APTISI Transactions on Technopreneurship (ATT)
Vol. 7, No. 3, November 2025, pp. 713−725
E-ISSN: 2656-8888 | P-ISSN: 2655-8807, DOI:10.34306

❒

713

Technopreneurial Filtering Technique for Speckle Noise
Reduction in Ultrasound Imaging of Polycystic
Ovary Syndrome
Pratibha Pandey1

, Sumit Chaudhary2

, Xin Nie3*

1, 2 Department of Computer Science and Engineering, Uttaranchal University, India
3 School of Computer Science and Engineering, Wuhan Institute of Technology, China
1 pratibhapandey8502@gmail.com, 2 drchaudharysumit@gmail.com, 3 niexin@whu.edu.cn

*Corresponding Author

Article Info

ABSTRACT

Article history:

Polycystic Ovary Syndrome (PCOS) is a common endocrine disorder in
women that is frequently diagnosed by ultrasound imaging by examining ovarian abnormality. Ultrasound images, though, are normally contaminated with
speckle noise that degrades image quality and poses difficulties for diagnosis.
Conventional denoising methods like mean and median filtering are unable to
eliminate noise properly while maintaining fine details. To solve this problem,
this paper introduces a new Attention-based Autoencoder (AAE) for denoising
PCOS ultrasound images. The model uses an attention mechanism to selectively
amplify significant image areas and suppress noise, enhancing image quality for
diagnosis. The introduced method was tested on a publicly available ultrasound
dataset with synthetic speckle noise at various levels (variance 0.5, 0.02, and
0.001). Experimental results prove that the suggested method performs better
than conventional denoising methods, with peak signal-to-noise ratio (PSNR)
values of 31.33, 34.25, and 36.23, respectively. The structural similarity index
measure (SSIM) also reveals notable improvements and corresponding scores
of 85.21, 92.33, and 99.25. Beyond technical performance, this work supports
the development of scalable, AI-driven PCOS diagnostic tools within a technopreneurship incubator model. These results indicate that AAEs can improve
ultrasound image quality, facilitating more accurate PCOS diagnosis.

Submission June 15, 2025
Revised July 4, 2025
Accepted September 7, 2025
Published September 24, 2025
Keywords:
Attention Module
Speckle Noise
Noise Reduction
Polycystic Ovary Syndrome

This is an open access article under the CC BY 4.0 license.

DOI: https://doi.org/10.34306/att.v7i3.767
This is an open-access article under the CC-BY license (https://creativecommons.org/licenses/by/4.0/)
©Authors retain all copyrights

1.

INTRODUCTION
The World Health Organization (WHO) reports that Polycystic Ovary Syndrome (PCOS) affects
around 8–13% of women around the world, making it a common health problem for women [1, 2]. There
are different symptoms that occur during PCOS, such as trouble getting pregnant, irregular periods, and excess
hair growth, etc. This is caused by small cysts in the ovaries, high testosterone levels, and insulin problems,
which can disrupt normal ovulation and overall health. It also affects other health-related conditions and causes
different diseases like heart disease, diabetes, and some types of cancer. So early identification of this disease
is mainly mandatory to prevent severe complications [3, 4].
For effective detection of this disease, all depends upon various methods such as blood tests, patient
physical body examinations, Magnetic Resonance Imaging (MRI), and ultrasound methods. Among all these,
Journal homepage: https://att.aptisi.or.id/index.php/att

714

❒

E-ISSN: 2656-8888 | P-ISSN: 2655-8807

ultrasound methods have become more popular for their non-invasive nature and accurately detecting key components from ovaries. But many times, these images are prone to different types of noise, such as shadow,
Gaussian, and speckle noise, which can make it challenging to accurately identify the disease from ovarian
ultrasound images [5–7]. By the shadow noise, it looks like dark areas that block certain regions of the data. In
contrast, Gaussian noise compromised the clarity of the image due to its blurry effect, and lastly, speckle noise
creates a smaller pattern of noise, which makes it difficult to differentiate small structures like cysts from noise.
These interferences lead to lowering the performance of the predicting models [8].
In recent years, many methods have been proposed for filtering ultrasound images, such as mean,
median, Gaussian, and adaptive filtering [9]. Although these methods remove the noise, they also remove
important features from the image, and these methods face difficulties in high levels of noise scenarios. As
these methods are heavily based on their parameters like kernel function, filter types, number of levels, etc.,
which leads to difficulty removing noise from the image effectively [10–12]. The Gaussian-based filtering
method is often used for smoothing the image, and sometimes it removes important components from the
image, and adaptive filtering is not so much used in practical applications as it does not effectively utilize
higher-order statistics from the input data. So, these methods have their own limitations, which affect the
effectiveness of filtering the ultrasound image accurately [13–16]. Also, another researchers [17] suggested a
self-supervised approach for eliminating noise from the ultrasound image.
There is a need for scalable, intelligent solutions that can improve medical imaging pipelines due to the
expanding convergence of AI and healthcare innovation, particularly under the purview of technopreneurship
incubator models [18]. This goal is supported by the proposed work, which provides a flexible framework that
can be incorporated into clinical decision-support tools and diagnostic startup ecosystems.
This paper addresses these problems that occur in the current methods by proposing an Attentionbased Autoencoder (AAE) framework [19, 20]. This framework effectively eliminates noise by preserving the
essential features of images. In this, the model is trained with different noise levels, which makes it more robust
and generalized in the context of noise elimination from ultrasound images. In this, the attention module gives
flexibility to pay attention to the most relevant features during noise elimination, which makes the denoised
version of the image more useful. The primary objective of this research is to develop an AAE for denoising
PCOS ultrasound images, ensuring improved image quality and diagnostic accuracy. The key goals of this
study are [21, 22]:
• Develop an autoencoder framework with attention mechanisms to effectively remove speckle noise from
PCOS ultrasound images [23].
• Ensure that critical anatomical features, such as ovarian cysts and follicular structures, remain intact after
noise removal [24].
• Assess the proposed model’s robustness by applying synthetic speckle noise at different variance levels
(0.5, 0.02, and 0.001) and comparing results with traditional denoising methods [25].
The rest of this paper is organized as follows: Section 2. presents a recent study on different filtering
techniques. Section 3. details the methodology, including the proposed method. Section 4. discusses the
experimental results of denoising models. Finally, Section 6. offers concluding remarks on the paper’s findings.
2.

