### Segmentation Loss Odyssey --- #### Distribution-based loss (最大化近似兩個幾率分佈 ) 1. ***Cross entropy (CE)***: minimize CE 等價於 minimize KL divergence / maximize likelihood estimation 2. ***TopK Loss***: 透過threshold來篩選出對模型較難學習的樣本 3. ***Focal Loss***: 透過權重改善 foreground-background class imbalance 4. ***Distance map penalized CE loss***: 有 distance penalty --- #### Region-based Loss (key element is Dice Loss) ***這部分我的理解是從 confusion matrix 那邊推導衍生出來的（比較針對 metrics 導向優化的loss function ex.Dice coefficient/IOU）*** 1. ***Sensitivity-specificity loss*** 2. ***Dice Loss (公式含義有點像 F1-score)*** 3. ***IOU Loss*** 4. ***Tversky Loss*** 5. ***Generalized Dice Loss*** 6. ***Focal Tversky Loss*** 7. ***Assymetric similarity loss*** 8. ***Penalty Loss*** --- #### Boundary-based Loss ***Minimize the distance between GT and predicted segmentation*** 1. ***Boundary (BD) Loss***： 透過積分boundary區域 改善unbalanced segmentation 2. ***Hausdorff Distance (HD) Loss*** --- #### Connection among Dice/BD/HD Loss - For Dice Loss, mismatch is weighted by the sum of the number of foreground pixels in the segmentationand the number of pixels in GT. - For BD loss, it is weighted by the distance transform map of GT - For HD loss, it use the distance transform map of GT for weighting and the distance transform map of the segnmentation --- #### Compound Loss 1. Combo Loss: weighted sum between weighted CE and Dice Loss 2. Exponential Logarithmic loss: exponential and log transforms to both Dice Loss and CE loss ---