the run wrote its first file at 08:53:03 and the result at 10:28. The seal, the three scripts and the result file were checked against their hashes. This is a description, with no registered verdict: the images had been seen and both banks were used to develop the rule, so the numbers are inflated, those of the 2,197 most of all. Against the expert hand tracing, without the frame images: on the 2,000 images of the first comparison (n = 1,839), the rule (package v2) had a mean Dice of 0.7740 and the U-Net 0.7548 (stored value), rule minus U-Net +0.0192 (95% CI +0.0085 to +0.0297); on the 2,197 (n = 2,017), 0.7739 and 0.7552, rule minus U-Net +0.0187 (95% CI +0.0082 to +0.0291). The two banks share no image. Read with the margin of 0.023 used in lot 1, both come out “superior”. With the frame images included, both come out “non-inferior” (rule minus U-Net -0.0039 and -0.0034); on the frame images alone the rule is far behind (0.56 against 0.83 on the 2,000; 0.56 against 0.81 on the 2,197). Images with a Dice below 0.05: rule 61, U-Net 76 on the 2,000; rule 70, U-Net 89 on the 2,197. By lesion size, the rule is ahead on small lesions and behind on medium and large ones. Lot 1 (reference drawn by a machine in 99% of images), sealed result shown only for comparison: rule 0.8157, U-Net 0.8619, rule minus U-Net -0.0462. The score of the three sets of predictions is in PLACAR_PREVISOES_MAOS_17NOV.txt: the Fable’s predictions for the 2,000 held; two of the Otsu’s items were refuted (the rule came out ahead by more than 0.017, and the swap T changed the mean Dice by 0.0028); Fabio predicted the direction and overestimated the values. The score sheet also declares an error in the Ot

Zenodo (CERN European Organization for Nuclear Research)
Cutaneous Melanoma Detection and Management
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon