To combine the strengths and none of the weaknesses of both losses, we can simply combine them by adding them.

def softdiceloss(predictions, targets, smooth: float = 0.001):
    batch_size = targets.shape[0]
    intersection = (predictions * targets).view(batch_size, -1).sum(-1)

    targets_area = targets.view(batch_size, -1).sum(-1)
    predictions_area = predictions.view(batch_size, -1).sum(-1)

    dice = (2. * intersection + smooth) / (predictions_area + targets_area + smooth)
    return 1 - dice.mean()

def dice_bce_loss(predictions, targets, weights = (1,1)):
    '''
    Combination between the bce loss and the soft dice loss. 
    The goal is to get the advantages
    from the soft dice loss without its potential instabilities.
    '''
    soft_dice_loss = softdiceloss(predictions, targets)
    bce_loss = nn.BCEWithLogitsLoss()(predictions, targets)
    combination = weights[0] * soft_dice_loss + weights[1] * bce_loss
    return combination