1.Dice Loss 與 Dice Coefficient
dice loss源自于dice coefficient分割效果評價标準, dice coefficient具體内容如下:
def dice_coefficient(y_true, y_pre):
eps=1e-5
intersection = tf.reduce_sum(y_true * y_pre)
union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pre) + eps
loss = 1. - (2 * intersection / union )
return loss
dice loss:
![](https://img.laitimes.com/img/9ZDMuAjOiMmIsIjOiQnIsIyZuBnL2MzNzQjMwITMwIDMxAjMwIzLc52YucWbp5GZzNmLn9Gbi1yZtl2Lc9CX6MHc0RHaiojIsJye.png)
适用場景: 醫學影像 圖像分割
2.Sensitivity-Specificity Loss
ss loss來源于準确度計算公式,綜合考慮靈敏度和準确度, 通過添加β系統平衡兩者之間的權重
具體損失:
def ssl(y_true,y_pred):
elpha = 0.1
TP = tf.reduce_sum(y_pred * y_true)
TN = tf.reduce_sum((1-y_pred )*(1- y_true))
FP = tf.reduce_sum(y_pred * (1-y_true))
FN = tf.reduce_sum((1-y_pred )* y_true)
eps = 1e-6
sensitivity = TP/(TP+FN)
specificity = TN/(TN+FP)
loss = elpha * sensitivity + (1-elpha) * specificity
return loss
使用場景: 需要側重TP,TN其中一個的的場合
3.Focal Tversky Loss
該損失由tversky loss改進,注重于困難樣本的訓練
首先, TI (tversky lndex):
A表示預測值,B表示真實值,|A-B|代表FP, |B-A|代表FN,調整α與β可調整權重.
然後,FTL( focal tversky loss):
FTL =Σ (1 − TI )^γ
γ取[1,3]