Real-Time microcalcification detection with SAAR algorithm

Abstract

Background

Detecting breast cancer, especially identifying microcalcifications in mammograms, is challenging due to the need for high sensitivity and efficient processing. This study presents a novel algorithm, Sigmoidal Slope Analysis and Aspect Ratio Evaluation (SAAR), designed for real-time application on edge devices. By employing a multi-step adaptive process with sigmoidal functions, SAAR enhances intensity contrast and prioritizes regions of interest, enabling fast, accurate detection of microcalcifications.

Objective

This study aims to develop and validate an efficient, edge-device-compatible method for detecting microcalcifications in mammographic images. The goal is to provide a tool that enhances diagnostic efficiency through real-time processing, thereby supporting early breast cancer detection in both clinical and remote settings.

Methods

The SAAR algorithm utilizes an adaptive slope detection technique based on the sigmoid function, dynamically adjusting to local intensity features. This approach allows for greater adaptability to image variations. The algorithm prioritizes regions of interest through a multi-step adaptive process, enhancing intensity differences to focus on potential microcalcifications.

Results

Testing on established mammography databases, such as MIAS, demonstrates the algorithm's effectiveness, with improved sensitivity compared to conventional methods. Designed for edge devices, the algorithm leverages their real-time processing capabilities, offering lower latency and enhanced privacy.

Conclusions

The integration of SAAR with edge devices represents a promising advancement in breast cancer detection. The adaptive nature of SAAR, coupled with the real-time processing capabilities of edge devices, provides a robust solution for enhancing microcalcification detection efficiency and sensitivity in mammography.

Keywords

Microcalcification detection mammography adaptive approach sigmoid function edge device

1 Introduction

Breast cancer is one of the most common malignant diseases among women, with increasing incidence and mortality rates.¹ Due to the absence of early symptoms, it is often detected at an advanced stage when metastases are already present, making treatment more difficult and reducing the chances of recovery. To facilitate early diagnosis, structured screening programs have been implemented, with mammography serving as the primary diagnostic tool for detecting tumors in women over 40.²

However, analyzing mammographic images is complex since tumor masses often lack well-defined edges and can vary in size.³ Microcalcifications, small calcium deposits in breast tissue, are often the only early indicator of breast cancer, making their detection crucial for timely diagnosis.⁴ On mammograms, microcalcifications appear as individual or clustered bright spots ranging in size from 50 microns to several millimeters.

Detecting microcalcifications is challenging for several reasons. Due to their small size and location within the breast, they may blend into the surrounding tissue, especially in cases of dense radiological composition.^5,6 When the mammogram background is heterogeneous or has high density, the contrast between microcalcifications and the surrounding tissue is reduced, making visual detection more difficult.⁷ Additionally, the presence of noise and film emulsion defects can create structures resembling microcalcifications, complicating their differentiation from actual deposits.⁸

To overcome these challenges, computer systems and artificial intelligence are used to automate microcalcification detection. Computer-aided diagnosis (CAD) algorithms enhance recognition by increasing contrast, filtering noise and classifying suspicious areas.⁹ The continued development of these technologies plays a crucial role in the early detection of breast cancer, enabling more accurate diagnoses and significantly contributing to reducing mortality and improving treatment outcomes.¹⁰

The breast cancer detection process can be further improved and accelerated by using edge devices, which execute diagnostic algorithms on data directly obtained from mammography machines. This approach enables fast data processing and real-time feedback, as the analysis is performed as close as possible to the point of data generation.

Local data processing on edge devices also provides significant advantages in terms of privacy and data integrity, as medical data is not stored on remote servers. Additionally, reducing the need to transfer large amounts of data between devices and centralized servers helps ease network congestion. These benefits make edge devices particularly attractive for use in mammographic diagnostics, as they enhance efficiency, security and patient satisfaction.¹¹ The diagram in Figure 1 illustrates the fundamental advantages of using edge devices in mammography.

Figure 1.

The fundamental advantages of utilizing edge devices in mammography.

The transfer of mammographic images to an edge device can be achieved through a direct connection via cables (USB, LAN) or wireless technologies such as Wi-Fi or Bluetooth. Most mammography devices use the DICOM (Digital Imaging and Communications in Medicine) standard for image transfer, so edge devices must also support this protocol to properly receive and process images.

In this paper, we present a novel method for detecting microcalcifications in mammograms, based on Sigmoid Slope Analysis and Aspect Ratio Evaluation (SAAR). The proposed multi-stage approach includes image preprocessing, edge enhancement and contour filtering based on size, location and distance from the image center to identify potential regions of interest (ROIs). Slope values within these ROIs are computed by analyzing pixel intensity differences with neighboring pixels, both vertically and horizontally. These values are transformed using a sigmoid function to emphasize relevant variations, enabling robust identification of microcalcifications.

The algorithm was validated using mammograms from the MIAS database, demonstrating its effectiveness in practical scenarios. Designed for real-time execution on edge devices, the SAAR algorithm enables rapid and localized image analysis, which is essential for timely diagnostics while preserving patient data privacy. The implementation was tested on the NVIDIA Jetson Nano, a platform tailored for edge computing and real-time medical image processing.

2 Literature review

2.1 Microcalcification detection in digital mammography

The development of computer algorithms for diagnostics using various methods and techniques has been the subject of numerous research studies over the past few decades. The primary goal of applying different image processing techniques to digital mammogram analysis is to identify potentially suspicious breast tissue. Since microcalcifications are a characteristic feature on mammograms and their structure significantly differs from that of other lesions, specialized diagnostic systems are typically developed for their detection. What follows is a small selection of scientific studies by renowned researchers; however, the total number of published papers describing various methods and algorithms for detecting microcalcifications in mammograms is considerably larger.

Wavelet-based techniques are widely used in microcalcification detection. Yoshida et al.^12,13 utilized asymmetric Daubechies wavelets with other techniques to achieve ∼90% detection accuracy. Similar methods are reported in.^14,15 Lahmiri¹⁶ combined wavelet decomposition with a k-Nearest Neighbors (k-NN) classifier. Subash Chandra Bose et al.¹⁷ also employed wavelet techniques. Wang et al.¹⁸ uses Dual-Tree Complex Wavelet Transform (DTCWT) with genetic optimization and Extreme Learning Machine (ELM) classifier, showing excellent region-of-interest detection performance.

Woods et al.¹⁹ and El-Naqa et al.²⁰ utilized enhanced k-NN and Support Vector Machine (SVM) methods, respectively. Authors of paper²¹ introduced a three-phase CAD system using SVM for cluster detection and Artificial Neural Network (ANN) for benign/malignant classification, reaching 98.4% sensitivity with only 0.85 FPs per image.

Verma and Zakos²² and Cheng et al.²³ combined fuzzy logic and neural networks. Subash Chandra Bose et al.¹⁷ integrated fuzzy logic, wavelet transforms and ANN for classification.

Nagel et al.²⁴ and Papadopoulos et al.²⁵ used rule-based and ANN combinations. Recent works^21,26–37 focus on Convolutional Neural Networks (CNNs) and deep learning. For example, Schonenberger et al.³⁵ use a deep CNN to classify BI-RADS patterns with 99.6% accuracy. Cai et al.²⁹ show that combining handcrafted and deep features improves diagnostic power. Paper²⁸ reports 99.99% accuracy and a very low false positive rate. Authors of papers³⁶ and³⁷ highlight the role of deep learning in BI-RADS 4 diagnosis and reducing unnecessary biopsies.

Studies³⁸ and³⁹ explore the clinical usability of Artificial Intelligence (AI) systems. Study³⁸ reports 94.7% sensitivity for AI-assisted mammography but emphasizes the necessity of radiologist interpretation. Study³⁹ uses Microsoft Custom Vision and finds that AI output needs human supervision due to limited sensitivity.

