Rethinking the Dice Loss for Deep Learning Lesion Segmentation in Medical Images

doi:10.1007/s12204-021-2264-x

摘要/Abstract

摘要： Deep learning is widely used for lesion segmentation in medical images due to its breakthrough performance. Loss functions are critical in a deep learning pipeline, and they play important roles in segmenting performance. Dice loss is the most commonly used loss function in medical image segmentation, but it also has some disadvantages. In this paper, we discuss the advantages and disadvantages of the Dice loss function, and group the extensions of the Dice loss according to its improved purpose. The performances of some extensions are compared according to core references. Because different loss functions have different performances in different tasks, automatic loss function selection will be the potential direction in the future.

关键词: Dice loss, deep learning, medical image, lesion segmentation

Abstract: Deep learning is widely used for lesion segmentation in medical images due to its breakthrough performance. Loss functions are critical in a deep learning pipeline, and they play important roles in segmenting performance. Dice loss is the most commonly used loss function in medical image segmentation, but it also has some disadvantages. In this paper, we discuss the advantages and disadvantages of the Dice loss function, and group the extensions of the Dice loss according to its improved purpose. The performances of some extensions are compared according to core references. Because different loss functions have different performances in different tasks, automatic loss function selection will be the potential direction in the future.

Key words: Dice loss, deep learning, medical image, lesion segmentation

中图分类号:

ZHANG Yue (张月), LIU Shijie (刘世界), LI Chunlai (李春来), WANG Jianyu (王建宇). Rethinking the Dice Loss for Deep Learning Lesion Segmentation in Medical Images[J]. J Shanghai Jiaotong Univ Sci, 2021, 26(1): 93-102.

