Loss Weightings for Improving Imbalanced Brain Structure Segmentation Using Fully Convolutional Networks

Abstract

Brain structure segmentation on magnetic resonance (MR) images is important for various clinical applications. It has been automatically performed by using fully convolutional networks. However, it suffers from the class imbalance problem. To address this problem, we investigated how loss weighting strategies work for brain structure segmentation tasks with different class imbalance situations on MR images. In this study, we adopted segmentation tasks of the cerebrum, cerebellum, brainstem, and blood vessels from MR cisternography and angiography images as the target segmentation tasks. We used a U-net architecture with cross-entropy and Dice loss functions as a baseline and evaluated the effect of the following loss weighting strategies: inverse frequency weighting, median inverse frequency weighting, focal weighting, distance map-based weighting, and distance penalty term-based weighting. In the experiments, the Dice loss function with focal weighting showed the best performance and had a high average Dice score of 92.8% in the binary-class segmentation tasks, while the cross-entropy loss functions with distance map-based weighting achieved the Dice score of up to 93.1% in the multi-class segmentation tasks. The results suggested that the distance map-based and the focal weightings could boost the performance of cross-entropy and Dice loss functions in class imbalanced segmentation tasks, respectively.

Deep 3D Neural Network for Brain Structures Segmentation Using Self-Attention Modules in MRI Images

“…Our intuition in this regard is that this is due, mainly, to the class imbalance problem and lack of enough data to train the model for those structures specifically. For this reason, it is important in the future to explore other methods that allow addressing this problem, for example, improving the calculation of the weights of the classes used in the loss functions similar to what is performed in [ 74 ], or using additional data augmentation techniques to increase the samples of classes with less information. Another factor that we considered in the analysis is the fact that deep learning methods based on transformers lack the inductive biases inherent in CNNs requiring large amounts of data to be able to generalize well [ 77 ], so their usage in small-size medical datasets remains difficult without any internal modification in their self-attention module.…”

“…Even the size difference between the structures and the background is usually significant. Therefore, multiple loss functions and weighting strategies for loss functions were proposed for improving imbalanced brain structure segmentation [ 74 ]. In the proposed approach, we used a combination of Dice Loss [ 75 ] and Focal Loss [ 76 ].…”

Deep Learning and Domain-Specific Knowledge to Segment the Liver from Synthetic Dual Energy CT Iodine Scans

“…The former could be addressed by scanning with smaller voxel dimensions [3] or by resampling scans into smaller voxel dimensions during the preprocessing steps. The latter could be addressed by implementing a class balancing scheme according to the pixel-wise frequency of each class in the dataset [52]. Since the goal of the current paper was to assess the value of synth-DECT scans, we did not implement class balancing schemes to mitigate the errors found at the margins of the scans.…”

Deep Learning for the Automatic Quantification of Pleural Plaques in Asbestos-Exposed Subjects

“…To reduce memory consumption without impacting performance, we selected a basis of 24 filters of 3 × 3 − 24 for the first layer, 48 for the second and so on, as proposed in [23]. The loss function (L) was a combination of binary cross-entropy (L BCE ) [24] and Dice loss (L DL ) [25] which is demonstrated to be well suited for imbalanced structure segmentation [26]. L was defined as:…”