Yu et al.⁴⁰ apply wavelets, backpropagation networks and Markov random fields for identifying true calcifications, enhancing accuracy through advanced modeling. Markov random fields, along with other image processing methods like clustering, feature extraction and classification, were also used in.⁴¹

Recent studies have increasingly adopted the Multiple Instance Learning (MIL) paradigm to enhance microcalcification detection in mammographic images. In study,⁴² an MIL framework was proposed using adaptive breast segmentation and feature extraction from regions of interest to classify images as “normal” or “abnormal”, comparing both supervised and weakly supervised anomaly detection strategies. Xu et al.⁴³ introduced an ordinal variant of MIL (Ordinal MIL), enabling three-class mammogram classification (normal, benign, malignant) without labeling individual instances, thereby improving early screening. The authors of paper⁴⁴ presented the WMDNet model, combining multi-view input and weighted MIL to address class imbalance and the low prevalence of abnormalities, achieving solid results on public datasets. Elmoufidi⁴⁵ integrated textural features extracted via bidimensional decomposition and employed an SVM classifier in an MIL context for ROI classification, reaching high accuracy and AUC on DDSM, MIAS and INbreast datasets. These studies highlighted the growing importance of MIL approaches in automated mammogram analysis.

The reviewed literature demonstrates the use of a remarkably diverse range of methods for microcalcification detection – spanning from classical image processing and traditional machine learning techniques, through hybrid and fuzzy systems, to modern approaches involving deep learning and artificial neural networks. Detection results vary significantly depending on the applied methodology, data characteristics and targeted clinical use cases. Nevertheless, each of the reviewed methods provides a good basis for further research and can contribute to the improvement of modern CAD systems. Given the rapid progress in artificial intelligence, deep convolutional neural networks in particular stand out for their potential to improve the accuracy, reliability and efficiency of early breast cancer detection tools.

2.2 Application of edge device

The practical application of edge devices and their numerous advantages have been the subject of various studies presented in published scientific papers.^46–53 These devices enable local data processing, thereby reducing latency, alleviating the load on network infrastructure and enhancing system security. In industrial settings, edge technology is employed for monitoring production processes, predictive maintenance and real-time optimization of machine operations. In the field of medicine, the use of edge devices contributes to faster and more reliable processing of data from sensors and medical equipment, enabling timely clinical decisions and improved diagnostics, especially in environments with limited access to centralized computing resources. The cited studies confirm the significance of the edge approach as a key factor in improving the efficiency and autonomy of various systems within modern smart environments.

3 Materials and methods

3.1 Dataset

The algorithm proposed in this paper was tested using the Mini-Mammographic Database, provided by the Mammographic Image Analysis Society (MIAS).⁵⁴ This database contains a total of 322 different mammogram images, categorized into seven groups based on various pathological types. All mammograms are digitized at a pixel resolution of 200 microns, resulting in 1024 × 1024 pixel images. The microcalcification category within the database includes 23 images.

The images in the database were reviewed by experienced radiologists, who recorded the location, size and type of pathological changes, as well as the breast tissue type, i.e., dense or fatty.^55,56 These records serve as the reference standard used to validate the accuracy of the proposed method.

The algorithm proposed in this paper was tested on mammographic images obtained from well-known databases, the MiniMammographic, the Mammographic Image Analysis Society (MIAS).⁵⁴ The images in the database was reviewed by experienced radiologists and the location, size and type of pathological changes as well as the tissue type of breast tissue, i.e., dense or fatty, were recorded.^55,56 These records serve as a reference standard to justify the proposed method.

3.2 Mammogram preprocessing

Mammogram preprocessing is a crucial step in the process of automatic analysis of medical images, as it enhances image quality and isolates relevant information for further processing. In this work, the preprocessing phase included two main stages. First, noise was removed from the mammogram using Discrete Wavelet Transform (DWT), improving image clarity while preserving important structural details. In the second stage, the breast region was extracted from the background, allowing irrelevant parts of the image to be discarded and enabling a more focused analysis of the region of interest.

3.2.1 Image denoising

Mammographic images are often degraded by noise during acquisition or transmission, which can obscure small, low-contrast features such as microcalcifications. Effective noise removal is therefore essential for reliable detection and diagnosis.

In this study, image denoising is performed using Discrete Wavelet Transform (DWT) due to its capability for multi-resolution analysis and better preservation of fine details. DWT-based denoising operates in both spatial and frequency domains, allowing selective removal of noise while maintaining the integrity of diagnostically relevant structures.⁵⁷

The general wavelet denoising procedure consists of the following steps⁵⁸:

Application of the wavelet transform to the noisy image to obtain wavelet coefficients across multiple decomposition levels;

Thresholding of the wavelet coefficients using a selected method (hard or soft thresholding);

Reconstruction of the denoised image using the inverse wavelet transform.