LITERATURE REVIEW
Research on PCOS ultrasound images often focuses more on classification than on noise elimination
[26]. However, noise, particularly speckle noise, is a significant issue in ultrasound imaging, and few studies have tackled this specific challenge [27–29]. Many recent studies only focus on additive Gaussian noise,
but for speckle noise, there are few studies. There are many studies that define a new novel method to denoise the ultrasound images [30]. For additive noise, developed the Non-Local Means (NLM) filtering method
[31], which is widely used for additive noise removal in MRI and CT imaging. While NLM is effective for
Gaussian noise, it struggles with speckle noise due to its different statistical properties. Consequently, various
NLM adaptations [32, 33] have emerged to better target speckle noise. Another study [34] proposed refining
NLM weights within a low-dimensional PCA subspace for improved speckle reduction. Although this strategy
enhances noise targeting, PCA’s processing complexity may limit its efficiency in real-time applications.
APTISI Transactions on Technopreneurship (ATT), Vol. 7, No. 3, November 2025, pp. 713–725

APTISI Transactions on Technopreneurship (ATT)

❒

715

The Optimized Bayesian Non-Local Means (OBNLM) filter [35] was invented to tackle patch similarity in noisy conditions. This method uses a Bayesian framework for improving patch matching by replacing
the traditional Euclidean distance with a Pearson distance. Although its effectiveness, this technique is computationally demanding for high-resolution photos. [36] introduced a hybrid methodology that fused local
statistical characteristics and NLM. This strategy tries to enhance accuracy in difficult noise situations by integrating local and global information. However, integrating numerous details may inadvertently increase the
risk of over-smoothing intricate visual elements. Another studied non-local Total Variation (TV) method utilizing NLM to address nonlinear noise. In contrast to traditional television approaches, these non-local television
methods enhance retention of image attributes. Nonetheless, television systems occasionally fail to accurately
capture intricate details in textured regions, perhaps leading to the omission of sensitive characteristics [37, 38].
[39] introduced an intelligent classification approach for ovarian detection that integrates texture and
intensity characteristics. They used autocorrelation, and sum variance from the Grey-Level Co-Occurrence
Matrix (GLCM), along with intensity from k-means clustering, to train a backpropagation neural network for
detecting ovarian issues. Despite its potential, the method’s dependence on manually crafted characteristics
restricts its applicability to alternative ultrasound scenarios. [40] concentrated on generating noise-free photos
by customizing filters for particular noise categories. A median filter was employed for salt-and-pepper noise,
while an adaptive Wiener filter was utilized for Gaussian noise, resulting in significant enhancements in Mean
Squared Error (MSE). However, this method might not work well in situations with different types of noise
mixed together, and needing to manually identify the types of noise could make it harder to use automatically.
[41] used OBNLM filtering with Convolutional Neural Networks (CNNs) to eliminate noise from ultrasound
pictures. Here, the OBNLM filter first cleans the image, and the CNN further refines the denoising, offering
effective denoising. However, using two distinct models (OBNLM and CNN) may increase computational
demands, impacting real-time applications. [42] used the Two-Dimensional Fractional Fourier Transform (2DFrFT) to denoise ultrasound images. To optimize a parameter, they employed a transfer learning-based VGG16 model. While this method targets specific artifact types and adapts to various noise angles, [42] study
tested the model on only 10 images, limiting generalizability and raising concerns about real-world scalability.
Some recent studies [43, 44] focused on mobile application in medical image analysis for more convenience.
[43] developed mobile application called Halodoc enhancing patient-doctor consultations while ensuring data
privacy and efficiency. Whereas [44] explored gradient centralization in conjunction for enhanced performance.
Recent studies [44–47] focused on technopreneurial innovation in the healthcare sector. The developments in healthcare [45–47] suggested the technopreneurial innovation through AI involvement to enhance
healthcare quality and sustainability and emphasized the importance of entrepreneurship for business growth.
Whereas [47, 48] suggested business models and stakeholder roles for digital health entrepreneurship and emphasized the need to understand the industry’s business strategies and regulatory environment.
3.

MATERIALS AND METHODS

Figure 1. Block Diagram of Proposed Denoise Method

716

E-ISSN: 2656-8888 | P-ISSN: 2655-8807

❒

In the above Figure 1 shows the block diagram of the proposed denoised model. Here the ultrasound
images are first pre-processed to maintain consistency, and after that, the speckle noise is added with different
levels (0.001, 0.02, and 0.5) to demonstrate the proposed method’s performance. After that, the noisy image is
input to the proposed denoising model. The proposed method preserves relevant details for a precise diagnosis
by producing a denoised image with less speckle noise. The attention mechanism of this architecture, which is
detailed in the block, allows the model to prioritize important features, improving its denoising performance at
various noise levels. Each step is explained in detail in the following subsections [49, 50].
3.1. Dataset Description and Noise Addition
A publicly accessible dataset of ultrasound images associated with PCOS was used in this paper. The
dataset includes 3,856 samples in total, each of which represents ultrasound images taken for PCOS analysis
and diagnosis. These photos offer vital visual information that helps detect cysts and other anomalies linked
to illness. Machine learning and image processing tasks benefit greatly from the dataset’s diversity, which
includes a wide range of patient profiles and variations in PCOS manifestation.
Even though the original ultrasound image dataset is clean, synthetic speckle noise was added for
this study to mimic real-world imperfections and improve the model’s resilience. Different noise intensities
were produced by adding speckle noise, a granular noise frequently seen in medical ultrasound imaging, at
different variance levels. For the suggested model to successfully remove noise from the photos, this step was
essential for training. Applying the model to real-world noisy ultrasound images allowed it to generalize better
because it was exposed to a wide range of scenarios through the simulation of noise at varying intensities. A
mathematical representation of the addition of speckle noise to a clear ultrasound image X is as follows:
Xnoisy = X + X ⊙ N

(1)

Where Xnoisy is the noisy image, X is the clean image, N ∼ N (0, σ 2 ) is the noise matrix sampled
from a Gaussian distribution with mean 0 and σ 2 , and ⊙ represents element-wise multiplication. The variance
σ 2 is varied to simulate different levels of noise intensity, enabling the model to handle diverse noise conditions.
The Figure 2 below illustrates a clean ultrasound image alongside versions with varying levels of added noise.

Figure 2. Clean Ultrasound with Noise Image with Varying Levels of Variance
3.2. Pre-Processing
In this, different preprocessing techniques are employed to enhance the quality of the image before
inputting it into the proposed method. As the image size varies so to make constant take a uniform size
(128128x3) for every image. To bring all image values in certain ranges, we used normalization techniques.
also used image smoothing to remove small artifacts while maintaining important features to improve performance. By enhancing the consistency and quality of the input data, these procedures optimize the deep learning
model’s performance.
3.3. Filtering Method
To effectively eliminate noise from the ultrasound images, we employed an AAE as our filtering
method. Autoencoders contain mainly two parts: an encoder that converts the input image into a low-dimensional
latent space and a decoder that reconstructs the image from this latent representation. In our case, we extended
this architecture by integrating an attention mechanism, allowing the model to focus on the most relevant features in the image, thereby improving noise removal performance.
The attention mechanism plays a crucial role in enhancing the performance of the proposed AAE by
selectively focusing on important image regions while suppressing noise. Traditional autoencoders compress
APTISI Transactions on Technopreneurship (ATT), Vol. 7, No. 3, November 2025, pp. 713–725

❒

APTISI Transactions on Technopreneurship (ATT)

717

the input into a latent representation and reconstruct the image, but they often fail to differentiate between
important structures (e.g., ovarian cysts) and unwanted noise patterns (e.g., speckle noise). To address this issue,
an attention gate network is incorporated, which dynamically adjusts the feature representation by highlighting
relevant information and down-weighing less significant regions.
Mathematically, for a given input image Xnoisy ∈ Rh×w , where h and w represent the height and width
′
′
of the noisy image, the encoder function fencoder (·) projects Xnoisy into a latent space Z ∈ Rh ×w ×d , where the
number of channels is represented by d, and h′ , w′ are the dimensions of the compressed representation. This
can be represented as:
Z = fencoder (Xnoisy )
(2)
′

′

Next, the attention mechanism computes an attention map A ∈ Rh ×w ×d , which assigns weights to
different regions of the latent space. The attention map is derived as:
A = σ(Wa · Z + bb )

(3)

Where Wa and bb are learnable weights and biases of the attention network, and σ(·) represents the softmax
function, ensuring that the attention weights sum to 1 across spatial dimensions, emphasizing the most important regions of the feature map. The attention map is then applied elementwise to the latent representation Z
using the Hadamard product (element-wise multiplication):
Z̃ = A ⊙ Z

(4)

h′ ×w′ ×d

Here, Z̃ ∈ R
is the attended latent representation, which retains important features while
down-weighting noise. This attended latent space is then passed to the decoder fdecoder (·) to reconstruct the
denoised image X̂:
X̂ = fdecoder (Z̃)
(5)
The objective is to minimize the difference between the original clean image X and the reconstructed
denoised image X̂. The loss function used to train the model is the mean squared error (MSE) between the
clean image and the denoised image, defined as:
n

1X
(Xi − X̂i )2
Lmse =
n i=1

(6)

After that, to enhance the attention mechanism’s ability to focus on relevant regions of the image,
we introduce an attention regularization term. This term penalizes the attention map when it becomes overly
diffuse, ensuring that the model focuses on the most critical areas. The regularization is based on the L1 norm
of the attention map, defined as:
n
1X
Latt =
∥Ai ∥1
(7)
n i=1
Where ∥Ai ∥1 is the L1 norm of the attention map for the i-th sample, encouraging sparsity in the
attention weights. The final loss function used to train the model combines the reconstruction loss (mean
squared error) and the attention regularization:
L = Lmse + λLatt

(8)

In here, λ is a hyperparameter of regularization that governs the trade-off between the reconstruction
of the clean image and the sparsity of the attention map. The autoencoder with attention is trained to minimize
this compound loss, learning to remove noise from ultrasound images, preserving relevant anatomical structures while eliminating real and synthetic speckle noise. This is a significant contribution to our work since
the regularization-driven attention mechanism enables the model to better separate noise from informative features, leading to improved denoising. The attention mechanism of the proposed AAE model is improving the
denoising ability of PCOS ultrasound images by selectively enhancing notable areas of images and suppressing noise. Unlike traditional filtering methods where uniform noise removal is enforced globally on the entire
image, attention adapts the network to selectively concentrate on diagnostically significant areas such as ovarian cysts and follicular patterns while resisting the speckle noise minimally. Through this selective attention,

718

❒

E-ISSN: 2656-8888 | P-ISSN: 2655-8807

it ensures that such critical anatomical data is properly retained to prevent being lost from the fine structures
needed in a correct PCOS diagnosis. Further, through dynamic feature weighing with the attention gate network, the model learns to differentiate noise from useful patterns and is hence less likely to over-smooth. This
results in higher image sharpness and improved diagnostic readability. Furthermore, the attention mechanism
enhances the model’s generalizability by being able to handle different levels of noise efficiently and thus being
trustworthy under different imaging conditions.
3.4. Approach to Denoising Raw PCOS Images
In this case, an AAE model is utilized for the denoising of ultrasound images. With the help of the
attention mechanism, the model can effectively remove the noise while preserving the important features, which
are helpful for PCOS detection. Compared to existing methods, these methods allow the use of adaptively
learning the noise pattern to improve denoising for images. Subsequently, the various steps are elucidated for
noise removal:
• Step 1: Firstly, collect all the ultrasound images and perform some preprocessing operations, namely
smoothing, normalization, and resizing, to help make the images clearer and better visuals. Then, speckle
noise, at several different levels, is added to the database for further evaluation.
• Step 2: After this, these noisy images are entered into the proposed model to begin the training. Since
the model is exposed to several different noise levels, the model can learn to separate noise from useful
features in the image.
• Step 3: Utilize attention to obtain the important parts of the image for feature extraction to enhance the
noise filtering while keeping the important details.
• Step 4: To measure how effectively the image has been denoised, this paper compares it with state-ofthe-art methods, using quality measures such as the Structural Similarity Index Measure (SSIM) and
Peak Signal-to-Noise Ratio (PSNR).
4.

RESULT AND DISCUSSION
This section describes the experimental procedure, including dataset preparation, noise addition,
model training, and evaluation using various performance metrics, along with a comparative analysis of results.
4.1. Experimental Setup
This research assesses the performance of a particular denoising technique against a publicly available
dataset of 3,356 PCOS ultrasound images. These images contain speckle noise at three levels: 0.001, 0.02, and
0.5. The dataset consists of 2,685 training images and 671 testing images. Training the model on each specific
noise level makes the model more robust to various kinds of noise. Such a strategy methodically evaluates the
model’s performance with regard to removing noise and correctly identifying data at different noise levels to
confirm good performance in practical scenarios.
4.2. System and Hyperparameter Specifications
The model was trained on an Intel® Xeon® Silver 4216 CPU, 128 GB RAM, and Nvidia® RTX
A4000 GPU with 16 GB VRAM. A loss function, Mean Squared Error (MSE), was used, and an optimizer
function to train the denoise model was Adaptive Moment Estimation (Adam), which utilizes momentum and
adaptive learning rate for faster convergence. The batch size is set at 64, and the learning rate is 0.0001 to strike
an equilibrium between stability and efficiency. The model was trained on 2,000 epochs so that the model
learns from the noise dataset and adjusts to different noise levels, which improves generalization for real-world
scenarios. The proposed model tuning process for ultrasound image denoising was configured with specific
performance parameters, resulting in optimal performance during the tuning process. The model was trained
with an input and output size of 128x128x3. A learning rate of [0.01, 0.001, 0.00005] was initially used to find
the most stable setting.
Higher learning rates, such as 0.01, led to unstable convergence, while excessively low values, such as
0.00005, resulted in slow learning. The chosen learning rate of 0.001 ensured smooth and stable convergence.
For the optimizer, three options SGD, RMSprop, and Adam were evaluated. SGD led to slow convergence
APTISI Transactions on Technopreneurship (ATT), Vol. 7, No. 3, November 2025, pp. 713–725

❒

APTISI Transactions on Technopreneurship (ATT)

719

and difficulty escaping local minimums, while RMSprop exhibited fluctuating validation loss. Adam was
selected as the final optimizer due to his ability to balance momentum and adaptive learning rates, leading
to stable and efficient training. Batch size selection involved testing 32 and 64, where smaller batch sizes
introduced high variance, while larger batch sizes increased memory consumption and slowed convergence.
The optimal batch size of 64 offered the best balance between stability and training efficiency. Regarding the
number of epochs, the model was initially trained for 2000 epochs, with performance monitored throughout.
Loss stabilization was observed after 1500 epochs, indicating potential for early stopping around 1800 epochs
in practical applications. These hyperparameter choices were made to ensure the model effectively removed
noise while preserving key anatomical structures in ultrasound images.
4.3. Performance Evaluation Metrics
In this paper two useful matrices are used to evaluate the model performance: PSNR and SSIM. PSNR
provides a quantitative evaluation of image quality following denoising by calculating the relationship between
the maximum power of an image and the power of corrupting noise. It is computed as:

P SN R = 10 · log10

M AXI2
M SE


(9)

Where M AXI2 is the maximum pixel value, and MSE represents the Mean Squared Error between
the denoised and clean ultrasound images. A high PSNR value gives better denoising performance and closer
to the original image.
SSIM, on the other hand, evaluates structural similarity between images by considering luminance,
contrast, and structural information. It ranges from −1 to 1, with values closer to 1 indicating higher similarity.
The SSIM is given by:
SSIM (x, y) =

(2µx µy + C1 )(2σxy + C2 )
2
(µx + µ2y + C1 )(σx2 + σy2 + C2 )

(10)

Where µx and µy are the means, σx2 and σy2 are variances, σxy is the covariance of images x and y,
while C1 and C2 are constants. Together PSNR and SSIM provide a comprehensive evaluation of the denoising
models’ performance in preserving image fidelity and structural details.
4.4. Experimental Results for Denoising
The proposed attention model was applied to noisy ultrasound images, effectively producing denoised
outputs. This approach, illustrated in Figure 3, was quantitatively assessed using PSNR and SSIM, with results
detailed in Table 1. The model’s performance was compared against five state-of-the-art denoising techniques,
including mean filter, Laplacian filter, bilateral filter, two-stage image denoising, and the OBNLM filter. Each
method contributes unique noise-reduction characteristics: the mean filter averages pixel values for smoothing,
the Laplacian filter enhances edges, and the bilateral filter preserves edges while removing noise. The TwoStage Image Denoising method employs Discrete Wavelet Transform (DWT) for decomposing and thresholding
wavelets, with noise patterns classified through a neural network to apply adaptive filters. The OBNLM filter,
combined with a CNN, efficiently removes speckle noise, yielding a refined, denoised output. Together, these
methods provide a comprehensive framework for denoising ultrasound images.
Table 1. Performance Evaluation of Proposed Denoise Method with State of Art Methods.
Performance after denoise
Performance before denoise
0.5
0.02
0.001
Performance metric
PSNR SSIM PSNR SSIM PSNR SSIM
Mean filter [39]
28.31 15.63 28.40 45.91 28.52 54.28
Laplacian filter [40]
28.35 18.23 28.49 55.25 28.46 60.80
Bilateral [35]
29.32 30.23 29.65 60.25 29.74 72.19
Two-Stage Denoising [42]
29.35 48.52 29.72 67.32 29.83 78.54
OBNLM Filter [41]
29.40 60.80 30.98 88.24 31.22 89.76
Proposed Method
31.33 85.21 34.25 92.33 36.23 99.25

