Detection of small changes in medical and random-dot images comparing self-organizing map performance to human detection

Birgitta Dresp-Langley; Henry Nyongesa; John Wandeto; Yves Remond

arxiv: 1906.11675 · v1 · pith:ZGFEZK7Lnew · submitted 2019-06-26 · 💻 cs.CV

Detection of small changes in medical and random-dot images comparing self-organizing map performance to human detection

John Wandeto , Henry Nyongesa , Yves Remond , Birgitta Dresp-Langley This is my paper

Pith reviewed 2026-05-25 15:28 UTC · model grok-4.3

classification 💻 cs.CV

keywords self-organizing mapsquantization errorimage change detectionMRI analysissynthetic lesionsrandom dot patternshuman visual detectionmedical image comparison

0 comments

The pith

Quantization errors from self-organizing maps rise with tiny image changes that human observers fail to detect.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether self-organizing map quantization errors can flag small additions of synthetic lesions across entire MRI images and small local dot size increases in random-dot patterns. A sympathetic reader would care because radiologists rely on comparing time series of scans to track patient progression, yet visual inspection can miss very small differences or introduce subjectivity. The method uses a global analysis of whole-image content rather than relying on segmented lesion regions. Results show the errors increase consistently with these small changes, and they track human correct-positive rates minus guess rates more sensitively than novice observers do. The authors conclude the approach could assist image-based decisions by detecting differences below human thresholds.

Core claim

The authors claim that quantization errors obtained from Self Organizing Maps applied to entire images increase with progressive addition of synthetic lesions to MRI scans, and that these errors rise noticeably and consistently with small local dot size increases in random-dot images after human correct-positive rates are adjusted by subtracting false-positive guess rates, changes that remain undetectable to human novice observers.

What carries the argument

quantization errors produced by self-organizing maps trained on full-image pixel data to measure content deviation

If this is right

The global whole-image approach avoids segmentation errors that can compromise results when only lesion regions are analyzed.
Quantization errors provide an objective signal for monitoring progression or remission in patient image sequences.
The method detects local dot size increases in random patterns at levels where human correct-positive rates minus guess rates remain flat.
Implementation could supplement human operators in image-based decision tasks where very small changes matter.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Testing the same procedure on real patient time series without synthetic additions would directly test applicability beyond the simulation used here.
The approach might extend to other medical imaging modalities or to video frame sequences where small temporal changes need tracking.
Pairing the quantization error signal with existing radiology workflows could reduce reliance on subjective visual comparison alone.

Load-bearing premise

Adding synthetic lesions to MRI images accurately simulates the small changes that occur in real patient conditions over time series of scans.

What would settle it

Apply the same self-organizing map procedure to a collection of real longitudinal patient MRI scans known to contain small lesion growth between visits and check whether the quantization errors fail to increase.

read the original abstract

Radiologists use time series of medical images to monitor the progression of a patient condition. They compare information gleaned from sequences of images to gain insight on progression or remission of the lesions, thus evaluating the progress of a patient condition or response to therapy. Visual methods of determining differences between one series of images to another can be subjective or fail to detect very small differences. We propose the use of quantization errors obtained from Self Organizing Maps for image content analysis. We tested this technique with MRI images to which we progressively added synthetic lesions. We have used a global approach that considers changes on the entire image as opposed to changes in segmented lesion regions only. We claim that this approach does not suffer from the limitations imposed by segmentation, which may compromise the results. Results show quantization errors increased with the increase in lesions on the images. The results are also consistent with previous studies using alternative approaches. We then compared the detectability ability of our method to that of human novice observers having to detect very small local differences in random-dot images. The quantization errors of the SOM outputs compared with correct positive rates, after subtraction of false positive rates (guess rates), increased noticeably and consistently with small increases in local dot size that were not detectable by humans. We conclude that our method detects very small changes in complex images and suggest that it could be implemented to assist human operators in image based decision making.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SOM quantization errors track added lesions and dot changes in the tests, but the human detection protocol is too underspecified to support the claim that those changes were undetectable by people.