In hard thresholding, all wavelet coefficients with absolute values below a specified threshold are set to zero, while others are left unchanged. In contrast, soft thresholding not only removes coefficients below the threshold but also shrinks the remaining coefficients toward zero by the threshold value. This approach helps suppress residual noise across all coefficients, assuming that each contains a mix of signal and noise.

The two thresholding functions can be expressed as:

\begin{aligned} Hard threshold : {\begin{cases} y = x, & | x | > T \\ y = 0, & | x | < T \end{cases} \end{aligned}

(1)

\begin{aligned} Soft threshold : y = \sin (x) (| x | - T) \end{aligned}

(2)

where x is the input coefficient, T is the threshold and y is the denoised output.

In this work, a two-level wavelet decomposition is applied using a biorthogonal wavelet (bior4.4), which is well-suited for medical imaging due to its symmetric property and good energy compaction. A threshold value of 3.9 was empirically selected to achieve an optimal trade-off between noise suppression and detail preservation.

The use of DWT in this context provides superior results by preserving high-frequency components critical for microcalcification detection. Moreover, its computational efficiency and suitability for implementation on edge devices make it an ideal choice for real-time, resource-constrained diagnostic environments.

3.2.2 Breast region extraction

In order to perform accurate analysis of mammographic images, it is essential to extract the breast region from the background. This step enables focusing on diagnostically relevant areas, avoiding unnecessary processing of background regions that may contain artifacts, labels, or film edges, especially in scanned analog images.

The method used in this work is described in detail in,⁴ while the key steps are summarized below.

1. Image Contrast Enhancement

Image contrast enhancement is performed using a simple logarithmic operation. This contrast-correction step is defined by the following expression:

\begin{aligned} g (x, y) & = \log [7 + f (x, y)] \end{aligned}

(3)

where

f (x, y)

is the original image and

g (x, y)

is the transformed image. Although this transformation is applied to the entire image, it significantly enhances the contrast in regions near the breast boundary, which are typically characterized by low density and poor detail definition. The purpose of this operation is to determine an approximate breast contour that closely follows the true boundary.

2. Image Quantization

Quantization is the process of representing the values of a sampled signal or image using a finite set of allowed values. In digital representation with n bits per sample and using only positive integers, there are $2^{n}$ possible quantization levels covering the range from 0 to $2^{n} - 1$ .

To obtain a binarized image, we used $n = 1$ bit per pixel, resulting in only two gray levels: “0” and “1”.

To correctly map the range of the analog signal to the input dynamic range of the quantizer, we set the lower threshold to 128 and the upper to 255. Thus, every pixel with a value between 128 and 255 is assigned a “1”, while all others are assigned “0”. This binarization step simplifies the separation of the breast region from the background.

3. Binary Image Enhancement

To remove small white areas that do not belong to the breast region and to smooth the contour, morphological erosion is applied using a diamond-shaped structuring element. This process eliminates isolated pixels and non-relevant objects.

4. Background Removal

To extract the breast region, a binary mask is created and applied to the image to remove background areas. The background removal procedure is described for the breast orientation shown in Figure 2 of this work. Since the approach is analogous for the opposite orientation, it is not necessary to describe both procedures separately.

Figure 2.

Mammogram preprocessing: (a) Original mammogram, (b) Extracted breast region.

To remove the breast background, it is necessary to create and apply a mask to the improved binarized version of the image. The mask creation algorithm is based on scanning the image row by row.

Mask Formation Algorithm:

Start with the first row.

Scan the row from right to left.

If a pixel is black (0), continue to the next.

Upon reaching the first white pixel (1), continue until a black pixel is found again.

From this point onward, set all remaining pixels in the row to black (0).

Move to the next row and repeat steps 2–5.

Once the binary mask is created, it is applied to the original image through element-wise multiplication, producing the final image with background removed and only the breast region preserved. The original mammogram and the mammogram with the background removed are shown in Figure 2, parts a) and b), respectively.

