12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.

org

Interactive Denoising of 3D Volumes Using Wavelets

Thomas Lang1 , A. Michael Stock2

1
Fraunhofer IIS, Division Development Center X-ray Technology, Passau, Germany, e-mail: [email protected]
2
FORWISS, University of Passau, Passau, Germany

Abstract
Conventional signal denoising methods either focus on statistical properties or employ a (complicated) set of parameters a user
needs to tweak appropriately. More recently, AI-based denoising methods are increasingly employed due to their robustness.
However, these methods typically require a large number of training datasets, which are costly to produce in a computed tomo-
More info about this article: https://www.ndt.net/?id=27718

graphy scenario. This work introduces a denoising procedure which combines classifier-based decision making with interactive
user input in the wavelet domain. These interactive user inputs allows for an easy and straight-forward specification of noisy
regions, which works especially well in the wavelet basis, without the need for statistical properties or large numbers of para-
meters. At the same time, the obtained results show that noise in the scanned data is reduced considerably while keeping the
structural similarity intact.

Keywords: Digital Signal Processing; Artifical Intelligence; Interactive denoising; Wavelets

1 Introduction
Every real-world data acquisition is subject to noise, and unfortunately computed tomography is no exception. Since noise often
prevents or even hinders further processing of the scanned data, denoising is a crucial part of many data processing pipelines.
Therefore, a broad variety of different techniques emerged over the last decades, both in terms of classical signal processing [1]
and also regarding methods using artificial intelligence [2]. Especially for the latter case, neural-network-based algorithms
gained popularity in recent years [3]. However, these AI-based procedures require a lot of training data, which is expensive to
generate especially in an industrial setting due to the high variability in the objects scanned. On the other hand, many classical
algorithms rely on statistical assumptions or on hyperparameters, which have to be tuned empirically. For industrial use cases
this process needs to be repeated for each individual scanned part. A more general approach to denoising is considering the
wavelet transform, which has the advantage that one obtains a separation of the structural parts and the edges of a dataset, for
which several methods exist with a focus on thresholding the high pass wavelet coefficients [4]. But thresholding by some global
threshold often produces unfavorable results. Thus, we will consider more flexibility than simple thresholding.
This work proposes the utilization of an interactively trainable classifier for detecting and removing noise within the wavelet
decomposition in order to generate a denoised dataset.

2 Wavelet-based Denoising
The wavelet transform is a well-known mathematical integral transform which decomposes a signal (in our case a three-
dimensional image) with regards to both space and frequency. It allows for a direct combination of both a frequency separation
and distinguishing between directional information. The respective wavelet function defines the basis functions of the decom-
position. Mathematically, the wavelet transform of a square-integrable function f ∈ L2 (Rd ) using a wavelet ψ ∈ L2 (Rd ) at some
scale s and a direction u is defined as [5]
 
1 ∗ t −u
Z
W f (u, s) = f (t) √ ψ dt.
Rd s s
The accompanying concept of multiresolution analysis on several classes of wavelets describes, that approximations of the de-
composed functions are retrievable at a range of different resolution levels. Furthermore, the wavelet transform for a wavelet sat-
isfying a mild admissibility condition is also invertible, i.e., the original function f can be retrieved from the wavelet-transformed
representation. An example application of a three-dimensional wavelet transform is given in Figure 1, where the Haar wavelet [6]
was applied to a 3D volume of a motor piston as depicted in Figure 1a. The resulting cubic volume (outlined) is of the same size
as the original volume, but the information was rearranged in accordance to the wavelet function, which generates a sequence of
smaller and smaller subcubes containing mainly the edges of the objects contained at a specified number of resolution levels. As
a result, the leftmost and topmost subcube contains a lowpass-filtered volume which is basically a downsampled version of the
original volume. The other cubes contain mainly the edges of the object along different directions at different resolutions (i.e.,
highpass-filtered volume).
Fortunately, however, the highpass parts are typically very sparse, cf. the visually "empty" parts in Figure 1b. These parts
typically contain not only the edges but also the measurement noise. It is a well known fact, that modifications of the highpass
parts, such as thresholding to eliminate the measurement noise [4], affect the noisy parts in the function by smearing out noisy

Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License.

https://doi.org/10.58286/27718
12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.org

(a) The input dataset (b) The wavelet-transformed data

Figure 1: Decomposition of a volumetric dataset of a motor piston (a) after applying two steps of a 3D Haar wavelet transform
in (b).

regions. This makes these regions more homogenous, and thus effectively removes the noise. However, a global threshold
may have detrimental effects on several regions as it does not respect its local properties in any way. To address this issue, the
following approach uses an AI-based local classifier for deciding when a local noisy region in the highpass coefficient subcubes
shall be removed, i.e., set to zero.
In the first step, the volume is wavelet-transformed. For this purpose, we made use of the Haar/db1 wavelet and a biorthogonal
wavelet (labeled bior2.6 within the Python pywavelets module) in this work. Secondly, for each resolution level an approxi-
mation of the volume gradient at that scale is computed from the highpass coefficients [7]. Even though this approximation is
mathematically proven only for Haar wavelets, empirical experiments showed its validity also for the bior2.6 wavelet. In the
next step, a local classifier is trained interactively on the gradient approximation similarly to the procedure in [8]. During train-
ing, a slice of the gradient is rendered and the user is prompted to select positions where noise occurs (assigned some according
label), and also select positions of features/edges to keep untouched (which get assigned another label). Subsequently, around
each marked position, features are extracted from a local environment of a configurable size (e.g., a 3 × 3 × 3 neighborhood)
centered around that position. For our purpose, the feature vector simply consists of the highpass coefficients within the envi-
ronment. Optionally, further textural or geometric features may be computed and used instead. The classifier is then trained to
classify these regions either as noisy or as non-noisy. For the choice of the classifier we used a single artificial neuron because
its training and application can be achieved in a particularly fast way. The parametrization of this neuron involved elastic net
regularization with an ℓ1 ratio of 0.25 and a z-scaling normalization before feeding the values to the classifier, all as specified
within scikit-learn [9]. We also considered Support Vector Machines and a small artificial neural network, but rejected these
two alternatives as no improvements were achieved when using them. Having a trained classifier at hand, the algorithm iterates
over local regions of the gradient approximation at each scale. For each position, the local region is classified as either noisy or
non-noisy. In case the noisy label is assigned, the highpass coefficient at that position is set to zero in each of the seven subcubes,
effectively removing the noise information from the wavelet decomposition. Finally, the inverse wavelet transform is applied to
return the denoised volume.
Figure 2 depicts the proposed interactive training procedure on the motor piston dataset shown in Figure 1a. A slice of the
original 3D dataset is shown in Figure 2d (left part). Applying the interactive classification to remove noise from the gradient
approximations (cf. Figures 2a to 2c) detects the positions in the highpass coefficients where noise is contained. Note the incre-
mental increase in resolution during the interactive training procedure. After setting these coefficients to zero and reconstructing
the volume from the modified wavelet decomposition produces the denoised volume as shown in Figure 2d (right part). In this
image it can be seen that the noisy background was made considerably more homogeneous, but also that the noisy interior of the
piston dataset was smoothed out. The edges were kept largely intact while introducing some minor discretization artefacts.

3 Results
The aforementioned procedure generates denoised datasets where noisy parts are homogenized which allow for easier processing
while also keeping object edges as sharp as possible. In this section the proposed interactive denoising method is applied on six
different scans as depicted in Figure 3. The first scan (cf. Figure 3a) is a relatively low-contrast scan of a human brain obtained
at the European Synchrotron Radiation Facility [11]. The data shown in Figure 3b depicts a slice of a bread bun integrating a

Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License.
12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.org