720

❒

E-ISSN: 2656-8888 | P-ISSN: 2655-8807

The above Table 1 compares the denoising performance of various methods across different noise
variance levels (0.5, 0.02, and 0.001) using PSNR and SSIM metrics. At a high variance level of 0.5, the
proposed method achieves a PSNR of 31.33, which is significantly higher than other methods, with a 7%
improvement over the second bilateral filter (29.32 PSNR). At a lower variance of 0.02, the proposed method
reaches a PSNR of 34.25 and an SSIM of 92.33, showing a 7.5% PSNR improvement over the next-best
OBNLM filter (30.98 PSNR). For the lowest noise variance (0.001), the proposed method yields a PSNR of
36.23 and an impressive SSIM of 99.25, indicating a major SSIM enhancement compared to the OBNLM filter,
which achieved 89.76 SSIM.
The proposed method gives better results as it uses an AAE structure, which effectively identifies and
preserves fine image details while focusing on regions needing denoising. In contrast, traditional methods such
as the mean and Laplacian filters struggle due to their simplistic operations; the mean filter blurs important
details by averaging pixel values, while the Laplacian filter, designed primarily for edge detection, amplifies
noise. The bilateral filter maintains edges reasonably well but cannot handle high noise levels as effectively.
The two-stage image denoising approach often over-smooths, sacrificing detail for noise reduction. Unlike
these methods, the proposed AAE adaptively enhances image quality, demonstrating clear advantages across
all tested noise levels.

Figure 3. Qualitative Result of Denoise Ultrasound Image on Varying Noise Levels
The above Figure 3 shows the performance of various denoising methods on an ultrasound image
with different noise variances levels: 0.5, 0.02, and 0.001. At the highest noise level (variance = 0.5), the noisy
image is extremely grainy, making it challenging to discern structural details. While some filters like the mean
and bilateral filters manage to reduce noise, they often blur essential features. The Laplacian filter emphasizes
edges but fails to address noise effectively at this high level. The OBNLM filter and the two-stage image
denoising approach achieve moderate noise reduction, but only the proposed method effectively removes most
of the noise while preserving image details, resulting in the clearest output in this challenging noise condition.

Figure 4. Line chart Comparison of Denoising Methods Based on PSNR and SSIM Across Different Noise
Levels
APTISI Transactions on Technopreneurship (ATT), Vol. 7, No. 3, November 2025, pp. 713–725

APTISI Transactions on Technopreneurship (ATT)

❒

721

As the noise variance decreases to 0.02 and then 0.001, all methods show improved results, with
more structural details becoming visible in the images. At variance 0.02, filters like the OBNLM and bilateral
filters maintain more clarity while reducing noise, but minor artifacts or blurring persist. By the time the
noise is minimal (variance = 0.001), the denoising methods produce images that are quite close to the original.
The proposed method continues to outperform others across all levels of noise, consistently delivering the
cleanest images with the highest structural fidelity, indicating its robustness and adaptability across different
noise conditions.
The above Figure 4 compares the performance of various image denoising methods at different noise
levels, showing their effectiveness in terms of PSNR on the left and SSIM on the right. In both plots, the
horizontal axis represents noise levels, with lower values indicating less noise, and the vertical axis represents PSNR and SSIM values, respectively. Higher PSNR and SSIM values mean better image quality after
denoising. The Proposed Method outperforms the others across all noise levels, achieving the highest PSNR
and SSIM values, indicating it preserves image quality better than traditional methods like the Mean Filter,
Laplacian Filter, Bilateral Filter, Two-Stage Denoising, and OBNLM Filter. The other methods show moderate
improvements as noise levels decrease, but they are generally less effective than the Proposed Method.
The below Figure 5 illustrates the performance of various denoising techniques on various levels of
noise. There is an axis for each noise level (0.5, 0.02, and 0.001), with the higher value on each axis representing
improved performance in retaining image quality upon denoising. The Proposed Method performs better than
other methods for all levels of noise, in terms of both PSNR and SSIM, as indicated by its broader spread
towards the corners of the chart. There are other techniques, including the Mean Filter, Laplacian Filter,
Bilateral Filter, Two-Stage Denoising, and OBNLM Filter, with lesser spreads, revealing lower PSNR and
SSIM values and hence less efficient denoising. The radar chart clearly demonstrates superior performance of
the Proposed Method in achieving both structural similarity and signal-to-noise ratio for various intensities of
noise.

Figure 5. Radar Chart Comparing Denoising Methods Based on PSNR and SSIM Across Different Noise
Levels
4.5. Discussion
This paper presents an AAE model for PCOS ultrasound image denoising that performs well across
noise levels. Despite promising quantitative and qualitative outcomes, the model has numerous drawbacks.
First, the dataset uses synthetic speckle noise rather than clinical noise, which may limit generalizability to
uncontrolled clinical situations. Second, the model was trained and evaluated on a single public dataset without
multi-institutional validation, which may impair its adaptability to different imaging settings. Finally, while
the suggested approach enhances denoising efficiency, resource-constrained devices like handheld ultrasound
scanners were not tested for real-time performance.
This technopreneurial work shows how AI-driven diagnostic tools might improve reproductive health
imaging. Policy measures are needed to apply these findings. Educational institutions should teach AI and
biological imaging to prepare technopreneurs. AI-enabled women’s health diagnostics firms can receive government grants or subsidies. Incubators and health-tech accelerators can also help clinicians and developers
collaborate on clinically relevant and economically feasible AI advances. These suggestions enable digital
transformation and medical AI adoption in entrepreneurial ecosystems. Additionally The suggested AI-based
denoising model makes it easier to diagnose Polycystic Ovary Syndrome (PCOS) early on, which is a direct
contribution to SDG 3 (Good Health and Well-being). It fits with SDG 9 (Industry, Innovation, and Infrastructure) since it encourages AI innovation in ultrasound imaging. Also, its possible use in health tech incubators

722

E-ISSN: 2656-8888 | P-ISSN: 2655-8807

❒

and digital education platforms fits with SDG 4 (Quality Education) by giving future healthcare technopreneurs
and physicians AI capabilities to help them make better diagnoses.
5.