3.3 SAAR algorithm

In the first step of mammogram processing, image preprocessing was performed, including noise removal and extraction of the breast region from the background. In the second step, an edge detection is performed using the Canny algorithm to improve the visibility of the contours.^59,60 ROIs are generated from the centre to the edge of the image. The radius of each circular ROI gradually increases by 10% of the image width and covers different scales. For each circular ROI, the algorithm applies an adaptive slope-based detection method. A local window around each pixel is analysed and an adaptive threshold is calculated based on the mean intensity and standard deviation.

The algorithm uses a sigmoid function to modify the scores obtained from the vertical and horizontal differences. Microcalcifications that exceed the adaptive threshold are identified and marked within the circular ROI. The algorithm is parameterised by values such as predefined intensity thresholds, adaptive slope thresholds and Gaussian blur filter parameters. The fine tuning of these parameters allows the adaptation to the characteristics of specific images. The vertical score is calculated by computing the sigmoid of the vertical difference between the intensity of a pixel (suspect) and its median grey vertical neighbor.

\begin{aligned} verDiff & = medianGreyVer - suspect \end{aligned}

(4)

\begin{aligned} verScore & = σ (verDiff) \end{aligned}

(5)

Similar to the vertical score, the horizontal score is calculated by applying the sigmoid function to the horizontal difference between the intensity of a pixel and its median grey horizontal neighbor.

\begin{aligned} horDiff & = medianGreyHor - suspect \end{aligned}

(6)

\begin{aligned} horScore & = σ (horDiff) \end{aligned}

(7)

where σ(x) is sigmoidal function.^61,62

The sigmoid function introduces nonlinearity and is used to map the vertical and horizontal differences to values between 0 and 1. In the proposed algorithm, the sigmoid function is used to transform the intensity differences between a given pixel and its neighbors. This difference is then used to calculate a score that is further used in the detection of contaminants. The sigmoid function causes the differences to be “squeezed” so that they are less sensitive to small deviations and important information is emphasised. The ratio between the vertical and horizontal scores is calculated to understand the orientation features of potential microcalcifications.

\begin{aligned} ratio = verScore / horScore \end{aligned}

(8)

The vertical and horizontal scores, enhanced by the sigmoid function, are used in a slope-based detection approach to identify potential microcalcifications. The adaptive thresholds for vertical and horizontal scores guide the detection process and make the algorithm adaptable to variations in image properties. Given a pixel with intensity suspect and a local window W around this pixel, the adaptive threshold Tadaptive is calculated using the mean (µ) and standard deviation (σ) of the intensities in W.

\begin{aligned} Tadaptive = μ + σ adaptSlopeThreshold \end{aligned}

(9)

The adaptSlopeThreshold function refers to the dynamic adaptation of the slope threshold in the microcalcification detection algorithm based on the local features of the image. The purpose of the adaptive slope threshold is to dynamically adjust the sensitivity of the microcalcification detection based on the local characteristics of the image. Higher pixel intensities lead to a larger scaling factor and thus to a higher adaptive slope threshold. This means that regions with higher intensities require a stronger intensity difference for a pixel to be considered a potential microcalcification. This adaptive approach is crucial for improving the sensitivity of microcalcification detection in mammography images, as it allows the algorithm to adapt to different intensity differences in different regions.

3.4 Implementation of the SAAR algorithm on edge device

The implementation of the proposed SAAR algorithm on edge devices involves a carefully designed process that ensures efficient execution. The algorithm is designed to take advantage of edge computing and provide real-time processing, reduced latency and improved privacy. Below we explain the specific steps required to deploy the SAAR algorithm on edge devices.

To run the SAAR algorithm on edge devices, the system is configured with the necessary hardware and software components. Edge devices, such as special processors or microcontrollers, are equipped with the necessary computing resources. The SAAR algorithm is modified to meet the limitations and capabilities of the edge devices. This adaptation includes optimising the algorithm for efficient resource usage and ensuring compatibility with the processing power and memory limitations of edge devices. Edge devices allow algorithms to run locally so that data does not need to be transferred to a centralised server. The SAAR algorithm takes advantage of this capability and enables on-device processing of mammography images. This local execution improves response times, a critical factor in diagnostic urgency.

