Lidar Augment: Searching for Scalable 3D LiDAR Data Augmentations

Using Polynomial Loss and Uncertainty Information for Robust Left Atrial and Scar Quantification and Segmentation

“…Recently, Leng et al (2022) [13] proposed PolyLoss, a new loss function that expresses the commonly used loss functions such as cross-entropy (Eq. 1) and focal loss (Eq.…”

“…Using Taylor expansion, cross-entropy can be represented as sum of polynomial bases (1 − p) j , as shown in Eq. 2, where p is the prediction probability of the target class [13]. By dropping the higher-order polynomials and adding terms that perturb the polynomial coefficients, they came up with a simplified version of the polynomial loss called Poly-1.…”

“…By dropping the higher-order polynomials and adding terms that perturb the polynomial coefficients, they came up with a simplified version of the polynomial loss called Poly-1. This loss function modifies the cross-entropy by only adding one hyper-parameter (ϵ) [13], as can be seen in Eq. 2.…”

“…2. PolyLoss has shown good performance on computer vision tasks by outperforming cross-entropy and focal losses [13].…”

“…Inspired by [13], in this paper, we proposed a loss function that uses Dice loss with PolyLoss (Eq. 5) for a LA cavity and scar segmentation.…”