MANAGERIAL IMPLICATIONS
The findings of this study offer significant implications for healthcare management, particularly in
hospitals, clinics, and diagnostic centers. The adoption of the proposed Attention-based Autoencoder (AAE)
model has the potential to enhance the quality of ultrasound imaging for PCOS diagnosis, thereby improving
clinical decision-making processes. From a managerial perspective, integrating such AI-based approaches
could contribute to reducing diagnostic errors, strengthening institutional credibility, and ensuring the provision
of higher-quality reproductive health services.
For health technology companies and emerging diagnostic startups, the study provides valuable insights into the commercialization of AI-driven medical imaging solutions. The proposed model can serve
as a foundation for developing automated diagnostic applications that may be incorporated into portable ultrasound devices or telemedicine platforms. This innovation can expand patient access to reliable diagnostic
tools, enhance market competitiveness, and foster the development of sustainable business models in the digital
healthcare ecosystem.
Furthermore, the study underscores the importance of aligning technological implementation with
regulatory frameworks, ethical standards, and patient data protection. Managers are required to ensure that
the deployment of AI-based diagnostic tools adheres to privacy regulations and clinical governance practices.
In this regard, the managerial implications extend beyond operational improvements and financial outcomes,
encompassing sustainability, trust-building among stakeholders, and contributions to SDG 3 (Good Health and
Well-being) and SDG 9 (Industry, Innovation, and Infrastructure).
6.

CONCLUSION
This paper offers a new AAE model for speckle noise removal in ultrasound images, with a focus
on PCOS detection. The presented method well removes noise while retaining important anatomy information, showing prominent improvement over the conventional denoising approach. By utilizing an attention
mechanism, the model conducts selective enhancement of key regions of the image, achieving greater PSNR
(31.33, 34.25, 36.23) and SSIM (85.21, 92.33, 99.25) values under varying amounts of noise. Comparative
analysis with current methods, such as mean filtering, bilateral filtering, and OBNLM filtering, illustrates the
better performance of the proposed method. These findings show the potential clinical value of using deep
learning-based denoising in ultrasound imaging to improve diagnostic accuracy for PCOS. Future studies may
investigate multi-scale attention mechanisms or hybrid models that integrate CNNs and transformers to further
enhance noise removal and image quality. In addition, applying the same concept to other medical imaging fields, such as thyroid, liver, and kidney ultrasound scanning, might also extend its reach in the field of
medicine. The use of ultrasound imaging in diagnosing PCOS is made easier by the proposed AAE, which
greatly enhances image quality while effectively eliminating speckle noise. This research aids in the wider
application of artificial intelligence in health systems, particularly concerning digital diagnostic devices. Later
studies may investigate attention mechanisms using transformers and adapt the technique to other branches of
medical imaging for more significant clinical influence.
7.
DECLARATIONS
7.1. About Authors
Pratibha Pandey (PP)

https://orcid.org/0000-0002-3122-4352

Sumit Chaudhary (SC)

https://orcid.org/0000-0002-5455-774X

Xin Nie (XN)

https://orcid.org/0000-0002-5005-9145

7.2. Author Contributions
Conceptualization: PP; Methodology: SC; Validation: PP and SC;; Writing Original Draft Preparation: PP and SC; Writing Review and Editing: SC and XN; Visualization: SC; All authors, PP, SC, and XN,
have read and agreed to the published version of the manuscript.
APTISI Transactions on Technopreneurship (ATT), Vol. 7, No. 3, November 2025, pp. 713–725

APTISI Transactions on Technopreneurship (ATT)