The SAAR algorithm is seamlessly integrated into the architecture of edge devices and utilises their unique characteristics. This integration involves tuning the algorithm to the operating environment of the edge device to ensure optimal performance. Fine-tuning of the algorithm parameters is performed to optimise its performance on edge devices. Parameters such as intensity thresholds and adaptive slope thresholds are adjusted to suit the specific characteristics of mammography images processed on edge devices.

By following these steps, the SAAR algorithm is successfully executed on edge devices, demonstrating its applicability in improving the efficiency of detecting microcalcifications in mammography images within a decentralised edge computing framework. The main goal of this paper is to propose an efficient algorithm that can be executed on edge devices and to propose the concept of integrating edge devices into mammography. Although we present a general concept, our research is specifically based on the NVIDIA Jetson Nano.

The NVIDIA Jetson Nano is a type of edge device specifically designed for real-time processing and deep learning on the device. This edge device is ideal for edge computing as it has a Graphics Processing Unit (GPU) that can accelerate deep learning. Despite the limited resources compared to server farms, the NVIDIA Jetson Nano remains powerful for edge devices. The SAAR algorithm, designed to run efficiently on edge devices, can utilise the NVIDIA Jetson Nano's resources for fast and efficient detection of microcalcifications.

The NVIDIA Jetson Nano utilises the CUDA architecture, which enables GPU processing programmability. The SAAR algorithm can be implemented and optimised on the CUDA architecture to take advantage of parallel GPU execution. When implementing the SAAR algorithm on the NVIDIA Jetson Nano, it is important to optimise the code for operation on a device with limited resources. Here are some basic parameters of the NVIDIA Jetson Nano platform⁶³:

GPU (Graphics Processing Unit): NVIDIA Maxwell architecture with 128 CUDA cores.

CPU (Central Processing Unit): Quad-core ARM Cortex-A57 processor running at 1.43 GHz.

RAM (Random Access Memory): 4 GB LPDDR4.

Storage: MicroSD card slot supporting cards up to 256 GB.

Network Connectivity: Gigabit Ethernet.

USB Ports: 4 × USB 3.0, 1 × USB 2.0 Micro-B.

Video I/O: HDMI and eDP (embedded DisplayPort) outputs.

GPIO (General-Purpose Input/Output): 40-pin GPIO header.

Dimensions: 100 mm × 80 mm.

Operating System: Support for Ubuntu-based Linux distribution.

4 Results and discussion

In this study, we proposed a method for identifying microcalcifications in mammography images by applying adaptive gradient computation using sigmoid functions to regions of interest defined by contours. The proposed SAAR algorithm consists of several stages, as outlined below.

First, an empty list is initialized to store the resulting images. For each detected contour, a binary mask is created to isolate microcalcifications. Then, adaptive slope calculation using sigmoid functions is applied to each binary mask. The slope function receives the Region of Interest (ROI), a predefined intensity value and a slope threshold as inputs. It iterates over all pixels in the ROI and considers only those below the given threshold. The pseudocode of the SAAR algorithm is shown in Figure 3.

Figure 3.

The pseudocode of the proposed SAAR algorithm.

Adaptive scores are calculated vertically and horizontally for each potential microcalcification using sigmoidal functions. The adaptive threshold for slope is introduced using sigmoid functions so that the threshold is dynamic based on pixel intensity. The final result is calculated using the adaptive threshold and the resulting matrix is returned. Figure 4 presents intermediate results for each iteration, highlighting the algorithm's progress. Figure 5 shows the final output, where regions with detected microcalcifications are clearly marked.

Figure 4.

Progress of the microcalcification detection using SAAR algorithm.

Figure 5.

The result of the final processing stage of the proposed SAAR algorithm.

Figure 6 illustrates additional experimental results that further demonstrate the performance of the proposed algorithm. In Figure 6 (a), a single cluster of microcalcifications is clearly visible and has been successfully detected. In Figure 6 (b), three distinct clusters are present; the algorithm correctly identified two of them, while one remained undetected. Additionally, one false positive region was marked, although it does not contain microcalcifications. These examples highlight both the effectiveness and the limitations of the method when applied to more complex cases.

