How to take into account the min area of ​​the predicted mask in the Dice metric?

Question

Neural networks

When predicting, the neural network sometimes gives a mask with false positive pixel values ​​​​from 0.5-0.75, etc. The average number of such pixels is from 100 to 500 per one falsely predicted mask. I want to put a filter in the Dice metric that would filter only pixels with a value of 0.75 in the predicted mask, and if the number of such pixels is less than 800, then we can assume that there is nothing in this mask, the entire mask is filled with 0, if more, then normal. After training, in manual testing on pictures, all this works, after writing ifs to filter by the values ​​of each pixel and min area (800 pieces minimum), but how to do it in the metric so that all this would work already during training?
Loss and metric:

def dice_loss(y_true, y_pred):
    smooth=1e-6
    y_true_f = K.flatten(y_true)
    y_pred_f = K.cast(y_pred, 'float32')
    y_pred_f = K.flatten(y_pred)
    intersection = y_true_f * y_pred_f
    score = K.mean(1. - (2. * K.sum(intersection) + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth))
    return score

def dice_metric(preds, trues):
    preds = K.cast(preds, 'float32')
    return 1 - dice_loss(preds, trues)

Tried to use inside the metric:

preds_f = K.cast(K.greater(K.flatten(preds), 0.75), 'float32')

There was little benefit from this.

I tried this: How to make it in the metric so that during training it would filter pixels with a value greater than 0.75 and which are at least 800 pieces in one mask?
K.cast(K.greater(K.sum(preds), 800.0), 'float32')