参考文献 54

[1]	GARC′IA-LORENZO D, FRANCIS S, NARAYANAN S, et al. Review of automatic segmentation methodsof multiple sclerosis white matter lesions on conventionalmagnetic resonance imaging [J]. Medical ImageAnalysis, 2013, 17(1): 1-18.
[2]	SARITHA S, PRABHA N A. A comprehensive review:Segmentation of MRI images — brain tumor [J]. InternationalJournal of Imaging Systems and Technology,2016, 26(4): 295-304.
[3]	WADHWA A, BHARDWAJ A, VERMA V S. A reviewon brain tumor segmentation of MRI images [J].Magnetic Resonance Imaging, 2019, 61: 247-259.
[4]	WEEDA M M, BROUWER I, DE VOS M L, et al.Comparing lesion segmentation methods in multiplesclerosis: Input from one manually delineated subjectis sufficient for accurate lesion segmentation [J]. NeuroImage:Clinical, 2019, 24: 102074.
[5]	BERNAL J, KUSHIBAR K, ASFAW D S, et al. Deepconvolutional neural networks for brain image analysison magnetic resonance imaging: A review [J]. ArtificialIntelligence in Medicine, 2019, 95: 64-81.
[6]	KERVADEC H, BOUCHTIBA J, DESROSIERS C, etal. Boundary loss for highly unbalanced segmentation[J]. Medical Image Analysis, 2021, 67: 101851.
[7]	CHEN C, QIN C, QIU H Q, et al. Deep learning forcardiac image segmentation: A review [J]. Frontiers inCardiovascular Medicine, 2020, 7: 25.
[8]	IS?IN A, DIREKOˇGLU C, S?AH M. Review of MRIbasedbrain tumor image segmentation using deeplearning methods [J]. Procedia Computer Science,2016, 102: 317-324.
[9]	JAAFRA Y, LAURENT J L, DERUYVER A, et al.Reinforcement learning for neural architecture search:A review [J]. Image and Vision Computing, 2019, 89:57-66.
[10]	MAKROPOULOS A, COUNSELL S J, RUECKERTD. A review on automatic fetal and neonatal brainMRI segmentation [J]. NeuroImage, 2018, 170: 231-248.
[11]	SCHMIDHUBER J. Deep learning in neural networks:An overview [J]. Neural Networks, 2015, 61: 85-117.
[12]	JADON S. A survey of loss functions for semanticsegmentation [EB/OL]. [2020-07-16]. https://arxiv.org/pdf/2006.14822.pdf.
[13]	MA J. Segmentation loss odyssey [EB/OL]. [2020-07-16]. https://arxiv.org/pdf/2005.13449.pdf.
[14]	MILLETARI F, NAVAB N, AHMADI S A. V-net:Fully convolutional neural networks for volumetricmedical image segmentation [C]//2016 Fourth InternationalConference on 3D Vision (3DV). Stanford,California, USA: IEEE, 2016: 565-571.
[15]	DROZDZAL M, VORONTSOV E, CHARTRAND G,et al. The importance of skip connections in biomedicalimage segmentation [M]//CARNEIRO G, MATEUSD, PETER L, et al. Deep learning and data labelingfor medical applications. Cham: Springer, 2016: 179-187.
[16]	FIDON L, LI W Q, GARCIA-PERAZA-HERRERA LC, et al. Generalised wasserstein Dice score for imbalancedmulti-class segmentation using holistic convolutionalnetworks [M]//CRIMI A, BAKAS S, KUIJFB, et al. Brainlesion: Glioma, multiple sclerosis,stroke and traumatic brain injuries. Cham, Switzerland:Springer, 2018: 64-76.
[17]	REZATOFIGHI H, TSOI N, GWAK J Y, et al. Generalizedintersection over union: A metric and a loss forbounding box regression [C]//2019 IEEE/CVF Conferenceon Computer Vision and Pattern Recognition(CVPR). Long Beach, CA, USA: IEEE, 2019: 658-666.
[18]	DOLZ J, DESROSIERS C, AYED I B. 3D fully convolutionalnetworks for subcortical segmentation in MRI:A large-scale study [J]. NeuroImage, 2018, 170: 456-470.
[19]	GUIZARD N, COUP?P, FONOV V S, et al. Rotationinvariantmulti-contrast non-local means for MS lesionsegmentation [J]. NeuroImage: Clinical, 2015, 8: 376-389.
[20]	HARMOUCHE R, SUBBANNA N K, COLLINS D L,et al. Probabilistic multiple sclerosis lesion classificationbased on modeling regional intensity variabilityand local neighborhood information [J]. IEEE Transactionson Biomedical Engineering, 2015, 62(5): 1281-1292.