Figure 6.

Additional examples of microcalcification detection: (a) successful identification of a single cluster; (b) partial detection in an image with three clusters - two were identified, one was missed, and one false positive region was detected.

The customized SAAR algorithm was rigorously tested on edge devices to evaluate its real-world performance. Adaptive slope-based detection that adjusts thresholds based on local image features proved more effective than conventional techniques, which often struggle with intensity variations and noise. The SAAR algorithm's adaptability enables it to cope with diverse image characteristics and enhances detection sensitivity. Dynamic thresholding ensures responsiveness to different intensity profiles, increasing microcalcification detection accuracy, a crucial factor in the early detection of breast abnormalities. Fine-tuning parameters, including predefined intensity values and adaptive slope thresholds, plays a critical role. Our findings show that optimal parameter selection significantly influences the algorithm's effectiveness, while its flexibility ensures broad applicability across diverse mammography datasets.

For the purposes of evaluation, a total of 46 mammographic images were selected from the MIAS database. The dataset includes 23 normal cases, i.e., mammograms without the presence of microcalcifications, and 23 abnormal cases containing visible microcalcifications. Within the abnormal cases, a total of 28 clusters of microcalcifications were identified, indicating that several mammograms contain more than one distinct cluster. Consequently, the evaluation was conducted on a dataset consisting of 28 microcalcification clusters and 23 normal mammograms, enabling balanced testing of the proposed detection method on both positive and negative examples.

The recognition performance is quantitatively assessed using the following evaluation metrics: Accuracy (AC), Sensitivity (SE), Specificity (SP), Precision and the F1 score. These measures are calculated based on classification outcomes including True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). The respective formulas are as follows:

Accuracy represents the proportion of correctly classified instances (both positive and negative) among all evaluated samples.

A C = \frac{T P + T N}{T P + T N + F P + F N}

Sensitivity, also referred to as recall or true positive rate, measures the model's ability to correctly identify positive instances, specifically microcalcifications in this context.

S E = \frac{T P}{T P + F N}

Specificity quantifies the model's ability to correctly identify negative cases, that is, to recognize normal images as non-pathological.

S P = \frac{T N}{T N + F P}

Precision denotes the proportion of correctly identified positive cases (true positives) among all instances that the system has classified as positive.

P r e c i s i o n = \frac{T P}{T P + F P}

F1 score represents the harmonic mean of Precision and Sensitivity, offering a single metric that balances the trade-off between these two measures. It is particularly useful when it is important to maintain a balance between correctly identifying true cases (sensitivity) and ensuring that the positive predictions are accurate (precision).

F 1 = 2 \times \frac{P r e c i s i o n \times R e c a l l}{P r e c i s i o n + R e c a l l}

Table 1 provides a summary of the performance achieved by the recognition model. The algorithm successfully identified 22 clusters of microcalcifications (TP), but failed to detect 6 (FN), and misclassified 2 normal images as positive (FP). Additionally, it correctly recognized 21 normal mammograms as negative (TN). The obtained results, including accuracy of 84.31 percent, sensitivity of 78.57 percent, and specificity of 91.3 percent, demonstrate a balanced and clinically meaningful detection capability. The high precision of 91.67 percent suggests that most detected clusters are indeed true positives, while the F1 score of 84.62 percent indicates a good compromise between precision and sensitivity. These values highlight the reliability and efficiency of the SAAR algorithm in real world scenarios, particularly when considering the constraints of edge computing environments.

Table 1.

Performance evaluation results.

Component	Value	Metric	Value (%)
True Positives (TP)	22	Accuracy	84.31%
True Negatives (TN)	21	Sensitivity	78.57%
False Positives (FP)	2	Specificity	91.3%
False Negatives (FN)	6	Precision	91.67%
		F1 Score	84.62%

True positives and false negatives were defined at the microcalcification cluster level (28 annotated clusters), while true negatives and false positives were defined at the normal mammogram level (23 normal images), ensuring methodological consistency between positive and negative case evaluation.