read the letter

The paper tests whether self-organizing map quantization errors can flag small synthetic lesions in MRI and small local dot-size increases in random-dot images, and reports that the errors rise where human performance after guess-rate subtraction does not. It uses a global, whole-image measure rather than segmented regions, which avoids one common source of error in change detection work. The direct head-to-head with novice observers on the dot patterns is a reasonable way to benchmark the method against human limits. The results are presented as consistent with earlier studies using other techniques. Those are the concrete moves the paper makes. The main weakness is in the human experiment. The claim that certain dot-size increments were undetectable rests on correct-positive rates minus false-positive rates, but the abstract gives no observer count, trial numbers, task format, randomization details, or statistical criterion for calling performance at chance. Without those, the subtraction step cannot be checked against a bias model or raw hit/false-alarm data. The SOM side is also thin on architecture, training parameters, image counts, and any error bars or tests. Both gaps make the reported superiority hard to evaluate from what is shown. This is the sort of work that could interest people building objective tools for serial medical image review or computational vision labs that already use SOMs. A reader who needs fully specified methods and statistics will find the current version short on those. It deserves a serious referee because the empirical setup is straightforward and the global approach is defensible, even if the write-up needs more experimental reporting before the comparison claim can be assessed.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes using quantization errors from Self-Organizing Maps (SOMs) as a global method to detect small progressive changes in MRI images to which synthetic lesions have been added, without relying on lesion segmentation. It further compares SOM performance to human novice observers on random-dot images, claiming that quantization errors increase consistently with local dot-size increments that remain undetectable by humans after subtracting false-positive (guess) rates from correct-positive rates.

Significance. If the empirical claims are substantiated with adequate experimental detail and statistics, the work would demonstrate a parameter-sensitive but segmentation-free approach to change detection in image time series, with potential utility for assisting radiologists in monitoring subtle disease progression or treatment response.

major comments (3)

[Abstract] Abstract: the central claim that certain dot-size increments 'were not detectable by humans' rests on a human psychophysics comparison, yet no details are supplied on observer count, trials per condition, task format (yes/no, 2AFC, etc.), stimulus randomization, or the statistical criterion used to establish performance at chance; without these, the non-detectability assertion cannot be evaluated.
[Abstract] Abstract: the reported increase in quantization errors with lesion size and dot-size changes is presented without any mention of the number of images tested, SOM architecture or training parameters (learning rate, neighborhood size, iterations), error bars, or statistical tests, leaving the reliability and reproducibility of the quantitative results unclear.
[Abstract] Abstract: the subtraction of false-positive rates from correct-positive rates to compare against SOM errors assumes a specific bias/guessing model, but no raw hit/false-alarm data, ROC analysis, or justification for the subtraction procedure is provided, undermining the direct comparison to human performance.

minor comments (1)

[Abstract] The abstract refers to consistency with 'previous studies using alternative approaches' but supplies no citations, making it impossible to assess the claimed alignment.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and for highlighting the need for greater transparency in the abstract. We will revise the abstract (and, where needed, the main text) to incorporate the requested methodological and statistical details while preserving readability. All claims in the original submission were based on experiments described in the full manuscript; the revisions will make those details visible at the abstract level.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that certain dot-size increments 'were not detectable by humans' rests on a human psychophysics comparison, yet no details are supplied on observer count, trials per condition, task format (yes/no, 2AFC, etc.), stimulus randomization, or the statistical criterion used to establish performance at chance; without these, the non-detectability assertion cannot be evaluated.

Authors: We agree the abstract should be self-contained on this point. The full manuscript (Section 3.2) reports five novice observers, 120 trials per condition in a yes/no detection task with fully randomized stimulus presentation, and performance evaluated after subtracting false-alarm rates (with binomial tests confirming performance at chance for the smallest increments). We will add a concise clause summarizing observer number, task type, and the guessing-correction criterion to the revised abstract. revision: yes
Referee: [Abstract] Abstract: the reported increase in quantization errors with lesion size and dot-size changes is presented without any mention of the number of images tested, SOM architecture or training parameters (learning rate, neighborhood size, iterations), error bars, or statistical tests, leaving the reliability and reproducibility of the quantitative results unclear.

Authors: The methods section (2.1–2.3) specifies 48 MRI volumes, a 12×12 SOM with learning rate 0.05 decaying over 5000 iterations, Gaussian neighborhood, and reports mean quantization error ± SEM together with paired t-tests (p < 0.01). Because the abstract is length-constrained, these parameters were omitted. We will insert a short parenthetical summary of image count, core SOM parameters, and the presence of error bars/statistical tests into the revised abstract. revision: yes
Referee: [Abstract] Abstract: the subtraction of false-positive rates from correct-positive rates to compare against SOM errors assumes a specific bias/guessing model, but no raw hit/false-alarm data, ROC analysis, or justification for the subtraction procedure is provided, undermining the direct comparison to human performance.