[21]	STYNER M, LEE J, CHIN B, et al. 3D segmentationin the clinic: A grand challenge II: MS lesion segmentation[J]. MIDAS Journal, 2008, 2008: 1-6.
[22]	WONG K C L, MORADI M, TANG H, et al. 3Dsegmentation with exponential logarithmic loss forhighly unbalanced object sizes [M]//FRANGI A F,SCHNABEL J A, DAVATZIKOS C, et al. Medicalimage computing and computer assisted intervention— MICCAI 2018. Cham, Switzerland: Springer, 2018:612-619.
[23]	LUCAS C, KEMMLING A, MAMLOUK A M, etal. Multi-scale neural network for automatic segmentationof ischemic strokes on acute perfusion images[C]//2018 IEEE 15th International Symposium onBiomedical Imaging (ISBI 2018). Washington, DC,USA: IEEE, 2018: 1118-1121.
[24]	WANG ZW, SMITH C D, LIU J D. Ensemble of multisizedFCNs to improve white matter lesion segmentation[M]//SHI Y H, SUK H I, LIUM X. Machine learningin medical imaging. Cham, Switzerland: Springer,2018: 223-232.
[25]	KARIMI D, SALCUDEAN S E. Reducing the hausdorffdistance in medical image segmentation with convolutionalneural networks [J]. IEEE Transactions onMedical Imaging, 2020, 39(2): 499-513.
[26]	YANG D, ROTH H, WANG X S, et al. Enhancingforeground boundaries for medicalimage segmentation [EB/OL]. [2020-07-16].https://arxiv.org/pdf/2005.14355.pdf.
[27]	ODA H, ROTH H R, CHIBA K, et al. BESNet:Boundary-enhanced segmentation of cells inhistopathological images [M]//FRANGI A F,SCHNABEL J A, DAVATZIKOS C, et al. Medicalimage computing and computer assisted intervention— MICCAI 2018. Cham, Switzerland: Springer, 2018:228-236.
[28]	SUDRE C H, LIWQ, VERCAUTEREN T, et al. GeneralisedDice overlap as a deep learning loss functionfor highly unbalanced segmentations [M]//CARDOSOM J, ARBEL T, CARNEIRO G, et al. Deep Learningin medical image analysis and multimodal learningfor clinical decision support. Cham, Switzerland:Springer, 2017: 240-248.
[29]	TAGHANAKI S A, ZHENG Y F, ZHOU S K, et al.Combo loss: Handling input and output imbalancein multi-organ segmentation [J]. Computerized MedicalImaging and Graphics, 2019, 75: 24-33.
[30]	SALEHI S S M, ERDOGMUS D, GHOLIPOUR A.Tversky loss function for image segmentation using 3Dfully convolutional deep networks [M]//WANG Q, SHIY H, SUK H I, et al. Machine Learning in MedicalImaging. Cham, Switzerland: Springer, 2017: 379-387.
[31]	RONNEBERGER O, FISCHER P, BROX T. U-net:Convolutional networks for biomedical image segmentation[C]//NAVAB N, HORNEGGER J, WELLS WM, et al. Medical image computing and computerassistedintervention—MICCAI 2015. Cham, Switzerland:Springer, 2015: 234-241.
[32]	WU Z F, SHEN C H, VAN DEN HENGELA. Bridging category-level and instance-level semanticimage segmentation [EB/OL]. [2020-07-16].https://arxiv.org/pdf/1605.06885.pdf.
[33]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focalloss for dense object detection [C]//2017 IEEE InternationalConference on Computer Vision (ICCV ).Venice, Italy: IEEE, 2017: 2980-2988.
[34]	WANG P, CHUNG A C S. Focal Dice lossand image dilation for brain tumor segmentation[M]//STOYANOV D, TAYLOR Z, CARNEIRO G, etal. Deep learning in medical image analysis and multimodallearning for clinical decision support. Cham,Switzerland: Springer, 2018: 119-127.
[35]	ZHOU Y J, HUANG W J, DONG P, et al. D-UNet:A dimension-fusion U shape network for chronicstroke lesion segmentation [J]. IEEE/ACM Transactionson Computational Biology and Bioinformatics,2019. https://doi.org/10.1109/TCBB.2019.2939522(published online).
[36]	ASLANI S, MURINO V, DAYAN M, et al. Scannerinvariant multiple sclerosis lesion segmentation fromMRI [C]//2020 IEEE 17th International Symposiumon Biomedical Imaging (ISBI). Lowa City, IA, USA:IEEE, 2020: 781-785.
[37]	HASHEMIS R, SALEHI S S M, ERDOGMUS D, et al.Asymmetric loss functions and deep densely-connectednetworks for highly-imbalanced medical image segmentation:Application to multiple sclerosis lesion detection[J]. IEEE Access, 2019, 7: 1721-1735.