❒

723

7.3. Data Availability Statement
The data presented in this study are available on request from the corresponding author.
7.4. Funding
This work was supported by a grant from the Hubei Key Laboratory of Intelligent Robot of China
(Grant No. HBIRL202009).
7.5. Declaration of Conflicting Interest
The authors declare that they have no conflicts of interest, known competing financial interests, or
personal relationships that could have influenced the work reported in this paper.
REFERENCES
[1] E. Stener-Victorin, H. Teede, R. J. Norman, R. Legro, M. O. Goodarzi, A. Dokras, J. Laven, K. Hoeger,
and T. T. Piltonen, “Polycystic ovary syndrome,” Nature Reviews Disease Primers, vol. 10, no. 1, p. 27,
2024.
[2] H. Singh, A. S. Ahmed, F. Melandsø, and A. Habib, “Ultrasonic image denoising using machine learning
in point contact excitation and detection method,” Ultrasonics, vol. 127, p. 106834, 2023.
[3] A. Chaudhuri, “Polycystic ovary syndrome: Causes, symptoms, pathophysiology, and remedies,” Obesity
Medicine, vol. 39, p. 100480, 2023.
[4] R. Saranya and R. Sridevi, “Design of automatic follicle detection and ovarian classification system for
ultrasound ovarian images,” Multimedia Tools and Applications, pp. 1–27, 2024.
[5] K. M. Hoeger, A. Dokras, and T. Piltonen, “Update on pcos: consequences, challenges, and guiding
treatment,” The Journal of Clinical Endocrinology & Metabolism, vol. 106, no. 3, pp. e1071–e1083,
2021.
[6] D. Jonas, H. D. Purnomo, A. Iriani, I. Sembiring, D. P. Kristiadi, and Z. Nanle, “Iot-based community
smart health service model: Empowering entrepreneurs in health innovation,” Aptisi Transactions on
Technopreneurship (ATT), vol. 7, no. 1, pp. 61–71, 2025.
[7] T. Pujiati, H. Setiyowati, B. Rawat, N. P. L. Santoso, and M. G. Ilham, “Exploring the role of artificial intelligence in enhancing environmental health: Utaut2 analysis,” Sundara Advanced Research on Artificial
Intelligence, vol. 1, no. 1, pp. 37–46, 2025.
[8] L. P. Dewanti, L. Sitoayu, and A. Idarto, “Digital tele-counseling for sustainable maternal health services
in indonesia focus on telelactation,” IAIC Transactions on Sustainable Digital Innovation (ITSDI), vol. 6,
no. 1, pp. 10–20, 2024.
[9] S. Ahmed, M. S. Rahman, I. Jahan, M. S. Kaiser, A. S. Hosen, D. Ghimire, and S.-H. Kim, “A review
on the detection techniques of polycystic ovary syndrome using machine learning,” IEEE Access, vol. 11,
pp. 86 522–86 543, 2023.
[10] S. A. Suha and M. N. Islam, “An extended machine learning technique for polycystic ovary syndrome
detection using ovary ultrasound image,” Scientific Reports, vol. 12, no. 1, p. 17123, 2022.
[11] W. M. H. W. Mahmud, A. Huong, and N. A. Jumadi, “Preliminary analysis of ultrasound features for
detection of polycystic ovary syndrome (pcos) in women,” in Biomedical Engineering: AI and Technological Innovations. Springer, 2025, pp. 137–151.
[12] A. Nazarudin, N. Zulkarnain, S. S. Mokri, W. M. D. W. Zaki, A. Hussain, M. F. Ahmad, and I. N. A. M.
Nordin, “Performance analysis of a novel hybrid segmentation method for polycystic ovarian syndrome
monitoring,” Diagnostics, vol. 13, no. 4, p. 750, 2023.
[13] R. Yang, Y. Zou, L. Li, W. V. Liu, C. Liu, Z. Wen, and Y. Zha, “Enhancing repeatability of follicle counting
with deep learning reconstruction high-resolution mri in pcos patients,” Scientific Reports, vol. 15, no. 1,
p. 1241, 2025.
[14] A. M. Rahmat, N. N. Noor, I. L. Maria, and G. Davies, “Risk factors for the incidence of drug-resistant
tuberculosis in the labuang baji hospital: A healthpreneur perspective,” Aptisi Transactions on Technopreneurship (ATT), vol. 6, no. 3, pp. 468–480, 2024.
[15] J. L. Willson, A. Nuche, and R. Widayanti, “Ethical considerations in the development of ai-powered
healthcare assistants,” International Transactions on Education Technology (ITEE), vol. 2, no. 2, pp.
109–119, 2024.

724

❒

E-ISSN: 2656-8888 | P-ISSN: 2655-8807

[16] A. S. Rafika, A. Faturahman, B. N. Henry, F. D. Yulian, and M. Hassan, “Ai-driven big data solutions for
personalized healthcare: Analyzing patient data to improve treatment outcomes,” Journal of Computer
Science and Technology Application, vol. 2, no. 1, pp. 36–45, 2025.
[17] C. Yu, F. Ren, S. Bao, Y. Yang, and X. Xu, “Self-supervised ultrasound image denoising based on
weighted joint loss,” Digital Signal Processing, vol. 162, p. 105151, 2025.
[18] Y.-H. Li, Y.-L. Li, M.-Y. Wei, and G.-Y. Li, “Innovation and challenges of artificial intelligence technology
in personalized healthcare,” Scientific reports, vol. 14, no. 1, p. 18994, 2024.
[19] L. Wang, X. Wang, A. Vizziello, and P. Gamba, “Rsaae: Residual self-attention-based autoencoder for
hyperspectral anomaly detection,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp.
1–14, 2023.
[20] P. Singh and A. Sharma, “Attention-based convolutional denoising autoencoder for two-lead ecg denoising and arrhythmia classification,” IEEE Transactions on Instrumentation and Measurement, vol. 71, pp.
1–10, 2022.
[21] G. Govindharajan, S. Subramanian, M. Doraipandian, and S. Rajendran, “Innovative ai-driven models
for predicting polycystic ovarian syndrome: An extensive review of machine learning and deep learning
frameworks,” Archives of Computational Methods in Engineering, pp. 1–26, 2025.
[22] K. van der Ham, F. Barbagallo, E. van Schilfgaarde, M. E. Lujan, J. S. Laven, and Y. V. Louwers, “The additional value of ultrasound markers in the diagnosis of polycystic ovary syndrome,” Fertility and Sterility,
vol. 123, no. 2, pp. 342–349, 2025.
[23] B. Karthik and K. R. CV, “Speckle noise removal from biomedical mri images and classification by
msvm,” in Green AI-Powered Intelligent Systems for Disease Prognosis. IGI Global, 2024, pp. 261–
278.
[24] S. Reka, T. S. Praba, M. Prasanna, V. N. N. Reddy, and R. Amirtharajan, “Automated high precision pcos
detection through a segment anything model on super resolution ultrasound ovary images,” Scientific
Reports, vol. 15, no. 1, p. 16832, 2025.
[25] N. A. El-Hag, H. M. El-Hoseny, and F. Harby, “Dnn-driven hybrid denoising: advancements in speckle
noise reduction,” Journal of Optics, pp. 1–10, 2024.
[26] P. Sharmila, P. Sivasakthi et al., “Classification of pcos in ultrasound images using deep learning methods,” in 2024 International Conference on IoT, Communication and Automation Technology (ICICAT).
IEEE, 2024, pp. 1089–1092.
[27] S. Sowmiya, S. Umapathy, O. Alhajlah, F. Almutairi, S. Aslam, and R. Ahalya, “F-net: Follicles net an
efficient tool for the diagnosis of polycystic ovarian syndrome using deep learning techniques,” Plos one,
vol. 19, no. 8, p. e0307571, 2024.
[28] P. Bedi, S. Goyal, A. S. Rajawat, and M. Kumar, “An integrated adaptive bilateral filter-based framework
and attention residual u-net for detecting polycystic ovary syndrome,” Decision Analytics Journal, vol. 10,
p. 100366, 2024.
[29] M. Harada, “Pathophysiology of polycystic ovary syndrome revisited: Current understanding and perspectives regarding future research,” Reproductive medicine and biology, vol. 21, no. 1, p. e12487, 2022.
[30] S. Cammarasana, P. Nicolardi, and G. Patanè, “Real-time denoising of ultrasound images based on deep
learning,” Medical & Biological Engineering & Computing, vol. 60, no. 8, pp. 2229–2244, 2022.
[31] X. Zhang, “Two-step non-local means method for image denoising,” Multidimensional Systems and Signal Processing, vol. 33, no. 2, pp. 341–366, 2022.
[32] A. Kaur and G. Dong, “A complete review on image denoising techniques for medical images,” Neural
Processing Letters, vol. 55, no. 6, pp. 7807–7850, 2023.
[33] B. Rawat, A. S. Bist, D. Supriyanti, V. Elmanda, and S. N. Sari, “Ai and nanotechnology for healthcare:
A survey,” APTISI Transactions on Management, vol. 7, no. 1, pp. 86–91, 2023.
[34] M. Kuswari, R. Gantino, and J. Maratis, “Maximizing healthcare service information system: Understanding the influence of integration on efficiency,” ADI Journal on Recent Innovation, vol. 6, no. 2, pp.
108–117, 2025.
[35] P. Coupé, P. Hellier, C. Kervrann, and C. Barillot, “Nonlocal means-based speckle filtering for ultrasound
images,” IEEE transactions on image processing, vol. 18, no. 10, pp. 2221–2229, 2009.
[36] Y. Zhan, M. Ding, L. Wu, and X. Zhang, “Nonlocal means method using weight refining for despeckling
of ultrasound images,” Signal Processing, vol. 103, pp. 201–213, 2014.
[37] L. Kask, N. Bloom, and R. Porta, “Health informatics: Utilization of information technology in health

