An Improved Dice Loss for Pneumothorax Segmentation by Mining the Information of Negative Areas

Abstract

The lesion regions of a medical image account for only a small part of the image, and a critical imbalance exists in the distribution of the positive and negative samples, which affects the segmentation performance of the lesion regions. Dice loss is beneficial for the image segmentation involving an extreme imbalance of the positive and negative samples but it ignores the background regions, which also contain a large amount of information. In this work, we propose an improved dice loss that can mine the information in background areas and modify network architecture to improve performance. The improved dice loss called weighted soft dice loss (WSDice loss). Our loss function gives a small weight to the background area of the label, so the background area will be added to the calculation when calculating dice loss. It can also soft the hard label in the lesion area to increase the robustness of the model to noise label. What's more, we propose to cascade Focal loss and WSDice loss. Focal Loss is a Distribution-based loss function, WSDice Loss is a Regionbased loss function, the optimization directions of them are different. The cascaded loss function can make full use of the advantages of both and greatly improve model performance. In addition, we add a simple but effective channel attention module to the decode module of U-net. We experimented on the ChestX-ray8 datasets. Compared with Dice loss, WSDice loss improves the dice coefficient by 1.59%, cascaded loss function can improve dice coefficient by 7.81%. The improved in model architecture can increase the dice coefficient by 1.36%.

A Novel Focal Phi Loss for Power Line Segmentation with Auxiliary Classifier U-Net

“…It should also be noted that none of these losses penalize the misclassifications of the background information, making it difficult to optimize them for accurate background predictions [ 26 ]. Moreover, these loss functions and their variants [ 21 , 22 , 23 ] have so far been applied to medical segmentation tasks only where the foreground/minority class is characterized by a small but compact RoI as opposed to PL detection tasks where the RoIs are thin and more widespread. Lastly, the class imbalance in PL datasets is more severe, i.e., 2:98 for the minority versus majority class, respectively, as compared to the medical datasets where the class imbalance levels are of the order of 20:80, approximately.…”

“…These loss functions also introduce additional weighting parameters to weight the constituent losses and thus, require tuning to achieve optimal performance. Furthermore, these losses have also been applied to medical segmentation tasks only [ 22 , 33 ] where the class imbalance levels are less severe and RoIs are more small and compact than the PL detection tasks. Apart from the tuning of weights, the learning rates (LRs) are also difficult to configure and optimize for the compound losses due to the varying nature of the constituent losses.…”

“…Nevertheless, it has been demonstrated that other kinds of losses can outperform the popular BCE loss for typical classification tasks [ 17 , 18 ]. Region-based losses such as the Dice loss (DL) [ 19 ], Tversky loss (TL) [ 20 ] and their variants [ 21 , 22 , 23 ], have outperformed the BBCE loss for handling the class imbalance issue in medical image segmentation [ 24 ] but these losses have not yet been extended to other semantic segmentation applications. Medical image segmentation is characterized via small and compact regions of interest (RoIs) and less severe class imbalance as compared to the thin RoIs and more severe class imbalance in PL detection.…”

SUNet: Change Detection for Heterogeneous Remote Sensing Images from Satellite and UAV Using a Dual-Channel Fully Convolution Network

“…While by compounding with CE loss, IoU-WCE loss is more stable than IoU loss, and the problem of non-convergence is successfully overcome. Finally, The poor effect of Dice loss may be because it directly ignoring the background regions [51], which causes information loss during training.…”

Coastline extraction from ALOS-2 satellite SAR images

“…Because we planned to create an ensemble of different models, we trained several models based on U-Net with EfficientNet backbone. Regarding the loss function, the commons are based on group characteristics such as intersect over union, Tversky loss, or Sorensen-Dice loss (Wang et al 2020), or pixel characteristics, like Binary cross-entropy (BCE) or Focal loss (Lin et al 2017b). Each of them has pros and cons.…”