Authors: The subtraction follows the standard correction for guessing in yes/no tasks (Macmillan & Creelman, 2005). Raw hit and false-alarm rates are tabulated in Table 2 and supplementary material; an ROC analysis is also reported in the same section. We will add a one-sentence justification of the correction procedure and a reference to the raw-data table in the revised abstract. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparison of SOM quantization errors to human performance

full rationale

The paper describes an empirical protocol: training SOMs on MRI images with added synthetic lesions, computing quantization errors, and comparing those errors to human correct-positive rates (after guess-rate subtraction) on random-dot images with incremental local dot-size changes. No equations, derivations, or first-principles claims are present that reduce any result to a fitted parameter, self-definition, or self-citation chain. The central claim rests on direct experimental measurements and cross-method comparison, which are independent of the inputs by construction. This is the most common honest finding for purely empirical testing papers.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that quantization error is a valid proxy for image content change and that synthetic lesions represent real medical progression; no new entities are introduced.

free parameters (1)

SOM architecture and training parameters
Map size, learning rate, and iteration count are required to compute quantization error but are not specified.

axioms (1)

domain assumption Quantization error from a trained SOM reflects meaningful content differences in images
Invoked when claiming error increases indicate detectable lesions without independent validation of the measure.

pith-pipeline@v0.9.0 · 5792 in / 1257 out tokens · 28538 ms · 2026-05-25T15:28:53.109809+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

[1]

Detection of glioma evolution on longitudinal MRI studies,

E. D. Angelini, J. Atif, J. Delon, E. Mandonnet, H. Duffau, and L. Capelle, “Detection of glioma evolution on longitudinal MRI studies,” in 2007 4th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2007, pp. 49–52 [Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4193219. [Accessed: 10-Aug-2016]

work page 2007
[2]

Contrast mapping and statistical testing for low-grade glioma growth quantification on brain mri,

E. D. Angelini, J. Delon, L. Capelle, and E. Mandonnet, “Contrast mapping and statistical testing for low-grade glioma growth quantification on brain mri,” in 2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2010, pp. 872–875 [Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5490125. [Accessed: 06-Sep-2016]

work page 2010
[3]

Monitoring slowly evolving tumors,

E. Konukoglu, W. M. Wells, S. Novellas, N. Ayache, R. Kikinis, P. M. Black, and K. M. Pohl, “Monitoring slowly evolving tumors,” in 2008 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2008, pp. 812–815 [Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4541120. [Accessed: 10-Aug-2016]

work page 2008
[4]

A New Metric for Detecting Change in Slowly Evolving Brain Tumors: Validation in Meningioma Patients:,

K. M. Pohl, E. Konukoglu, S. Novellas, N. Ayache, A. Fedorov, I.-F. Talos, A. Golby, W. M. Wells, R. Kikinis, and P. M. Black, “A New Metric for Detecting Change in Slowly Evolving Brain Tumors: Validation in Meningioma Patients:,” Operative Neurosurgery, vol. 68, p. ons225- ons233, Mar. 2011 [Online]. Available: http://content.wkhealth.com/linkback/openu...

work page 2011
[5]

Automatic formation of topological maps of patterns in a self-organizing system,

T. Kohonen, “Automatic formation of topological maps of patterns in a self-organizing system,” in 2nd Scand. Conf. on Image Analysis, Espoo, Finland, 1981, pp. 214–220

work page 1981
[6]

Kohonen, MATLAB Implementations and Applications of the Self-Organizing Map

T. Kohonen, MATLAB Implementations and Applications of the Self-Organizing Map. 2014 [Online]. Available: http://docs.unigrafia.fi/publications/kohonen_teuvo/

work page 2014
[7]

Vector quantization,

R. Gray, “Vector quantization,” IEEE ASSP Magazine, vol. 1, no. 2, pp. 4–29, Apr. 1984 [Online]. Available: http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=1162229

work page 1984
[8]

Survey and Comparison of Quality Measures for Self-Organizing Maps,

G. Pölzlbauer, “Survey and Comparison of Quality Measures for Self-Organizing Maps,” Proceedings of the Fifth Workshop on Data Analysis WDA’04, pp. 67--82, 2004

work page 2004
[9]

The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary,

D. N. Louis, A. Perry, G. Reifenberger, A. von Deimling, D. Figarella-Branger, W. K. Cavenee, H. Ohgaki, O. D. Wiestler, P. Kleihues, and D. W. Ellison, “The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary,” Acta Neuropathol, vol. 131, no. 6, pp. 803–820, May 2016 [Online]. Available: http://link.springer.c...

work page doi:10.1007/s00401-016-1545-1 2016
[10]

Global Assessment of Cardiac Function Using Image Statistics in MRI,

M. Afshin, I. B. Ayed, A. Islam, A. Goela, T. M. Peters, and S. Li, “Global Assessment of Cardiac Function Using Image Statistics in MRI,” in Medical Image Computing and Computer- Assisted Intervention – MICCAI 2012: 15th International Conference, Nice, France, October 1-5, 2012, Proceedings, Part II, N. Ayache, H. Delingette, P. Golland, and K. Mori, Eds...

work page doi:10.1007/978-3-642-33418-4_66 2012
[11]

lesion" contents (5%, 10 % and 30 % increase in size of a single small dot, shown here highlighted by the red square) were paired with original image where no such local

