CoChat
Sign Up
Cited by 107Domain Adaptation and Few-Shot LearningData Stream Mining TechniquesMedical Image Segmentation Techniques

The DOI matches its Crossref record; 59 of the 61 references it cites were found in a registry or an index.

Dynamic memory to alleviate catastrophic forgetting in continual learning with medical imaging

Open access
Perkonigg, Matthias, Hofmanninger, Johannes, Herold, Christian J., Brink, James A., Pianykh, Oleg, Prosch, Helmut, Langs, Georg·2021·Nature Communications
DOI 10.1038/s41467-021-25858-zPMC PMC8479083
Open this paper in CoChat — free accountOpen PDF

Abstract

Medical imaging is a central part of clinical diagnosis and treatment guidance. Machine learning has increasingly gained relevance because it captures features of disease and treatment response that are relevant for therapeutic decision-making. In clinical practice, the continuous progress of image acquisition technology or diagnostic procedures, the diversity of scanners, and evolving imaging protocols hamper the utility of machine learning, as prediction accuracy on new data deteriorates, or models become outdated due to these domain shifts. We propose a continual learning approach to deal with such domain shifts occurring at unknown time points. We adapt models to emerging variations in a continuous data stream while counteracting catastrophic forgetting. A dynamic memory enables rehearsal on a subset of diverse training data to mitigate forgetting while enabling models to expand to new domains. The technique balances memory by detecting pseudo-domains, representing different style clusters within the data stream. Evaluation of two different tasks, cardiac segmentation in magnetic resonance imaging and lung nodule detection in computed tomography, demonstrate a consistent advantage of the method.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-021-25858-z

Cites

59 of 59 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Different model training and adaptation strategies. 1 Continual DL for image analysis adapts to new data properties and, at the same time, retains the capability to work with older data. a Static DL: after training and deploying a DL model, technology changes and the accuracy of the model decreases. b A naive continual learning approach to solve this limitation is to continue training a model on a stream of data. However, this leads to forgetting of old data properties, and a corresponding decrease in performance on these data. c As a continual learning approach, dynamic memory recognizes new emerging domains, and samples data in a continuous stream accordingly. The ML model adapts to new technology but stays accurate on the diverse set of scanners previously seen. This is important to ensure backward compatibility of the model and to build a more stable model that adapts to new scanners faster. 2 Validation on a separate test set evaluates the DL model performance at the end of the period.
Fig. 1 Different model training and adaptation strategies. 1 Continual DL for image analysis adapts to new data properties and, at the same time, retains the capability to work with older data. a Static DL: after training and deploying a DL model, technology changes and the accuracy of the model decreases. b A naive continual learning approach to solve this limitation is to continue training a model on a stream of data. However, this leads to forgetting of old data properties, and a corresponding decrease in performance on these data. c As a continual learning approach, dynamic memory recognizes new emerging domains, and samples data in a continuous stream accordingly. The ML model adapts to new technology but stays accurate on the diverse set of scanners previously seen. This is important to ensure backward compatibility of the model and to build a more stable model that adapts to new scanners faster. 2 Validation on a separate test set evaluates the DL model performance at the end of the period.CC BY · PubMed Central
Fig. 2 Cardiac MR segmentation. Dice score (DSC) on the validation set during training for M = 128 for DM and DM-PD, compared to random replacement and naive continual learning. The timeline at the bottom represents a continuous data stream and the change of domains in the stream. a A drop of DSC for Scanner A can be observed when Scanner B occurs in the stream, DM and DM-PD were able to recover from this drop using the memory; b all methods are stable during training with Scanner C, and c as soon as Scanner D data flows in, we see a quick rise of DSC for Scanner D validation samples. Naive and random replacement lose some DSC points during that period, while DM remains stable. This significant and rapid change shows that Scanner D is different compared to the others and not as closely related as Scanner B and C.
Fig. 2 Cardiac MR segmentation. Dice score (DSC) on the validation set during training for M = 128 for DM and DM-PD, compared to random replacement and naive continual learning. The timeline at the bottom represents a continuous data stream and the change of domains in the stream. a A drop of DSC for Scanner A can be observed when Scanner B occurs in the stream, DM and DM-PD were able to recover from this drop using the memory; b all methods are stable during training with Scanner C, and c as soon as Scanner D data flows in, we see a quick rise of DSC for Scanner D validation samples. Naive and random replacement lose some DSC points during that period, while DM remains stable. This significant and rapid change shows that Scanner D is different compared to the others and not as closely related as Scanner B and C.CC BY · PubMed Central
Fig. 3 Qualitative examples for cardiac segmentation. Results for DM and DM-PD with M = 128 compared to naive continual learning and random memory replacement. Mislabelled refers to pixels that were labelled, but the class membership was confused by the model. The base model was trained in a static training approach on data from Scanner A only.
Fig. 3 Qualitative examples for cardiac segmentation. Results for DM and DM-PD with M = 128 compared to naive continual learning and random memory replacement. Mislabelled refers to pixels that were labelled, but the class membership was confused by the model. The base model was trained in a static training approach on data from Scanner A only.CC BY · PubMed Central
Fig. 4 Lung nodule detection. Average precision (AP) measured on the validation set during training for DM and DM-PD with M = 128 as well as a random replacement memory and naive continual learning approach. The timeline at the bottom represents the changes of domains in the data stream. a As soon as Scanner F data occurred in the stream, the validation performed on the Scanner F (and also Scanner G) domain increased for all approaches. b A clear drop in performance (AP) occurred for the naive and random replacement approach for Scanner E, F and G, after some steps of training on Scanner H data, which marked catastrophic forgetting. For DM, the performance first dropped slightly, but recovered after some training steps, because samples from the memory are used for training. The DM-PD performance remained stable for Scanners E, F and G. c At the end of continual training, a better performance was achieved for Scanner E, F and G, when dynamic memory was utilized. For Scanner H, the performance for all three approaches was similar. d The performance for Scanner E was stable during the entire continual training process for all approaches, showing a base training that was saturated for scanner E.
Fig. 4 Lung nodule detection. Average precision (AP) measured on the validation set during training for DM and DM-PD with M = 128 as well as a random replacement memory and naive continual learning approach. The timeline at the bottom represents the changes of domains in the data stream. a As soon as Scanner F data occurred in the stream, the validation performed on the Scanner F (and also Scanner G) domain increased for all approaches. b A clear drop in performance (AP) occurred for the naive and random replacement approach for Scanner E, F and G, after some steps of training on Scanner H data, which marked catastrophic forgetting. For DM, the performance first dropped slightly, but recovered after some training steps, because samples from the memory are used for training. The DM-PD performance remained stable for Scanners E, F and G. c At the end of continual training, a better performance was achieved for Scanner E, F and G, when dynamic memory was utilized. For Scanner H, the performance for all three approaches was similar. d The performance for Scanner E was stable during the entire continual training process for all approaches, showing a base training that was saturated for scanner E.CC BY · PubMed Central
Fig. 5 Quantitative and qualitative results for lung nodule detection. a Precision–recall curves of the final naive, DM and DM-PD trained model compared to a base model trained on only scanner E data. Shaded areas represent confidence intervals for n = 5 independent training runs. b Samples of lung nodule detection of the final naive, DM and DM-PD trained model compared to a base model trained on only scanner E data on all four domains. Green boxes indicate true-positives, yellow boxes false-negatives and red boxes false-positives.
Fig. 5 Quantitative and qualitative results for lung nodule detection. a Precision–recall curves of the final naive, DM and DM-PD trained model compared to a base model trained on only scanner E data. Shaded areas represent confidence intervals for n = 5 independent training runs. b Samples of lung nodule detection of the final naive, DM and DM-PD trained model compared to a base model trained on only scanner E data on all four domains. Green boxes indicate true-positives, yellow boxes false-negatives and red boxes false-positives.CC BY · PubMed Central
Fig. 6 Enhancing memory diversity. Comparison of memory composition for DM vs. DM-PD for M = 128. a TSNE over the Gram matrices of the training set shows a distinction of the domains. Markers show positions of memory elements at the end of training for one run. DM-PD shows a more equal distribution over the whole training set. b Five pseudo-domains detected were detected for a single training run of DM-PD; bars show the amount of domain elements assigned to the pseudo-domains. c Amount of domain elements in memory after training for DM and DM-PD. Error bars represent standard deviation and mean (middle line), n = 5 independent runs.
Fig. 6 Enhancing memory diversity. Comparison of memory composition for DM vs. DM-PD for M = 128. a TSNE over the Gram matrices of the training set shows a distinction of the domains. Markers show positions of memory elements at the end of training for one run. DM-PD shows a more equal distribution over the whole training set. b Five pseudo-domains detected were detected for a single training run of DM-PD; bars show the amount of domain elements assigned to the pseudo-domains. c Amount of domain elements in memory after training for DM and DM-PD. Error bars represent standard deviation and mean (middle line), n = 5 independent runs.CC BY · PubMed Central
Fig. 7 Updating dynamic memory M of size 8 with a new sample b, taken from the input-mini-batch. (1) The image is given to the style model and the Gram matrices G1, G2 are calculated. (2) Based on those matrices the Gram distance δ between b and each element in the memory is calculated. (3) b replaces the memory element with minimal distance δ.
Fig. 7 Updating dynamic memory M of size 8 with a new sample b, taken from the input-mini-batch. (1) The image is given to the style model and the Gram matrices G1, G2 are calculated. (2) Based on those matrices the Gram distance δ between b and each element in the memory is calculated. (3) b replaces the memory element with minimal distance δ.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 61 references

