PoreTimeNet segments a whole micro-CT sequence at once, cutting carbonate porosity error from 0.97 to 0.28 points

Segmentation of time-resolved micro-CT moves from independent per-image processing to joint processing of several three-dimensional images within one temporal window, bringing pore-volume growth and rock boundaries closer to the reference segmentations. Previous 3D deep networks segmented each image separately without information from other images in the sequence, so errors varied between images and could be mistaken for physical change; thresholding performed poorly on noisy, low-contrast images and needed manual adjustment per image. On a sequence of oil displacement by CO2-saturated brine with carbonate dissolution, compared with a 3D model sharing the same spatial architecture but no temporal information, mean absolute porosity error fell from 0.97 to 0.28 percentage points and rock boundary F1 rose from 0.833 to 0.941.

The segmentations feed directly into flow and reactive-transport calculations, so differences in pore geometry appear as differences in estimated permeability and mineral dissolution rate. The 3D model underestimated the final-to-initial permeability ratio by about 27%, overestimating permeability by 13% at 0 min and underestimating it by 18% at 646 min; the self-configuring 3D network nnU-Net underestimated that ratio by 41%. Single-phase flow and reactive transport were simulated with GeoChemFoam at 0, 187 and 646 min; the 4D model estimated the permeability ratio with a mean absolute relative error of 6.7%, and at 646 min pore volume was 1.71 times its initial value in the reference segmentation, 1.74 for 4D, 1.57 for 3D and 1.53 for nnU-Net.

The gain comes from information supplied by other images in the sequence, not from additional data corruption or larger network capacity. Adding input corruption such as noise, masking or greyscale jitter to the 3D model left porosity error, fluid Dice and rock boundary F1 close to the original 3D model; a 4D model retaining the full architecture but processing only the current time step also stayed close to the 3D model in accuracy. At inference, repeating the current image to fill the input sequence while keeping the trained 4D weights unchanged dropped mean rock boundary F1 from 0.941 to 0.808 and raised porosity error, with the largest drops at the four time steps where the 3D model had the lowest boundary accuracy.

Evaluation covers six time-resolved sequences: a dissolution sequence in one Ketton sample, drainage and imbibition in the same sample, oil flooding in Luxembourg sandstone, and mixed-wet and water-wet dissolution sequences from a second Ketton sample; transfer tests use no fine-tuning and no target-based checkpoint selection.

The authors list faster laboratory CT imaging as a potential use and note that the model uses relative image positions rather than elapsed time in seconds, so it could be retrained for more frequent imaging; that use has not been tested experimentally. In the Ketton drainage sequence, adding input corruption to the 3D model recovered much of the Dice accuracy gain while 4D porosity predictions remained more consistent; oil boundary F1 in Seq2 was slightly higher for 3D than for 4D.