APTISI Transactions on Technopreneurship (ATT), Vol. 7, No. 3, November 2025, pp. 713–725

APTISI Transactions on Technopreneurship (ATT)

❒

725

care and patient management,” International Journal of Cyber and IT Service Management, vol. 4, no. 1,
pp. 53–58, 2024.
[38] A. K. Bedi and R. K. Sunkaria, “Ultrasound speckle reduction using adaptive wavelet thresholding,”
Multidimensional Systems and Signal Processing, vol. 33, no. 2, pp. 275–300, 2022.
[39] V. Kiruthika, S. Sathiya, and M. Ramya, “Machine learning based ovarian detection in ultrasound images,”
International Journal of Advanced Mechatronic Systems, vol. 8, no. 2-3, pp. 75–85, 2020.
[40] S. B. Choubey, A. Choubey, D. Nandan, and A. Mahajan, “Polycystic ovarian syndrome detection by
using two-stage image denoising.” Traitement du signal, vol. 38, no. 4, 2021.
[41] G. Vasavi and S. Jyothi, “Noise reduction using obnlm filter and deep learning for polycystic ovary
syndrome ultrasound images,” in International Conference On Computational And Bio Engineering.
Springer, 2021, pp. 203–212.
[42] M. Jha, R. Gupta, and R. Saxena, “Noise cancellation of polycystic ovarian syndrome ultrasound images
using robust two-dimensional fractional fourier transform filter and vgg-16 model,” International Journal
of Information Technology, vol. 16, no. 4, pp. 2497–2504, 2024.
[43] R. K. Hudiono and S. Watini, “Remote medical applications of artificial intelligence,” International Transactions on Artificial Intelligence, vol. 1, no. 2, pp. 182–187, 2023.
[44] V. Agarwal, M. Lohani, and A. S. Bist, “A novel deep learning technique for medical image analysis using
improved optimizer,” Health Informatics Journal, vol. 30, no. 2, p. 14604582241255584, 2024.
[45] U. Rahardja, P. A. Sunarya, Q. Aini, S. Millah, and S. Maulana, “Technopreneurship in healthcare: Evaluating user satisfaction and trust in ai-driven safe entry stations,” Aptisi Transactions on Technopreneurship
(ATT), vol. 6, no. 3, pp. 404–417, 2024.
[46] L. Patru, F.-A. Birchi, and C. L. Patru, “The relationship between digital technology and the development
of the entrepreneurial competencies of young people in the medical field,” Electronics, vol. 12, no. 8, p.
1796, 2023.
[47] A. V. Samrot, M. Bavanilatha et al., “Medical biotechnology-approaches to become an entrepreneur in
medical biotechnology,” in Opportunities for Biotechnology Research and Entrepreneurship. Bentham
Science Publishers, 2024, pp. 291–308.
[48] S. N. Weimar, R. S. Martjan, and O. Terzidis, “Conceptualizing the landscape of digital health entrepreneurship: a systematic review and research agenda,” Management Review Quarterly, vol. 75, no. 2,
pp. 1619–1667, 2025.
[49] A. E. Ilesanmi and T. O. Ilesanmi, “Methods for image denoising using convolutional neural network: a
review,” Complex & Intelligent Systems, vol. 7, no. 5, pp. 2179–2198, 2021.
[50] H. Thanganadar, S. M. Yaseen, S. K. Shukla, A. S. Bist, S. N. Shavkatovich, and P. Vijayakumar, “6g
wireless communication cyber physical system based smart healthcare using quantum optimization with
machine learning,” Wireless Personal Communications, pp. 1–20, 2024.