Research works printed but not linked: 2

  1. Lopez-Paz, D. & Ranzato, M. Gradient episodic memory for continual learning. Adv. Neural Inf. Process. Syst. 30, 6467–6476 (2017).
  2. Glocker, B., Robinson, R., Castro, D. C., Dou, Q. & Konukoglu, E. Machine learning with multi-site imaging data: an empirical study on the impact of scanner effects. Preprint at https://arxiv.org/abs/1910.04597 (2019). · arXiv

Cited by

Sorted by citations

Items 101–107 of 107

Related

See all

Metadata

Authors
Perkonigg, M., Hofmanninger, J., Herold, C. J., Brink, J. A., Pianykh, O., Prosch, H., & Langs, G.
Affiliations
  • Medical University of Vienna (AT)
  • Harvard University (US)
Corresponding author
Langs, Georg
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2021-09-28
Volume
12
Issue
1
Pages
5678-5678
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-021-25858-z
PMC
PMC8479083
OpenAlex
W3202253200
Cited by
106
References
61
Citations by year
2026: 24 · 2025: 28 · 2024: 15 · 2023: 23 · 2022: 14 · 2021: 2
FWCI
4.75
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Boehringer Ingelheim
  • European Commission (780495, 765148)
  • Austrian Science Fund (P34198, P35189)
  • Oesterreichische Nationalbank (18207)
  • Novartis Pharmaceuticals Corporation
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Data Stream Mining Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
Keywords
catastrophic forgetting, continual learning, dynamic memory, domain shift, diagnostic imaging, cardiac image segmentation, lung nodule detection, magnetic resonance imaging, computed tomography, model updating
MeSH terms
Machine Learning, Diagnostic Imaging, Humans, Learning, Lung, Memory, Tomography, X-Ray Computed, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review