Plug-and-Play with 2.5D Artifact Reduction Prior for Fast and Accurate Industrial Computed Tomography Reconstruction

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Plug-and-Play with 2.5D Artifact Reduction Prior for Fast and Accurate Industrial Computed Tomography Reconstruction".

Jane: Here's the gist of the paper. The gist The proposed method improves XCT reconstruction quality and defect detection in case studies from additively manufactured parts.

Tom: Plug-and-Play with 2.5D Artifact Reduction Prior The paper proposes a Plug-and-Play (PnP) reconstruction method that utilizes a 2.5D artifact reduction CNN as the prior, which leverages inter-slice information from adjacent slices to capture richer spatial context while remaining computationally efficient.

Jane: This approach is an improvement over previous methods that use a 2D CNN for artifact reduction, which only captures slice-independent information from the 3D reconstruction.

Tom: The method is presented in Algorithm 1, which iteratively updates the volume by applying the prior and a regularization sub-problem.

Paper summary: Tom: So, we're looking at this paper, "Plug-and-Play with 2 point 5D Artifact Reduction Prior for Fast and Accurate Industrial Computed Tomography Reconstruction" <ref:2506.14719#pg1,Plug-and-Play with 2.5D Artifact Reduction Prior for Fast and>. The main idea here is tackling the slow and expensive process of getting high-quality three dee reconstructions from X-ray scans by using a plug-and-play approach with a new kind of prior <ref:2506.14719#pg1>.

Jane: It sounds like they’re trying to cut down on those many X-ray measurements needed for good quality, especially when you're dealing with dense materials, which is really important for industrial stuff. So, what exactly is this 2 point 5D artifact reduction CNN they propose <ref:2506.14719#pg1>?

Tom: They are using a 2 point 5D artifact reduction CNN as the prior instead of the 2D one that was used in previous work; it uses inter-slice information from adjacent slices to capture richer spatial context without being too computationally expensive <ref:2506.14719#pg1>.

Lu: That’s interesting because most prior work, like what they mention in thirteen, is plug-and-play for model-based reconstruction, but this specifically modifies the prior itself by incorporating multi-slice information along the channel dimension to estimate the center slice.

Meng: But if we're talking about deploying this in a real setting, how does that 2 point 5D structure actually translate to something that runs fast enough when we're trying to ship models <ref:2506.14719#pg1>? Does it add a lot of overhead?

Lalam: The paper says they use a UNet architecture from thirty-eight, but instead of inputting just one slice, they provide a stack of five neighboring slices as input, which modifies the number of channels in the first layer.

Jane: That modification sounds like it’s trying to give the network more context about what’s happening between slices, which is something we always want when reconstructing three dee volumes <ref:2506.14719#pg1>. So, what's the naive question a listener might have about why this helps?

Tom: A listener might ask why this is better than just using a standard 2D prior; the paper claims that this 2 point 5D prior doesn't just improve image quality generally, it enables the model to directly suppress commonly occurring XCT artifacts, like beam hardening <ref:2506.14719#pg1>.

Lu: That’s a big claim because it says they eliminate the need for artifact correction pre-processing, which is a huge time saver for anyone in industrial inspection.

Meng: Does that mean we don't even have to run an extra step before we can use the final reconstruction? Because if we’re running this on real traffic, that pre-processing still takes time and resources.

Paper summary: Tom: The paper shows experiments on both experimental and synthetic cone-beam XCT data demonstrating that this 2 point 5D prior better preserves fine structural details, specifically pore size and shape, which leads to more accurate defect detection compared to the 2D priors they tested <ref:2506.14719#pg1>.

Jane: So, if we look at the numbers, what kind of performance gain are they seeing when comparing this method against the previous 2D priors <ref:2506.14719#pg1>?

Tom: They show that on experimental XCT data using a 2 point 5D prior trained entirely on simulated scans, the method shows strong performance and better SNR and CNR for experimental XCT short and sparse scans of an Al-Ce part with one hundred forty-five views.

Lu: That’s where the generalization aspect comes in; they trained the prior only on simulated data, but it performs well enough on experimental scans too, which highlights its ability to generalize across domains.

Meng: If it's trained only on synthetic scans, how sure are we that this won't break when we introduce real-world noise or different material types? Does it handle Out-of-Distribution data well?

Lalam: The paper notes that the 2 point 5D PnP framework demonstrates strong generalization capabilities to Out-of-Distribution data with differing noise levels, but they also flag that the increase in false positives for 2 point 5D PnP arises from “porelike” structures induced by view sparsity artifacts that appear consistently across neighboring slices <ref:2506.14719#pg1>.

Jane: So, even though it's good at generalizing, there’s a specific trade-off mentioned regarding those porelike structures when dealing with differing noise or different data sources?

Tom: Exactly, and they also show that the 2 point 5D PnP performs better on In-Distribution data compared to the 2D PnP, especially in matching the reference reconstruction pixel intensity both at the edges and the center of the reconstruction, which reduces cupping artifacts <ref:2506.14719#pg1>.

