⚡ electrical engineering

Deep Slice Interpolation for Reducing Through-Plane Anisotropy and Noise in Head CT

This paper presents a deep learning system that synthesizes intermediate CT slices to halve through-plane spacing, simultaneously reducing anisotropy and noise in head CT scans while demonstrating superior performance over classical and video interpolation baselines through rigorous evaluation of various loss functions.

Original authors: Luis Cortés Ferre, Miguel A. Gutiérrez-Naranjo, Marcin Balcerzyk

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Luis Cortés Ferre, Miguel A. Gutiérrez-Naranjo, Marcin Balcerzyk

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a loaf of bread that you want to slice. In a perfect world, you'd use a very sharp knife to cut thin, even slices. But in the real world of medical CT scans, the "knife" (the scanner) often leaves gaps. It takes a picture of the brain, skips a chunk of space, takes another picture, and skips again.

This creates a "loaf" where the slices are thin and crisp from side-to-side (in-plane), but thick and chunky from top-to-bottom (through-plane). If you try to look at this loaf from the side (a side view), it looks blocky and jagged, like a staircase. This makes it hard for doctors to see small details or measure things accurately.

This paper presents a smart digital slicer that fixes this problem. Here is how it works, explained simply:

1. The Magic Trick: Guessing the Missing Slice

The system looks at two existing slices of a brain scan (let's call them Slice A and Slice C). It uses a powerful AI brain (a deep learning model) to guess what the missing slice in the middle (Slice B) should look like.

The Analogy: Imagine you are looking at a flipbook animation. You have Page 1 and Page 3, but Page 2 is missing. A human might guess what happens in between, but this AI is incredibly good at it. It doesn't just blur the two images together; it "hallucinates" the missing details to create a smooth, new image that fits perfectly between the two.
The Result: It effectively cuts the loaf in half, turning thick, chunky slices into thin, smooth ones. This removes the "staircase" effect and makes 3D views look much better.

2. The Bonus: It's Also a Noise Eraser

Here is the clever part: while the AI is guessing the missing slice, it accidentally (or rather, mathematically) does something else very useful. It removes the "grain" or "static" from the image.

The Analogy: Think of a grainy, noisy photo. If you ask a smart friend to describe the scene based on two other grainy photos, they will likely describe the actual scene (the person, the tree, the car) rather than the random static on the film. They ignore the noise because the noise is different in every photo, but the scene stays the same.
The Science: The paper explains that because the AI is trained to find the "average" truth between two images, it naturally filters out the random static (noise) that appears in medical scans. So, you get a sharper image and a cleaner image at the same time.

3. The "Recipe" Problem: Finding the Right Loss Function

To teach the AI, the researchers had to give it a "scorecard" (called a loss function) to tell it how well it was doing. They tried many different scorecards:

Pixel-by-Pixel: "Did you get every single dot of color right?"
Structure: "Does the shape of the brain look right, even if the colors are slightly off?"
Mixtures: A combination of both.

The Discovery: They found that some scorecards were very fragile. One popular scorecard (SSIM) was like a tightrope walker: if the AI took a slightly wrong step (a specific setting for learning speed or batch size), it would fall off the tightrope and the whole training would crash. They had to add a "safety net" (a numerical fix) to keep it stable.

They concluded that a mix of "Structure" and "Pixel" scoring worked best as a reliable, all-around recipe.

4. What They Tested

The Data: They used thousands of real head CT scans from a public dataset (RSNA 2019), focusing on brain scans.
The Competition: They compared their AI against:
- Old-school methods: Like simple averaging (blending two slices) or mathematical curves (splines).
- Video AI: They tried using AI models designed to guess missing frames in videos (like a movie).
The Winner: Their custom AI beat everyone. It created slices that looked more like the real thing than the old methods or the video AI did.

5. Important Limits (What They Didn't Say)

The paper is very careful about what it claims:

Head Only: This AI was trained only on heads. It doesn't know how to fix a knee or a liver.
No Doctor Check: They did not ask real doctors to look at the images and say, "Yes, this helps me diagnose patients." They only measured the math (how close the pixels were).
Not a Magic Wand: If a tiny blood vessel exists in the missing gap but isn't hinted at in the slices above or below, the AI cannot invent it. It can only guess what is likely there based on the neighbors.

Summary

The paper presents a tool that takes chunky, blocky brain scans and uses AI to fill in the gaps, creating smooth, thin slices. As a happy side effect, it also cleans up the "static" noise in the images. It works better than old math tricks and even better than AI designed for movies, but it currently only works on heads and hasn't been tested by doctors yet.

Technical Summary: Deep Slice Interpolation for Reducing Through-Plane Anisotropy and Noise in Head CT

Problem Statement
Head computed tomography (CT) volumes are typically reconstructed with sub-millimeter in-plane resolution but significant through-plane spacing (2–5 mm). This anisotropy degrades multiplanar reconstructions, introduces stair-step artifacts in coronal and sagittal views, and hinders volumetric measurements and downstream algorithms that assume near-isotropic voxels. While classical interpolation methods (e.g., cubic spline) exist, they cannot recover anatomical detail absent from the acquired slices. Furthermore, existing deep learning approaches for medical super-resolution often lack systematic comparisons of loss functions within a fixed architecture, and the relationship between interpolation and implicit noise reduction remains under-explored in this specific context.

