Analysis of the impact of loss functions in U-Net architecture for segmentation of right ventricle

The present paper sheds light on the effect of loss functions in U-Net architecture for the segmentation of the right ventricle. Five loss functions namely binary cross-entropy, dice, inverse dice, dice combo and combined have been tested using optimisers such as Adam, stochastic gradient descent and root mean square propagation. The accuracy of the U-Net model is measured using the popular dice coefficient metric. The two loss functions dice and dice combo achieved maximum dice coefficients of 0.7825 and 0.7633 with stochastic gradient descent respectively. The result also shows that the loss functions such as dice and dice combo give acceptable dice coefficients with all three chosen optimisers. The loss functions binary cross entropy, combined and inverse dice have achieved moderate dice coefficients value with Adam and root mean square propagation optimisers but have shown very poor performance with stochastic gradient descent optimiser. The dice and dice combo loss functions with stochastic gradient descent optimiser are good candidates for segmentation of the right ventricle in U-Net architecture.