Lu: That focus on matching intensities at both edges and center sounds like it directly addresses some fundamental issues with how these reconstructions are built up iteratively.

Meng: From a production standpoint, if we're using this to detect defects, what does that improved accuracy actually mean for the end user who is inspecting the part? Does it translate to fewer missed defects or just cleaner images?

Tom: The conclusion says that 2 point 5D PnP achieves better image quality metrics like NRMSE and SSIM for almost all In-Distribution testing scans, and it consistently achieves significantly higher recall and precision for InD data, which implies it enables more accurate defect detection over all defect diameters <ref:2506.14719#pg1>.

Paper summary: Jane: So if you're only listening to the show, what does this mean for your job or what you do in your day-to-day? Does this change how you think about XCT reconstruction problems?

Lu: It suggests that moving from a 2D prior to a 2 point 5D prior is a practical way to get better image quality metrics while maintaining computational complexity lower than full three dee networks, which is what they were aiming for in the paper <ref:2506.14719#pg1>.

Meng: I just see it as a way to get usable results faster, because the runtime comparison shows that both the 2D PnP and this proposed 2 point 5D PnP take approximately forty-eight minutes using four Nvidia H100 GPUs, which is significantly less expensive than some other methods like MBIR that require about six hours per volume.

Lalam: The paper emphasizes this trade-off between computational cost and performance, showing the 2 point 5D design offers a favorable balance by capturing important inter-slice context while remaining significantly more memory-efficient and easier to train than full three dee networks <ref:2506.14719#pg1>.

Tom: So, to wrap up on this "Plug-and-Play with 2 point 5D Artifact Reduction Prior for Fast and Accurate Industrial Computed Tomography Reconstruction", it’s essentially about using that multi-slice information efficiently within a plug-and-play framework to get better defect detection without needing extra pre-processing steps <ref:2506.14719#pg1,Plug-and-Play with 2.5D Artifact Reduction Prior for Fast and>.

Jane: It sounds like they've shown that this approach is robust enough for real applications, even though they also noted the limitation that the 2 point 5D prior can introduce false positives related to view sparsity artifacts when dealing with varying noise levels <ref:2506.14719#pg1>.

Lu: That’s a fair caveat; the method doesn't entirely eliminate artifacts caused by sparse views, it just makes them better at suppressing beam hardening and improving detail preservation overall.

Meng: For me, what matters is that we have a method that scales well enough and runs in under an hour on modern hardware, which this 2 point 5D design achieves compared to those slower methods <ref:2506.14719#pg1>.

Tom: So the main implication is that for industrial inspection where speed and accuracy are both needed, this 2 point 5D PnP reconstruction method provides a solid path forward by improving metrics like recall and precision on In-Distribution data with high confidence from their training on synthetic scans <ref:2506.14719#pg1>.

Conclusion: Tom: So, we’re wrapping up on this paper, "Plug-and-Play with two point 5D Artifact Reduction Prior for Fast and Accurate Industrial Computed Tomography Reconstruction <ref:2506.14719#pg3>." It boils down to using a modified plug-and-play reconstruction method that uses a two point five dimensional prior instead of the usual two dimensional one.

Jane: That’s right, and the authors are looking at how you can get better three dee reconstructions from X-ray scans faster by incorporating information from adjacent slices.

Lu: What’s really interesting here is how they use that multi-slice information along the channel dimension to estimate the center slice; it means they get richer spatial context without needing a full three dimensional network.

Meng: From what I’m seeing, they're trying to keep the computational cost down because a full three dimensional network can be really heavy to run in a production environment.

Lalam: I’ve read through the literature, and this approach is trying to solve two things at once: getting better image quality metrics like NRMSE and SSIM on In-Distribution data, and making it scalable for real-world use.

Tom: So what does this mean for someone who only listens to the show? It means that industrial inspection can get more accurate defect detection because the method has a higher recall and precision than previous methods.

Jane: Exactly, so if you're in manufacturing or inspection, this suggests you might be able to find smaller defects that you were missing before.

Lu: But there’s a caveat they mention too; the authors note that using this two point five dimensional prior can sometimes create false positives related to view sparsity artifacts when the noise levels are different.

Meng: That’s important because it means we can't just swap in this model and expect perfect results everywhere without checking those noise levels first.

Lalam: I'm pretty sure the confidence in their performance on In-Distribution data is high, but they did show strong generalization to Out-of-Distribution data with different noise levels too.

Tom: So the implication here is that we’re moving towards a faster and more accurate way of doing XCT reconstruction for industrial parts, provided you keep an eye on those artifacts when you deploy it.

Jane: Right, so they’ve shown a practical path to improving three dee reconstruction quality without needing massive amounts of data or overly complex models.

Lu: And the next thing we can explore is how they are scaling this prior modeling beyond just XCT data.