(a) Iteration #1 (b) Iteration #2

(c) Iteration #3 (d) Before and after denoising

Figure 2: Interactive denoising on a motor piston dataset using three wavelet decomposition levels applied to all resolution level,
according to the wavelet subcubes, from coarse (a) to fine (c) using interactive annotation (blue markers for noise, red for object
points), leading to a classification result in the right parts of these images. The noisy input is shown in (d) (left), the denoised
image to its right.

lot of fine structures while also exhibiting considerable noise in the background. Figure 3c features a synthetic dataset where 50
percent of Gaussian random noise was added. Additionally, we also considered an actual CT scan of a motor piston as shown in
Figure 2d on the left. The remaining two images show industrial CT scans of an electric motor and a Region of Interest from a
CT scan of a car. In all images contrast was increased to enhance the visibility of the noise.

(a) A human brain (b) A bread bun (c) Noisy synthetic data (d) Electric motor (e) RoI from a car scan

Figure 3: The datasets considered in this study.

For each of these scans, the proposed interactive denoising method was performed. Specifically, for each scan, two different
wavelets were considered, where for each wavelet we repeated the experiment using one, two, or three wavelet levels, i.e., that
many wavelet decomposition levels were generated and an interactive model was trained on the gradient approximation for
each level independently. Within each such configuration of a wavelet and a number of decomposition levels, three different
environment sizes (3 × 3 × 3, 5 × 5 × 5, and 7 × 7 × 7) were used. For all these environments, we tracked both the runtime of our
algorithm (excluding the interactive part). Additionally, the results are compared to state-of-the-art methods, namely common
wavelet thresholding [4], TV-denoising according to Chambolle [10], and non-local means (NLM) denoising [1].
All results are listed in Tables 1 and 2. As marked in these tables, the best method was in most cases the TV-thresholding, as it
minimizes the TV norm and we specifically measure the TV norm quotient. However, empirically the proposed method performs
an equivalently good denoising and is often much faster. Additionally, our denoising methodology leaves the object information,
represented by the "edge information" in the wavelet subcubes, untouched. Thus, the object is still mostly unaffected while the
noise is reduced. Consequently, the grayscale value difference between the object and the noise in the reconstructed volume
decreases only slightly, as does the TV norm, which is evidenced by the results in Table 1 and 2. Interestingly, the best results
using the proposed method are produced when using the Haar wavelet. This may be attributed to the fact that the gradient

Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License.
12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.org

approximation was only proven for the Haar wavelet, not general wavelets.
A notable exception where the classifier-based wavelet denoising produces superior results is the CT scan of a motor piston and a
synthetic specimen generated from that piston. There, the combination of noise and artefacts heavily affect both the surrounding
air but also the interior of the object. Thus, large portions of the entire scan are homogenized (cf. Figure 2d) and thus the TV
norm decreases.
Naturally, a much more explorative quantitative study of the denoising behavior would be of great interest. However, as evaluating
interactive algorithms is particularly hard, we consider this out of scope for this work but a topic for our future research. This
may also include the influence of the denoising on dimensional metrology, where we suspect that dimensional measurements
results might deteriorate, mainly due to discretization artefacts as introduced by the block-like wavelet transform and subsequent
pointwise deletion of noise points.
In order to give a few qualitative impressions of the interactive denoising, we consider the following scans in the sequel:

1. A human brain, as low-contrast scans might get worse by denoising, especially at the borders between different regions.
2. An electric motor, containing many fine structures but also considerable noise.
3. A Region of Interest of a car CT scan, for a realistic application scenario in the industrial domain.
For each scan, we applied the proposed interactive denoising technique using local environments of size 3 × 3 × 3, using three
wavelet deomposition levels, and performing the decomposition with the Haar wavelet. Note that even though we only display
slices, the denoising method actually works in true 3D.
Starting with the scan of a human brain, the central slice of the input data is shown in Figure 4a. After the wavelet decomposition,
the noisy parts were marked interactively and assigned a respective label. Similarly, the brain regions were marked in the
wavelet gradient containing the noise and only the borders of the brain lobes (as expected from a gradient). After the voxelwise
classification in the gradient, a volumetric dataset was reconstructed from the modified wavelet representations. The central slice
of the generated volume is shown in Figure 4b. From a visual inspection of the generated result, the noise outside the organ has
changed. However, the pointwise difference image (cf. Figure 4c) clearly shows that also the interior regions were homogenized,
where no gradient from white matter to gray matter occured.
Next, the denoising of an industrial CT scan of an electric motor was performed. The original data is depicted in Figure 5a as
it was scanned at the ESRF in Grenoble. As both the denoised image (cf. Figure 5b) and the difference image (cf. Figure 5c)
indicate, the noisy regions are affected and homogenized as desired, while the edges and the industrial structures are well-
preserved. Unfortunately, there are also some pixelization artefacts introduced, which are generated most likely due to more
aggressive noise deletion in these regions within the gradient.
Finally, we pursue the interactive denoising of a Region of Interest taken from a computed tomography scan of an entire car,
see Figure 6a. The difficulty there lies in the high levels of measurement noise and the presence of artefacts due to many metal
components. Even in this case, our method was able to successfully reduce the measurement noise considerably while leaving
the regions of the car components untouched (cf. Figure 6b) which is visually confirmed by the difference image in Figure 6c.

4 Conclusion
In this work, a novel denoising procedure was proposed, specifically targeted towards three-dimensional voxel datasets as they
are generated in X-ray computed tomography. The combination of a hierarchical wavelet transform and a locally trained classifier
instead of a regular threshold applied to the wavelet representation, proved to be highly efficient in terms of runtime and simplicity
in usage. Furthermore, on selected datasets it outperformed or came close to three state-of-the-art methods at the cost of requiring
interactive user input.

Acknowledgements
We would like to thank Christoph Heinzl for proofreading this paper.

References
[1] A. Buades, B. Coll, J. Morel, A review of image denoising algorithms, with a new one, Multiscale Modeling and
Simulation: A SIAM Interdisciplinary Journal, 4(2), 2005, pp. 490-530
[2] P. Kaur, G. Singh, P. Kaur, A Review of Denoising Medical Images Using Machine Learning Approaches, Current
Medical Imaging Reviews 14, 2018, pp. 675-685
[3] C. Tian, L. Fei, W. Zheng, et al, Deep learning on image denoising: An overview, Neural Networks 131, 2020, 251-275
[4] S. Ruikar, D. Doye, Wavelet Based Image Denoising Technique, Int J Adv Comput Sci Appl 2(3), 2011, pp. 49-53
[5] S. Mallat, A Wavelet Tour of Signal Processing: The Sparse Way, Academic Press, 1999.
[6] A. Haar, Zur Theorie der orthogonalen Funktionensysteme, Math Ann 69, 1910, pp. 331–371.

Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License.
12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.org

[7] T. Sauer, M. Stock, Haar Wavelets, Gradients and Approximate Total Variation Regularization, arXiv:2208.05389, 2022.
[8] T. Lang, AI-Supported Interactive Segmentation of 3D Volumes, PhD Thesis, University of Passau, 2021.
[9] F. Pedregosa et al, Scikit-learn: Machine Learning in Python, JMLR 12, 2011, pp. 2825-2830
[10] A. Chambolle, An algorithm for total variation minimization and applications, J Math Imaging Vis (20), 2004, pp. 89-97.
[11] C. Walsh, P. Tafforeau, W. Wagner, et al, Imaging intact human organs with local resolution of cellular structures using
hierarchical phase-contrast tomography, Nat Methods, 2021

(a) Input data (b) Denoised data (c) Difference image

Figure 4: Denoising the "brain" dataset, displaying the central slice along one axis.

(a) Input data (b) Denoised data (c) Difference image

Figure 5: Denoising the "electric motor" dataset, displaying the central slice along one axis.

(a) Input data (b) Denoised data (c) Difference image

Figure 6: Denoising the "car" dataset, displaying the central slice along one axis.

Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License.
12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.org

Wav. Thresh. TV NLM

Scan Wavelet Levels K TVQ Time [s]
TVQ Time [s] TVQ Time [s] TVQ Time [s]
3 0.777 131.10
1 5 0.762 201.05
7 0.742 344.96
3 0.753 131.42
Haar 2 5 0.788 215.85
7 0.741 413.06
3 0.758 125.79
3 5 0.746 189.60
7 0.707 344.34
Brain 0.662 10.86 0.268 1,093.06 0.995 18,445.06
3 0.824 156.62
1 5 0.947 221.72
7 0.871 407.56
3 0.938 167.97
bior2.6 2 5 0.910 242.42
7 0.872 477.40
3 0.916 150.19
3 5 0.815 218.18
7 0.930 377.31
3 0.939 17.07
1 5 0.918 24.68
7 0.916 41.90
3 0.924 8.96
Haar 2 5 0.918 14.99
7 0.911 28.45
3 0.936 20.29
3 5 0.941 32.40
7 0.929 53.98
Bread bun 0.973 2.16 0.752 9.94 1.000 4,532.86
3 0.930 19.78
1 5 0.910 28.12
7 0.910 47.36
3 0.915 10.79
bior2.6 2 5 0.907 17.60
7 0.895 45.06
3 0.908 25.96
3 5 0.901 37.99
7 0.894 63.70
3 0.300 37.96
1 5 0.294 76.49
7 0.316 115.62
3 0.182 44.95
Haar 2 5 0.168 71.19
7 0.192 126.01
3 0.114 48.79
3 5 0.144 72.63
7 0.180 132.14
Noisy piston 0.218 6.76 0.769 34.59 1.000 281.90
3 0.375 46.92
1 5 0.404 65.46
7 0.467 111.68
3 0.415 55.64
bior2.6 2 5 0.415 85.33
7 0.539 150.72
3 0.694 59.58
3 5 0.451 83.77
7 0.449 144.14
Table 1: TV quotient results after denoising for several wavelets and different environment sizes of K × K × K.

Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License.
12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.org

Wav. Thresh. TV NLM

Scan Wavelet Levels K TVQ Time [s]
TVQ Time [s] TVQ Time [s] TVQ Time [s]
3 0.547 46.90
1 5 0.562 68.70
7 0.573 122.10
3 0.380 53.48
Haar 2 5 0.381 78.92
7 0.430 126.19
3 0.350 39.87
3 5 0.365 62.40
7 0.405 109.97
Piston 0.903 4.41 0.511 118.33 1.000 311.95
3 0.787 50.99
1 5 0.675 78.99
7 0.663 136.78
3 0.671 55.28
bior2.6 2 5 0.674 78.30
7 0.656 141.09
3 0.749 46.80
3 5 0.660 74.05
7 0.654 123.30
3 0.947 112.70
1 5 0.945 149.15
7 0.949 263.32
3 0.949 118.37
Haar 2 5 0.972 179.79
7 0.955 287.46
3 0.991 93.13
3 5 1.059 149.12
7 1.077 250.93
Electric motor 1.000 24.40 0.327 1,179.82 0.967 7,945.28
3 0.976 120.80
1 5 1.014 174.19
7 1.015 427.90
3 0.996 125.37
bior2.6 2 5 1.024 193.78
7 1.036 471.94
3 1.109 120.48
3 5 1.109 162.63
7 1.094 407.11
3 0.984 59.21
1 5 0.979 99.07
7 0.976 419.03
3 1.007 143.85
Haar 2 5 1.006 232.33
7 0.999 415.95
3 0.990 142.94
3 5 1.040 214.81
7 1.026 379.71
Car RoI 0.983 14.48 0.930 61.96 1.000 22,604.38
3 1.031 176.51
1 5 1.026 225.41
7 1.026 405.02
3 1.039 181.30
bior2.6 2 5 1.014 249.08
7 1.010 481.89
3 1.044 189.16
3 5 1.029 272.12
7 1.004 394.12
Table 2: TV quotient results (continued).

Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License.