[38]	XUE Y Z, FARHAT F G, BOUKRINA O, et al. Amulti-path 2.5 dimensional convolutional neural networksystem for segmenting stroke lesions in brain MRIimages [J]. NeuroImage: Clinical, 2020, 25: 102118.
[39]	LI H L, PARIKH N A, WANG J H, et al. Objectiveand automated detection of diffuse white matter abnormalityin preterm infants using deep convolutionalneural networks [J]. Frontiers in Neuroscience, 2019,13: 610.
[40]	GROS C, DE LEENER B, BADJI A, et al. Automaticsegmentation of the spinal cord and intramedullarymultiple sclerosis lesions with convolutional neural networks[J]. NeuroImage, 2019, 184: 901-915.
[41]	RACHMADI M F, VALD′ES-HERN′ANDEZ M D C,AGAN M L F, et al. Segmentation of white matter hyperintensitiesusing convolutional neural networks withglobal spatial information in routine clinical brain MRIwith none or mild vascular pathology [J]. ComputerizedMedical Imaging and Graphics, 2018, 66: 28-43.
[42]	NAIR T, PRECUP D, ARNOLD D L, et al. Exploringuncertainty measures in deep networks for multiplesclerosis lesion detection and segmentation [J]. MedicalImage Analysis, 2020, 59: 101557.
[43]	XU B T, CHAI Y Q, GALARZA C M, et al. Orchestralfully convolutional networks for small lesionsegmentation in brain MRI [C]//2018 IEEE 15th InternationalSymposium on Biomedical Imaging (ISBI2018). Washington, DC, USA: IEEE, 2018: 889-892.
[44]	NACEUR M B, AKIL M, SAOULI R, et al. Fully automaticbrain tumor segmentation with deep learningbasedselective attention using overlapping patches andmulti-class weighted cross-entropy [J]. Medical ImageAnalysis, 2020, 63: 101692.
[45]	KUZINA A, EGOROV E, BURNAEV E. Bayesiangenerative models for knowledge transfer in MRI semanticsegmentation problems [J]. Frontiers in Neuroscience,2019, 13: 844.
[46]	GHAFFARI M, SOWMYA A, OLIVER R. Automatedbrain tumor segmentation using multimodalbrain scans: A survey based on models submitted tothe BraTS 2012—2018 Challenges [J]. IEEE Reviewsin Biomedical Engineering, 2020, 13: 156-168.
[47]	KUMAR A, UPADHYAY N, GHOSAL P, et al.CSNet: A new DeepNet framework for ischemic strokelesion segmentation [J]. Computer Methods and Programsin Biomedicine, 2020, 193: 105524.
[48]	WANG G T, SONG T, DONG Q, et al. Automatic ischemicstroke lesion segmentation from computed tomographyperfusion images by image synthesis andattention-based deep neural networks [J]. Medical ImageAnalysis, 2020, 65: 101787.
[49]	HU S Y, WENG W H, LU S L, et al. Multimodalvolume-aware detection and segmentation for brainmetastases radiosurgery [M]//NGUYEN D, XING L,JIANG S. Artificial Intelligence in Radiation Therapy.Cham, Switzerland: Springer, 2019: 61-69.
[50]	MAIER O, MENZE B H, VON DER GABLENTZ J,et al. ISLES 2015: A public evaluation benchmark forischemic stroke lesion segmentation from multispectralMRI [J]. Medical Image Analysis, 2017, 35: 250-269.
[51]	C?IC?EK ¨ O, ABDULKADIR A, LIENKAMP S S, etal. 3D U-Net: Learning dense volumetric segmentationfrom sparse annotation [M]//OURSELIN S,JOSKOWICZ L, SABUNCU M R, et al. Medical imagecomputing and computer-assisted intervention —MICCAI 2016. Cham, Switzerland: Springer, 2016:424-432.
[52]	MYRONENKO A. 3D MRI brain tumor segmentationusing autoencoder regularization [M]//CRIMI A,BAKAS S, KUIJF H, et al. Brainlesion: Glioma, multiplesclerosis, stroke and traumatic brain injuries.Cham, Switzerland: Springer, 2018: 311-320.
[53]	MOSTAPHA M, STYNER M. Role of deep learningin infant brain MRI analysis [J]. Magnetic ResonanceImaging, 2019, 64: 171-189.
[54]	WARING J, LINDVALL C, UMETON R. Automatedmachine learning: Review of the state-of-the-art andopportunities for healthcare [J]. Artificial Intelligencein Medicine, 2020, 104: 101822