Spectral decoupling for training transferable neural networks in medical imaging
Summary: Many neural networks for medical imaging generalize poorly to data unseen during training. Such behavior can be caused by overfitting easy-to-learn features while disregarding other potentially informative features. A recent implicit bias mitigation technique called spectral decoupling prov...
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2022-02-01
|
| Series: | iScience |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2589004222000372 |
| _version_ | 1819289150244782080 |
|---|---|
| author | Joona Pohjonen Carolin Stürenberg Antti Rannikko Tuomas Mirtti Esa Pitkänen |
| author_facet | Joona Pohjonen Carolin Stürenberg Antti Rannikko Tuomas Mirtti Esa Pitkänen |
| author_sort | Joona Pohjonen |
| collection | DOAJ |
| description | Summary: Many neural networks for medical imaging generalize poorly to data unseen during training. Such behavior can be caused by overfitting easy-to-learn features while disregarding other potentially informative features. A recent implicit bias mitigation technique called spectral decoupling provably encourages neural networks to learn more features by regularizing the networks' unnormalized prediction scores with an L2 penalty. We show that spectral decoupling increases the networks′ robustness for data distribution shifts and prevents overfitting on easy-to-learn features in medical images. To validate our findings, we train networks with and without spectral decoupling to detect prostate cancer on tissue slides and COVID-19 in chest radiographs. Networks trained with spectral decoupling achieve up to 9.5 percent point higher performance on external datasets. Spectral decoupling alleviates generalization issues associated with neural networks and can be used to complement or replace computationally expensive explicit bias mitigation methods, such as stain normalization in histological images. |
| first_indexed | 2024-12-24T03:02:16Z |
| format | Article |
| id | doaj.art-af6d19ffc4ff43f0bb7df1dc13d598ae |
| institution | Directory Open Access Journal |
| issn | 2589-0042 |
| language | English |
| last_indexed | 2024-12-24T03:02:16Z |
| publishDate | 2022-02-01 |
| publisher | Elsevier |
| record_format | Article |
| series | iScience |
| spelling | doaj.art-af6d19ffc4ff43f0bb7df1dc13d598ae2022-12-21T17:18:09ZengElsevieriScience2589-00422022-02-01252103767Spectral decoupling for training transferable neural networks in medical imagingJoona Pohjonen0Carolin Stürenberg1Antti Rannikko2Tuomas Mirtti3Esa Pitkänen4Research Program in Systems Oncology, University of Helsinki, Helsinki, Finland; Corresponding authorResearch Program in Systems Oncology, University of Helsinki, Helsinki, FinlandResearch Program in Systems Oncology, University of Helsinki, Helsinki, Finland; Department of Urology, Helsinki University Hospital, Helsinki, FinlandResearch Program in Systems Oncology, University of Helsinki, Helsinki, Finland; Department of Pathology, Helsinki University Hospital, Helsinki, FinlandInstitute for Molecular Medicine Finland (FIMM), HiLIFE, University of Helsinki, Helsinki, Finland; Research Program in Applied Tumor Genomics, University of Helsinki, Helsinki, Finland; Corresponding authorSummary: Many neural networks for medical imaging generalize poorly to data unseen during training. Such behavior can be caused by overfitting easy-to-learn features while disregarding other potentially informative features. A recent implicit bias mitigation technique called spectral decoupling provably encourages neural networks to learn more features by regularizing the networks' unnormalized prediction scores with an L2 penalty. We show that spectral decoupling increases the networks′ robustness for data distribution shifts and prevents overfitting on easy-to-learn features in medical images. To validate our findings, we train networks with and without spectral decoupling to detect prostate cancer on tissue slides and COVID-19 in chest radiographs. Networks trained with spectral decoupling achieve up to 9.5 percent point higher performance on external datasets. Spectral decoupling alleviates generalization issues associated with neural networks and can be used to complement or replace computationally expensive explicit bias mitigation methods, such as stain normalization in histological images.http://www.sciencedirect.com/science/article/pii/S2589004222000372Medical testsMedical imagingAlgorithmsArtificial intelligence |
| spellingShingle | Joona Pohjonen Carolin Stürenberg Antti Rannikko Tuomas Mirtti Esa Pitkänen Spectral decoupling for training transferable neural networks in medical imaging iScience Medical tests Medical imaging Algorithms Artificial intelligence |
| title | Spectral decoupling for training transferable neural networks in medical imaging |
| title_full | Spectral decoupling for training transferable neural networks in medical imaging |
| title_fullStr | Spectral decoupling for training transferable neural networks in medical imaging |
| title_full_unstemmed | Spectral decoupling for training transferable neural networks in medical imaging |
| title_short | Spectral decoupling for training transferable neural networks in medical imaging |
| title_sort | spectral decoupling for training transferable neural networks in medical imaging |
| topic | Medical tests Medical imaging Algorithms Artificial intelligence |
| url | http://www.sciencedirect.com/science/article/pii/S2589004222000372 |
| work_keys_str_mv | AT joonapohjonen spectraldecouplingfortrainingtransferableneuralnetworksinmedicalimaging AT carolinsturenberg spectraldecouplingfortrainingtransferableneuralnetworksinmedicalimaging AT anttirannikko spectraldecouplingfortrainingtransferableneuralnetworksinmedicalimaging AT tuomasmirtti spectraldecouplingfortrainingtransferableneuralnetworksinmedicalimaging AT esapitkanen spectraldecouplingfortrainingtransferableneuralnetworksinmedicalimaging |