Asymmetric Loss Functions and Deep Densely Connected Networks for Highly Imbalanced Medical Image Segmentation: Application to Multiple Sclerosis Lesion Detection.
Asymmetric loss function
Convolutional neural network
Deep learning
FC-DenseNet
Focal loss
Fβ scores
Lesion segmentation
Multiple Sclerosis
Patch prediction fusion
Tversky index
Journal
IEEE access : practical innovations, open solutions
ISSN: 2169-3536
Titre abrégé: IEEE Access
Pays: United States
ID NLM: 101639462
Informations de publication
Date de publication:
2019
2019
Historique:
entrez:
19
9
2019
pubmed:
19
9
2019
medline:
19
9
2019
Statut:
ppublish
Résumé
Fully convolutional deep neural networks have been asserted to be fast and precise frameworks with great potential in image segmentation. One of the major challenges in training such networks raises when data is unbalanced, which is common in many medical imaging applications such as lesion segmentation where lesion class voxels are often much lower in numbers than non-lesion voxels. A trained network with unbalanced data may make predictions with high precision and low recall, being severely biased towards the non-lesion class which is particularly undesired in most medical applications where false negatives are actually more important than false positives. Various methods have been proposed to address this problem including two step training, sample re-weighting, balanced sampling, and more recently similarity loss functions, and focal loss. In this work we trained fully convolutional deep neural networks using an asymmetric similarity loss function to mitigate the issue of data imbalance and achieve much better trade-off between precision and recall. To this end, we developed a 3D fully convolutional densely connected network (FC-DenseNet) with large overlapping image patches as input and an asymmetric similarity loss layer based on Tversky index (using
Identifiants
pubmed: 31528523
doi: 10.1109/ACCESS.2018.2886371
pmc: PMC6746414
mid: NIHMS1518239
doi:
Types de publication
Journal Article
Langues
eng
Pagination
721-1735Subventions
Organisme : NIBIB NIH HHS
ID : R01 EB018988
Pays : United States
Organisme : NINDS NIH HHS
ID : R01 NS079788
Pays : United States
Références
Curr Opin Neurol. 2018 Jun;31(3):233-243
pubmed: 29634596
IEEE Trans Med Imaging. 2019 May;38(5):1116-1126
pubmed: 30387726
IEEE Trans Med Imaging. 2016 May;35(5):1229-1239
pubmed: 26886978
Neuroimage Clin. 2015 May 13;8:376-89
pubmed: 26106563
J Magn Reson Imaging. 2016 Jun;43(6):1445-54
pubmed: 26606758
IEEE Trans Pattern Anal Mach Intell. 2020 Feb;42(2):318-327
pubmed: 30040631
Med Image Anal. 2017 Feb;36:61-78
pubmed: 27865153
IEEE Trans Pattern Anal Mach Intell. 2017 Apr;39(4):640-651
pubmed: 27244717
IEEE Trans Med Imaging. 2015 Jun;34(6):1349-61
pubmed: 25616008
Neuroimage. 2018 Apr 15;170:434-445
pubmed: 28223187
Neuroimage. 2017 Mar 1;148:77-102
pubmed: 28087490
Neuroimage. 2015 Mar;108:214-24
pubmed: 25562829
Clin Med Res. 2003 Jan;1(1):57-60
pubmed: 15931287
Neuroimage. 2017 Jul 15;155:159-168
pubmed: 28435096
Cell. 1996 May 3;85(3):299-302
pubmed: 8616884
Deep Learn Med Image Anal Multimodal Learn Clin Decis Support (2017). 2017;2017:240-248
pubmed: 34104926
Neuroimage. 2018 Apr 15;170:446-455
pubmed: 28445774
Neuroimage. 2011 Jul 15;57(2):378-90
pubmed: 21497655
IEEE Trans Med Imaging. 2017 Nov;36(11):2319-2330
pubmed: 28678704
IEEE Trans Med Imaging. 2016 May;35(5):1252-1261
pubmed: 27046893
Neuroimage. 2016 Apr 1;129:460-469
pubmed: 26808333
IEEE Trans Med Imaging. 2016 May;35(5):1240-1251
pubmed: 26960222
Med Image Anal. 2013 Jan;17(1):1-18
pubmed: 23084503
Med Image Anal. 2017 Jan;35:18-31
pubmed: 27310171