Brain age prediction non-deterministic ?

[tletut0ur]

Hi, I am using the Brain Age prediction model and found out that the predicted_age output can fluctuate for the same image if ran multiple times (not by much, it's usually by something like a year). What could explain this element of randomness if the weights are already computed ? Is there some random initialization in pre-preocessing ?
I am using python 3.11 with tensorflow==2.15.0, antspynet==0.3.2 and antspyx==0.6.3

[ntustison]

This is just a previously trained model I randomly found and ported into ANTsXNet but I don't use it at all. You might want to look at the original publication (see the help information) to see if there are any random network components. Also, I did include some randomness in the number of simulations but that is off by default.

[tletut0ur]

After some investigation. I can almost conclusively say that this is due to preprocessing. I ran the following code on the IXI 100_Guys image to compute brain age 5 times in a row in different conditions:

from antspynet.utilities import preprocess_brain_image
import ants
import antspynet


t1_file = r"C:\path\to\datasets\IXI\100_Guys\SCANS\T1\NIfTI\IXI100-Guys-0747-T1.nii.gz"
t1 = ants.image_read(t1_file)
t1_preprocessing = preprocess_brain_image(t1,
    truncate_intensity=(0.01, 0.99),
    brain_extraction_modality="t1",
    template="croppedMni152",
    template_transform_type="antsRegistrationSyNQuickRepro[a]",
    do_bias_correction=True,
    do_denoising=True,
    verbose=False)
t1_preprocessed = t1_preprocessing["preprocessed_image"] * t1_preprocessing['brain_mask']
computed_ages_preprocessed = []
for i in range(5):
    age = antspynet.brain_age(
        t1_preprocessed, 
        number_of_simulations=0, 
        sd_affine=0.01, 
        verbose=False, 
        do_preprocessing=False
    )
    computed_ages_preprocessed.append(age["predicted_age"])
computed_ages_raw = []
for i in range(5):
    age = antspynet.brain_age(
        t1, 
        number_of_simulations=0, 
        sd_affine=0.01, 
        verbose=False, 
        do_preprocessing=True
    )
    computed_ages_raw.append(age["predicted_age"])
print("computed ages with explicit preprocessing", computed_ages_preprocessed)
print("computed ages with implicit preprocessing", computed_ages_raw)

What I call "explicit preprocessing" is when I run preprocess_brain_image once, outside the loop that computes the brain age. While "implicit preprocessing" just uses the do_preprocessing=True of the argument, which runs it 5 different times. The output (without many preceding warnings about tensorflow) is :

computed ages with explicit preprocessing [37.15592, 37.15592, 37.15592, 37.15592, 37.15592]
computed ages with implicit preprocessing [36.90978, 36.60119, 37.04059, 36.89186, 36.80053]

I'll now try to see if there is a specific step of the preprocessing that causes this. Is there any way to expose a seed argument to the preprocessing ?

PS: I also investigated this tensorflow warning oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable "TF_ENABLE_ONEDNN_OPTS=0". but it did not change anything.

EDIT : I was able to get constant output by adding these 3 lines at the start of my script

import os

os.environ["ANTS_RANDOM_SEED"] = "123"
os.environ["ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS "] = "1"

in accordance with what was said in this thread
It seems to work before and after importing ants and antspynet. I have not tried it after loading data.