U-WNO:U-Net-enhanced Wavelet Neural Operator for fetal head segmentation

arXiv:2411.16890 · eess.IV, cs.CV · Submitted 2024-11-25 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "U-WNO:U-Net-enhanced Wavelet Neural Operator for fetal head segmentation".

Dev: Here's the gist of the paper. This article describes the development of a novel U-Net-enhanced Wavelet Neural Operator (UWNO), which combines wavelet decomposition, operator learning, and an encoder-decoder mechanism to perform accurate fetal head segmentation in two-dimensional ultrasound images.

Rosa: Model Foundation The U-WNO is a combination of operator learning, Convolutional Neural Network (ConvNet), and the encoder-decoder mechanism.

Dev: Operator learning focuses on learning the mapping function of an operator that transforms one function into another.

Rosa: Some noteworthy neural operators include the Wavelet Neural Operator (WNO) (Tripura and Chakraborty, 2022) and Fourier Neural Operator (FNO) (Li et al, 2021; Wen et al, 2022).

Title and authors: Rosa: So the title itself tells you right away what it’s about, using the U-Net enhanced Wavelet Neural Operator for fetal head segmentation. It suggests a specific combination of techniques is key to getting that accurate result.

Dev: That combination means they aren't just using a standard deep learning model; they are specifically leveraging wavelet transforms to look at the data in different scales, which is important for ultrasound because those images have a lot of fine detail we need to capture.

Taro: When you combine wavelet decomposition with neural operators, you're essentially trying to teach the AI how to process information across different resolutions simultaneously, which should help it see things that are either really big features or really small details in the ultrasound scan.

Rosa: Exactly, and they are using this framework because it lets them learn those complex functional mappings much more effectively than traditional methods, which is what operator learning is all about.

The paper's summary: Dev: Okay, so the summary lays out the core idea: they’re developing a novel U-Net enhanced Wavelet Neural Operator to perform accurate fetal head segmentation in two-dimensional ultrasound images. It frames it as a way to bridge theoretical advancements with practical applications in medical imaging.

Rosa: They explain that this approach uses wavelet transforms for time frequency localization, and then combines down-sampling and up-sampling operations within the network structure to generate the segmentation map. This is how they achieve that accurate tracking of patterns in both the spatial domain and by learning those functional mappings.

Taro: The paper highlights how this integration allows the model to handle multi-scale data much better, which is crucial when you're dealing with complex structures like a fetal head where you have features at many different sizes.

Dev: And they mention that the neural network uses these wavelet-based representations specifically to learn the underlying operators that map inputs to outputs, which enhances its ability to model and predict those complex systems with high accuracy.

The paper's improvements: Rosa: Now, looking at what they improved, one big point is how they enhanced the concept of neural operators by leveraging wavelet transforms to capture complex patterns in both spatial and temporal data, which is a significant step for modeling these kinds of problems.

Dev: They also talk about the kernel acting as a convolutional operator specifically in the wavelet domain, represented by that mathematical expression K psi u x equals W minus one K dot W u x, which shows how they've adapted the standard operator idea to work with wavelets.

Taro: And from an autonomy researcher's view, it sounds like this structure gives them a way to handle those infinite-dimensional feature vector spaces that can pop up in these kinds of complex imaging problems without getting lost in the noise.

Rosa: Beyond the math, they used a specific dataset called the two-dimensional USG dataset, which has approximately nine hundred seventy-six images and their corresponding masks for fetal head segmentation <ref:2411.16890#pg3,approximately 976 images and their corresponding masks>. They resized all those images to a shape of one hundred twenty-eight by one hundred twenty-eight pixels for training <ref:2411.16890#pg3>.

Conclusion: Dev: So to wrap up, the U-WNO framework is essentially this combination of wavelet decomposition, operator learning, and the encoder-decoder structure designed specifically for fetal head segmentation using ultrasound images. It’s a method that combines frequency domain analysis with neural network architectures.

Rosa: The main result they reported was a maximum Dice Score of approximately zero point six five on their test set, which the authors suggest shows strong performance for this segmentation task. That score is what we use to measure how well the model is separating the fetal head from everything else in the image.

Taro: And when you look at that result, it suggests that even with this complex setup, they managed to get a solid score, which means the method actually works quite well for tracking those patterns. But we should remember what they flagged: their limitation is dealing with data that might be significantly noisy or when the images are resized to smaller dimensions due to computational constraints.

Dev: That makes sense from an engineering standpoint; if you're working in a real-time loop, you have to worry about those noise issues and how the system behaves when the input data isn't perfect. So, they’re focused on making sure it handles those real-world imperfections gracefully.

Rosa: Overall, this paper shows a way to use these advanced mathematical tools to get more precise segmentation in ultrasound. It opens up pathways for using AI in clinical settings to assist healthcare professionals by accurately tracking fetal development during routine scans.

Taro: It also serves as a good educational tool, allowing trainees to see how complex images are handled by an AI model, which is valuable for understanding the interpretation process.

Dev: So we’ve talked about the U-WNO:U-Net enhanced Wavelet Neural Operator for fetal head segmentation. It’s a method that uses wavelets to handle multi-scale data, and it shows promise in getting better results on fetal head segmentation in ultrasound.

Pranava Seth, Deepak Mishra, Veena Iyer

Thapar Institute of Engineering & Technology · Indian Institute of Space Science & Technology · Indian Institute of Public Health

eess.IV, cs.CV

Submitted: 2024-11-25

Updated: 2026-11-09

Comments: The authors have decided to withdraw this preprint as the work is no longer being actively pursued in its current form.

Code: https://github.com/pranava1709/U-WNO

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 76/100

The gist: This article describes the development of a novel U-Net-enhanced Wavelet Neural Operator (UWNO), which combines wavelet decomposition, operator learning…

Key concepts

Wavelet Neural Operator (WNO)
This concept uses wavelet transforms within a neural network framework to learn complex mappings between different functions. Wavelets are mathematical tools that allow for multiresolution analysis, meaning the model can examine image details at various scales simultaneously, which is crucial for identifying subtle features like fetal head structures.
Operator Learning
Operator learning focuses on training a neural network to learn the mathematical function that transforms one input function into another. In this context, it means teaching the AI how to map an ultrasound image directly to its corresponding segmentation mask, rather than just classifying pixels sequentially.
U-Net Architecture
The U-Net is a specific encoder-decoder neural network structure widely used for image segmentation. It works by progressively reducing the spatial resolution of an input image through an encoder and then using upsampling layers in a decoder to reconstruct the output, allowing it to produce precise boundary maps for the fetal head.
Dice Score (DS)
The Dice Score is a primary metric used to evaluate how well the model has segmented the fetal head. It measures the overlap between the predicted segmentation and the ground truth mask by comparing true positives, false positives, and false negatives. A higher score indicates better accuracy in identifying where the fetal head is located.

Terminology

Summary

This article describes the development of a novel U-Net-enhanced Wavelet Neural Operator (UWNO), which combines wavelet decomposition, operator learning, and an encoder-decoder mechanism to perform accurate fetal head segmentation in two-dimensional ultrasound images. p. 0

Where the paper says it

org; Abstract This article describes the development of a novel U-Net-enhanced Wavelet Neural Operator (UWNO),which combines wavelet decomposition, operator learning, and an encoder-decoder mechanism. This approach harnesses the superiority of the wavelets in time frequency localization of the functions, and the combine down-sampling and up-sampling operations to generate the segmentation map to enable accurate tracking of patterns in spatial domain and effective learning of the functional mappings to perform regional segmentation.

Model Foundation

The U-WNO is a combination of operator learning, Convolutional Neural Network (ConvNet), and the encoder-decoder mechanism. Operator learning focuses on learning the mapping function of an operator that transforms one function into another. p. 0

Where the paper says it

The model is a combination of operator learning, Convolutional Neural Network (ConvNet), and the encoder-decoder mechanism. Operator learning focuses on learning the mapping function of an operator that transforms one function into another.

Some noteworthy neural operators include the Wavelet Neural Operator (WNO) (Tripura and Chakraborty, 2022) and Fourier Neural Operator (FNO) (Li et al, 2021; Wen et al, 2022). p. 0
Where the paper says it

Operator learning focuses on learning the mapping function of an operator that transforms one function into another. Some noteworthy neural operators include the Wavelet Neural Operator (WNO) (Tripura and Chakraborty, 2022), Fourier Neural Operator (FNO) (Li et al, 2021; Wen et al, 2022) and a few others, designed to efficiently approximate complex functions, particularly for solving 1

WNO combines the principles of wavelet transforms with neural network architecture to efficiently learn and approximate complex functions. p. 1
Where the paper says it

By operating on the frequency components of the data, FNO can learn and approximate operators that map inputs to outputs with high accuracy and reduced computational cost. WNO combines the principles of wavelet transforms with neural network architecture to efficiently learn and approximate complex functions, particularly in the context of solving partial differential equations (PDEs) and other mathematical problems.

Data Preparation

The 2D - USG dataset is used as the main training dataset for U-WNO. The image dataset comprises ultrasound images sourced from HC18 challenge. p. 2

Where the paper says it

2 Dataset The 2D - USG dataset is used as the main training dataset for U-WNO. The image dataset comprises ultrasound images sourced from HC18 challenge1 .

Each image has been resized to the shape of 128x128 pixels and is stored in PNG format. p. 2
Where the paper says it

The image dataset comprises ultrasound images sourced from HC18 challenge1 . Each image has been resized to the shape of 128*128 pixels and is stored in PNG format.

The dataset includes approximately 976 images and their corresponding masks, which were converted into filled masks, with the fetal head designated as the target label. p. 2
Where the paper says it

The dataset includes two-dimensional ultrasound scans from various stages of pregnancy across different trimesters. It consists of approximately 976 images and their corresponding masks, which were converted into filled masks, with the fetal head designated as the target label, as shown in Figure 1.

Architecture Details

The U-WNO combines wavelet-based feature extraction, ConvNet, and U-Net into a unified architecture. The WNO framework enhances the concept of neural operators by leveraging wavelet transforms to capture complex, multi-scale patterns in both spatial and temporal data. p. 2

Where the paper says it

operator is particularly effective for regional segmentation tasks, as illustrated in the subsequent steps. The WNO framework enhances the concept of neural operators by leveraging wavelet transforms to capture complex, multi-scale patterns in both spatial and temporal data.

The kernel acts as a convolutional operator in the wavelet domain, facilitated by the mathematical representation Kψ(u)(x) = W−1 (K · W(u))(x). p. 3
Where the paper says it

the kernel acts as a convolutional operator in the wavelet domain, facilitating efficient, localized learning at varying scales. Mathematically, this is represented by the kernel application on wavelet-transformed coefficients: Kψ(u)(x) = W−1 (K · W(u))(x), where W denotes the wavelet transform and W−1 its inverse.

The encoder progressively reduces spatial dimensions while increasing feature channels through convolutional layers and max-pooling operations. p. 3
Where the paper says it

The encoder, or contracting path, progressively reduces the spatial dimensions of the input image while increasing the number of feature channels. This is achieved through a series of convolutional layers and max-pooling operations.

The decoder reconstructs the spatial dimensions using upsampling, concatenation with the encoder's feature map, and convolutional layers.

Performance Metrics

The Dice Score (DS) is used as the primary metric to assess model performance. The Dice score is calculated based on True Positives (TP), False Positives (FP), and False Negatives (FN) using the formula DS = 2 · TP / (2 · TP + FP + FN). p. 4

Where the paper says it

In this context, each pixel is classified as either part of the object (foreground) or background. The Dice score is calculated based on True Positives (TP), False Positives (FP), and False Negatives (FN) as follows: DS = 2 · TP 2 · TP + FP + FN where: • TP (True Positives) is the number of correctly predicted positive pixels.

The model achieved a maximum DS of approximately 0. p. 4
Where the paper says it

• FN (False Negatives) is the number of positive pixels that are not predicted as positive. Here’s a refined version with enhanced grammar, structure, and readability: We achieved a maximum DS of approximately 0.

65, indicating strong performance in the segmentation task. p. 4
Where the paper says it

Here’s a refined version with enhanced grammar, structure, and readability: We achieved a maximum DS of approximately 0. 65, indicating that integrating the Wavelet Neural Operator (WNO) with U-Net as a unified operator yields strong performance in the segmentation task.

Clinical and Educational Impact

U-WNO’s application in ultrasound imaging could greatly impact clinical practice by assisting healthcare professionals in monitoring fetal development. Accurate segmentation also serves as a valuable educational tool, teaching healthcare professionals to interpret ultrasound images. p. 5

Where the paper says it

U-WNO’s deployment, following thorough testing, has the potential to reduce human error and increase efficiency in ultrasound procedures within clinical settings. Accurate segmentation also serves as a valuable educational tool, teaching healthcare professionals to interpret ultrasound images, particularly in challenging cases like late-stage pregnancy.

The model’s ability to effectively segment images under challenging conditions opens pathways for its adoption in multiple fields. p. 4
Where the paper says it

Its ability to effectively segment images under challenging conditions opens pathways for its adoption in multiple fields. Future work could focus on further model refinement and exploring its adaptability to diverse data types, enhancing its practical utility.

Future Directions

Future work could focus on further model refinement and exploring its adaptability to diverse data types, enhancing its practical utility. The operator’s design is particularly noteworthy for managing infinite-dimensional feature vector spaces. p. 5

Where the paper says it

As AI evolves, models like U-WNO illustrate how advanced computational techniques can address real-world challenges. The operator’s design is particularly noteworthy for managing infinite-dimensional feature vector spaces, a challenging aspect in computer vision.

This amalgamation of techniques also paves the way for future technologies, such as digital twin systems. p. 5
Where the paper says it

Integrating wavelet decomposition with deep learning techniques enhances feature extraction and representation. This amalgamation of techniques also paves the way for future technologies, such as digital twin systems.

References

Dias T, Sairam S, Kumarasiri S (2014) Ultrasound diagnosis of fetal renal abnormalities. Best Pract Res Clin Obstet Gynaecol 28. p. 5

Where the paper says it

References Dias T, Sairam S, Kumarasiri S (2014) Ultrasound diagnosis of fetal renal abnormalities. Best Pract Res Clin Obstet Gynaecol 28.

https://doi. p. 5
Where the paper says it

Best Pract Res Clin Obstet Gynaecol 28. https://doi.

p. 5
Where the paper says it

Best Pract Res Clin Obstet Gynaecol 28. https://doi.

p. 5
Where the paper says it

Best Pract Res Clin Obstet Gynaecol 28. https://doi.

org/10.1016/j.bpobgyn.2014.01.009

Fiorentino MC, Villani F, di Cosmo M, et al (2022) A review on deep-learning algorithms for fetal ultrasound-image analysis. Med Image Anal 83:102629. p. 6

Where the paper says it

Fiorentino MC, Villani F, di Cosmo M, et al (2022) A review on deep-learning algorithms for fetal ultrasound-image analysis. Med Image Anal 83:102629.

https://doi.org/10.1016/j.media.2022.102629

Li Z, Kovachki N, Azizzadenesheli K, et al (2021) Fourier neural operator for parametric partial differential equations. Available at: urlhttps://arxiv. p. 6

Where the paper says it

102629 Li Z, Kovachki N, Azizzadenesheli K, et al (2021) Fourier neural operator for parametric partial differential equations. Available at: urlhttps://arxiv.

p. 6
Where the paper says it

102629 Li Z, Kovachki N, Azizzadenesheli K, et al (2021) Fourier neural operator for parametric partial differential equations. Available at: urlhttps://arxiv.

p. 6
Where the paper says it

102629 Li Z, Kovachki N, Azizzadenesheli K, et al (2021) Fourier neural operator for parametric partial differential equations. Available at: urlhttps://arxiv.

p. 6
Where the paper says it

102629 Li Z, Kovachki N, Azizzadenesheli K, et al (2021) Fourier neural operator for parametric partial differential equations. Available at: urlhttps://arxiv.

org/abs/2010. p. 6
Where the paper says it

Available at: urlhttps://arxiv. org/abs/2010.

08895

Sobhaninia Z, Rafiei S, Emami A, et al (2019) Fetal ultrasound image segmentation for measuring biometric parameters using multi-task deep learning. Available at: urlhttps://arxiv.org/abs/1909. p. 6

Where the paper says it

Available at: urlhttps://arxiv. org/abs/1909.

00273

Tripura T, Chakraborty S (2022) Wavelet neural operator: A neural operator for parametric partial differential equations. Available at: urlhttps://arxiv.org/abs/2205. p. 6

Where the paper says it

Available at: urlhttps://arxiv. org/abs/2205.

02191

Wells PNT (2006) Ultrasound imaging. Phys Med Biol 51(13):R83. p. 6

Where the paper says it

02191 Wells PNT (2006) Ultrasound imaging. Phys Med Biol 51(13):R83.

https://doi.org/10.1088/0031-9155/51/13/R06, URL https://dx. p. 6
Where the paper says it

https://doi. 1088/0031- 9155/51/13/R06, URL https://dx.

doi.org/10.

Wells PNT (2006) Ultrasound imaging

Wen G, Li Z, Azizzadenesheli K, et al (2022) U-FNO – an enhanced fourier neural operatorbased deep-learning model for multiphase flow. Available at: urlhttps://arxiv.org/abs/2109. p. 6

Where the paper says it

Available at: urlhttps://arxiv. org/abs/2109.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved fetal head segmentation accuracy by integrating wavelet decomposition, which enables multiresolution analysis of ultrasound images to isolate critical features missed by conventional methods.

  2. Enhanced model robustness against challenging imaging conditions such as resizing images to smaller dimensions due to computational constraints and handling significant noise in the data, resulting in a maximum Dice Score of approximately 0.65 on the test set.

  3. Increased efficiency in clinical settings by providing a tool for healthcare professionals to accurately track and label key anatomical features during routine scans, thereby reducing human error and improving operational efficiencies.

  4. Enabled dual functionality serving both clinical and educational needs, allowing the system to support trainee education by helping them navigate complex images and identify regions of interest with greater confidence.

Abstract

This article describes the development of a novel U-Net-enhanced Wavelet Neural Operator (U-WNO),which combines wavelet decomposition, operator learning, and an encoder-decoder mechanism. This approach harnesses the superiority of the wavelets in time frequency localization of the functions, and the combine down-sampling and up-sampling operations to generate the segmentation map to enable accurate tracking of patterns in spatial domain and effective learning of the functional mappings to perform regional segmentation. By bridging the gap between theoretical advancements and practical applications, the U-WNO holds potential for significant impact in multiple science and industrial fields, facilitating more accurate decision-making and improved operational efficiencies. The operator is demonstrated for different pregnancy trimesters, utilizing two-dimensional ultrasound images.

Sources

Related papers