Methodology
The authors propose a deep learning system to synthesize an intermediate axial slice ( $I_{k+1}$ ) from two neighboring slices ( $I_k, I_{k+2}$ ), effectively halving the through-plane spacing.

Architecture: The system utilizes a U-Net architecture with an EfficientNetV2-S encoder. The encoder is initialized with ImageNet-pretrained weights (adapted for 2-channel input), while the decoder is initialized from scratch. The model processes 256 × 256 patches, which are reassembled into 512 × 512 slices during inference using a 13-patch tiling strategy (9 overlapping center crops + 4 corner patches).
Dataset: The model is trained on the RSNA 2019 Intracranial Hemorrhage Detection dataset, restricted to non-contrast head CT. The data is windowed to [−20 HU, 107 HU] (center 44 HU, width 128 HU) and normalized to [0, 1]. The dataset is split patient-wise into training, validation, and test sets. The test set (30 patients) is stratified to ensure representation of all five hemorrhage subtypes.
Loss Function Evaluation: A systematic evaluation of eleven loss configurations was conducted under a fixed training protocol. The study compares:
- Pixel-wise losses: Mean Squared Error (MSE) and Mean Absolute Error (L1).
- Structural losses: Structural Similarity Index (SSIM) and Multi-Scale SSIM (MS-SSIM).
- Hybrid losses: Combinations of structural and pixel-wise terms (e.g., MS-SSIM + L1).
Numerical Conditioning: The authors identified and addressed training instability in SSIM-family losses. They found that default stability constants ( $K_2 = 0.03$ ) lead to division-by-near-zero errors in low-variance brain regions, causing NaN gradients. They implemented two remedies for training: increasing $K_2$ to 0.4 and enforcing full float32 precision for SSIM forward passes.
Statistical Protocol: Evaluation is performed at the patient level (30 independent patients) using non-parametric bootstrap confidence intervals and paired Wilcoxon signed-rank tests to compare against classical baselines and pretrained video frame interpolation (VFI) models (RIFE, FILM).

Key Contributions

Performance Benchmarking: The proposed deep learning system outperforms classical interpolation (cubic spline, mean) and pretrained VFI models (RIFE, FILM) on all structural metrics (SSIM, MS-SSIM) in a controlled head CT domain.
Loss Function Analysis: The paper provides a systematic evaluation of loss functions, identifying MS-SSIM + L1 (with a 0.5/0.5 weighting) as a robust, balanced default. It highlights that no single loss dominates all metrics; L1 minimizes pixel error (MAE), while MS-SSIM + L1 yields the highest PSNR.
Training Stability Insights: The study documents systematic training divergence in SSIM-family losses, attributing it to numerical conditioning issues in the SSIM denominator. It provides actionable remedies (adjusted $K_2$ and precision handling) and notes that residual instability persists at smaller batch sizes.
Implicit Denoising: The authors demonstrate that regression-based synthesis inherently produces denoised outputs. By leveraging the conditional-expectation property of regression losses (where the minimizer of MSE is the conditional mean), the model learns the underlying anatomical signal rather than the acquisition noise, even when trained on noisy targets.

Results

Structural Metrics: On the held-out test set, the learned models significantly outperform baselines. The reference model (MS-SSIM + L1) achieves a patient-level SSIM of 0.888 and MS-SSIM of 0.936, compared to ~0.78 for classical methods and ~0.85 for VFI models.
Implicit Denoising: Quantitative analysis of uniform white-matter regions of interest (ROIs) shows that pixel-wise regression models (L1, MSE, MS-SSIM + L1) reduce in-ROI variance by approximately 26% compared to the acquired reference slices. This confirms the theoretical prediction that the model suppresses unpredictable noise.
Pathology Impact: Interpolation quality is slightly lower on slices containing hemorrhage compared to normal slices across all methods, a gap that is statistically significant for SSIM metrics.
External Validation: An out-of-distribution case study on a head CT series with mixed slice spacing demonstrated that the model's residuals follow the theoretical predictions of a conditional-expectation estimator, specifically attenuating data-side parity artifacts without explicit training on such patterns.

Significance and Claims
The paper claims that deep learning-based slice interpolation offers a dual benefit: improving 3D visualization by reducing anisotropy and simultaneously providing implicit denoising. This is achieved through a single inference pass without requiring clean training targets.

The authors emphasize that their findings are specific to non-contrast head CT within a defined Hounsfield Unit window. They explicitly state that the work does not claim generalization to other anatomical regions, modalities, or acquisition protocols. The significance lies in providing a reliable, denoised interpolation method that outperforms existing classical and transfer-learning baselines, while offering practical guidance on the numerical stability of structural similarity losses in medical imaging. The paper concludes that while the interpolation quality is high, clinical validation via radiologist reader studies is a necessary next step before deployment.