Confustion about predictive structure-wise variance about bayesian_unet

[ThisGame42]

Hi @yuta-hi thanks very much for this detailed explanation. Now I understand how PSV is derived. Much appreciated!

Closing this issue as well. Have a good day!

[yuta-hi]

The point of this equation is to calculate the uncertainty per unit volume.
Here's a simple code for it.

n_class = 10

predictor = BayesianUNet(ndim=2,
                         out_channels=n_class)

model = MCSampler(predictor,
                  mc_iteration=args.mc_iteration,
                  activation=partial(F.softmax, axis=1),
                  reduce_mean=partial(F.argmax, axis=1),
                  reduce_var=partial(F.mean, axis=1)) # NOTE: Average the predicted variance of each class

infer = Inferencer(test_iter, model, device=args.gpu[0])

label, variance = infer.run()

# NOTE: 
# label.shape: (number of slices, 512, 512)
# variance.shape: (number of slices, 512, 512)

# convert the pixel-wise variance to the structure-wise variance
for i in range(1, n_class): # Zero is assumed to be the background.
    mask = label == i
    psv = np.mean(variance[mask]) # NOTE: Normalize to the variance per unit volume
    print('PSV (#%02d) = %f' % (i, psv))