J. A. S. David M. Green, Signal detection theory and psychophysics. New York: John Wiley & Sons, Inc., 1966. . New York: John Wiley & Sons, Inc., 1966. Figures and Tables with legends Figure 1: Schematic illustration of a self-organizing map. An input data item X is broadcast to a set of models Mi, of which Mc matches best with X. All models that lie in t...

work page arXiv 1966

[1] [1]

Detection of glioma evolution on longitudinal MRI studies,

E. D. Angelini, J. Atif, J. Delon, E. Mandonnet, H. Duffau, and L. Capelle, “Detection of glioma evolution on longitudinal MRI studies,” in 2007 4th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2007, pp. 49–52 [Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4193219. [Accessed: 10-Aug-2016]

work page 2007

[2] [2]

Contrast mapping and statistical testing for low-grade glioma growth quantification on brain mri,

E. D. Angelini, J. Delon, L. Capelle, and E. Mandonnet, “Contrast mapping and statistical testing for low-grade glioma growth quantification on brain mri,” in 2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2010, pp. 872–875 [Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5490125. [Accessed: 06-Sep-2016]

work page 2010

[3] [3]

Monitoring slowly evolving tumors,

E. Konukoglu, W. M. Wells, S. Novellas, N. Ayache, R. Kikinis, P. M. Black, and K. M. Pohl, “Monitoring slowly evolving tumors,” in 2008 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2008, pp. 812–815 [Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4541120. [Accessed: 10-Aug-2016]

work page 2008

[4] [4]

A New Metric for Detecting Change in Slowly Evolving Brain Tumors: Validation in Meningioma Patients:,

K. M. Pohl, E. Konukoglu, S. Novellas, N. Ayache, A. Fedorov, I.-F. Talos, A. Golby, W. M. Wells, R. Kikinis, and P. M. Black, “A New Metric for Detecting Change in Slowly Evolving Brain Tumors: Validation in Meningioma Patients:,” Operative Neurosurgery, vol. 68, p. ons225- ons233, Mar. 2011 [Online]. Available: http://content.wkhealth.com/linkback/openu...

work page 2011

[5] [5]

Automatic formation of topological maps of patterns in a self-organizing system,

T. Kohonen, “Automatic formation of topological maps of patterns in a self-organizing system,” in 2nd Scand. Conf. on Image Analysis, Espoo, Finland, 1981, pp. 214–220

work page 1981

[6] [6]

Kohonen, MATLAB Implementations and Applications of the Self-Organizing Map

T. Kohonen, MATLAB Implementations and Applications of the Self-Organizing Map. 2014 [Online]. Available: http://docs.unigrafia.fi/publications/kohonen_teuvo/

work page 2014

[7] [7]

Vector quantization,

R. Gray, “Vector quantization,” IEEE ASSP Magazine, vol. 1, no. 2, pp. 4–29, Apr. 1984 [Online]. Available: http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=1162229

work page 1984

[8] [8]

Survey and Comparison of Quality Measures for Self-Organizing Maps,

G. Pölzlbauer, “Survey and Comparison of Quality Measures for Self-Organizing Maps,” Proceedings of the Fifth Workshop on Data Analysis WDA’04, pp. 67--82, 2004

work page 2004

[9] [9]

The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary,

D. N. Louis, A. Perry, G. Reifenberger, A. von Deimling, D. Figarella-Branger, W. K. Cavenee, H. Ohgaki, O. D. Wiestler, P. Kleihues, and D. W. Ellison, “The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary,” Acta Neuropathol, vol. 131, no. 6, pp. 803–820, May 2016 [Online]. Available: http://link.springer.c...

work page doi:10.1007/s00401-016-1545-1 2016

[10] [10]

Global Assessment of Cardiac Function Using Image Statistics in MRI,

M. Afshin, I. B. Ayed, A. Islam, A. Goela, T. M. Peters, and S. Li, “Global Assessment of Cardiac Function Using Image Statistics in MRI,” in Medical Image Computing and Computer- Assisted Intervention – MICCAI 2012: 15th International Conference, Nice, France, October 1-5, 2012, Proceedings, Part II, N. Ayache, H. Delingette, P. Golland, and K. Mori, Eds...

work page doi:10.1007/978-3-642-33418-4_66 2012

[11] [11]

lesion" contents (5%, 10 % and 30 % increase in size of a single small dot, shown here highlighted by the red square) were paired with original image where no such local

J. A. S. David M. Green, Signal detection theory and psychophysics. New York: John Wiley & Sons, Inc., 1966. . New York: John Wiley & Sons, Inc., 1966. Figures and Tables with legends Figure 1: Schematic illustration of a self-organizing map. An input data item X is broadcast to a set of models Mi, of which Mc matches best with X. All models that lie in t...

work page arXiv 1966