Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations

By C. Sudre, Wenqi Li, Tom Kamiel Magda Vercauteren, Sébastien Ourselin, M. Jorge Cardoso. DLMIA/ML-CDS@MICCAI, 2017

2,736 citations recorded by the source corpus.

Deep-learning has proved in recent years to be a powerful tool for image analysis and is now widely used to segment both 2D and 3D medical images. Deep-learning segmentation frameworks rely not only on the choice of network architecture but also on the choice of loss function. When the segmentation process targets rare observations, a severe class imbalance is likely to occur between candidate labels, thus resulting in sub-optimal performance. In order to mitigate this issue, strategies such as the weighted cross-entropy function, the sensitivity function or the Dice loss function, have been proposed. In this work, we investigate the behavior of these loss functions and their sensitivity to learning rate tuning in the presence of different rates of label imbalance across 2D and 3D segmentation tasks. We also propose to use the class re-balancing properties of the Generalized Dice overlap, a known metric for segmentation assessment, as a robust and accurate deep-learning loss function for unbalanced tasks.

Subfields: Computer Vision and Pattern Recognition, Neurology

View Paper: https://www.semanticscholar.org/paper/2d8edc4e38bf9907170238726ec902cb3739393b

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Neurology

Citations in the corpus

MASK-RL: Multiagent Video Object Segmentation Framework Through Reinforcement Learning (2020, background): “…given in Table VI, showing that the dice loss leads to better performance than cross-entropy loss (as already found in recent works, e.g., [34] and [40], especially in cases of unbalanced data sets) and that increasing the number of input frames does not affect performance signiﬁcantly—at least,…”
Continual Segmentation with Disentangled Objectness Learning and Class Recognition (2024, positive): “Then, we optimize this stage using ground truth masks: where L mask is the sum of binary cross-entropy loss and Dice loss [43].”
Unified Focal loss: Generalising Dice and cross entropy-based losses to handle class imbalanced medical image segmentation (2021, methodology): “…(FN): The Dice loss ( L DSC ), can therefore be defined as: Other variants of the Dice loss include the Generalised Dice loss (Crum et al., 2006; Sudre et al., 2017) where the class weights are corrected by the inverse of their volume, and the Generalised Wasser-stein Dice loss (Fidon et al.,…”
Unknown Object Segmentation from Stereo Images (2021, methodology): “loss [56] Ldice = 1 − 2yŷ+1 y+ŷ+1 as cost function, the optimal assignment is computed by the Hungarian method [57].”
Automated brain extraction of multisequence MRI using artificial neural networks (2019, positive): “sed at a specific epoch and 𝛼0=10−4 is the initial learning rate. Motivated by successful recent work (Drozdzal, et al., 2016; Isensee, et al., 2017; Kayalibay, et al., 2017; Milletari, et al., 2016; Sudre, et al., 2017) a soft dice loss formulation for training the network was used. �𝐷(�,�)= − 2 ”
Deep learning in radiology: an overview of the concepts and a survey of the state of the art with focus on MRI (2018, methodology): “weighted cross entropy was used in [73] for brain structure segmentation in MRI or Dice coefficient-based loss for brain tumor segmentation in MRI [74].”
nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation (2018, methodology): “Based on past experience [13,1] we favor this formulation over other variants [8,15].”
No New-Net (2018, methodology): “While several formulations of the dice loss exist in the literature [25,16,17], we prefer to use a multi-class adaptation of [16] which has given us good results in segmentation challenges in the past [8,6].”
MONAI: An open-source framework for deep learning in healthcare (2022, background): “The added functions introduce specialized losses not present in the base library, such as Dice and its variants (generalized, masked, focal, etc.) [22], focal loss, Tversky loss, contrastive loss, those used for image registration (bending energy, multi-scale loss, etc.), and several others.”
Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory and Practice (2019, methodology): “Indeed, in the risk minimization framework for medical image segmentation, there are numerous examples where each of these measures are optimized [8,15,19].”