To further strengthen statistical validity, 95% confidence intervals were calculated using the Wilson score method, which is appropriate for binomial proportions and small sample sizes. Sensitivity (22/28) yielded a 95% confidence interval of 60.46%–89.79%, while specificity (21/23) produced a 95% confidence interval of 73.20%–97.58%. Precision (22/24) showed a 95% confidence interval of 74.15%–97.68%, and overall accuracy (43/51) demonstrated a 95% confidence interval of 71.99%–91.83%. Agreement beyond chance was assessed using Cohen's kappa coefficient (κ = 0.688), indicating substantial agreement. Additionally, the Matthews correlation coefficient (MCC = 0.697) confirmed balanced classifier performance even in the presence of class distribution differences. These additional statistical measures further support the robustness and reliability of the proposed SAAR algorithm under realistic evaluation conditions.

Minimizing false negatives is especially important in medical imaging, as missed microcalcifications could delay diagnosis. On edge devices, careful adjustment of detection thresholds is required to balance sensitivity and processing efficiency. Likewise, the presence of false positives must be managed to maintain user trust and avoid unnecessary interventions. High specificity and precision contribute to system reliability, especially in environments where fast and dependable feedback is essential.

The application of adaptive thresholding introduces flexibility, allowing system performance to be tailored to specific deployment environments. In constrained hardware scenarios, a certain compromise may be accepted, favoring sensitivity instead of specificity, to ensure that clinically relevant findings are not overlooked. This balance supports early stage identification and enables timely responses, even under limited computational conditions.

Although other methods may achieve higher detection accuracy, they often rely on high-performance servers and demand extensive processing time, which hinders deployment in time-sensitive or resource-limited settings. In contrast, the proposed approach leverages the SAAR algorithm on edge devices, allowing real-time analysis without the need for remote infrastructure. This makes the method particularly valuable in low-resource environments, such as mobile clinics or facilities without stable internet access. Modern edge platforms, such as the NVIDIA Jetson Nano, offer sufficient processing power for image analysis, while maintaining low energy consumption and cost-efficiency, and ensuring patient data privacy.

Furthermore, while many published studies have applied sophisticated algorithms, few have considered constraints typical of edge computing. Our work addresses this gap by delivering a lightweight, adaptive algorithm capable of reliable real-time performance under limited hardware conditions. Due to fundamental differences in architecture and requirements between conventional server-based solutions and edge-oriented systems, direct comparative analysis was intentionally avoided. Instead, we focused on highlighting the advantages and trade-offs of our approach, emphasizing its applicability in real-world diagnostic workflows.

The promising results achieved through the implementation of the SAAR algorithm demonstrate that effective detection is possible with appropriate algorithmic support. Continued refinement of thresholding strategies, particularly those that take device-specific limitations into account, will further improve system performance. Additionally, incorporating lightweight machine learning models, where feasible, may help boost precision and recall, ensuring that the system remains responsive and accurate in real-time clinical use.

5 Conclusions

Detecting breast cancer, particularly identifying microcalcifications in mammograms, is a critical step in early diagnosis and improving clinical outcomes. However, analyzing mammographic images presents significant challenges, as microcalcifications are often subtle, small and may be obscured by dense breast tissue or image noise and artifacts, making their detection difficult. The proposed Sigmoidal Slope Analysis and Aspect Ratio Evaluation (SAAR) algorithm offers a promising approach to overcoming these challenges by enhancing sensitivity and enabling real-time analysis on edge devices.

By integrating this algorithm with edge computing, we introduce a method that facilitates faster diagnostic workflows while preserving patient privacy and ensuring data integrity. This study emphasizes the importance of adapting diagnostic methods to local processing, as edge devices reduce the need for extensive data transfers, ease network congestion and improve responsiveness. Furthermore, the ability to execute complex algorithms directly on edge devices addresses the limitations of relying on centralized server infrastructures, particularly in resource-constrained environments.

Nonetheless, while the SAAR algorithm has shown encouraging results, certain limitations remain. Future research will focus on improving its robustness through dynamic parameter adjustments tailored to specific image characteristics. We also plan to explore more advanced machine learning models to strike a balance between detection accuracy and real-time execution. Additionally, further development will address the optimization of image transfer protocols from mammography systems to edge devices, along with the implementation of enhanced security mechanisms to safeguard patient data.

In conclusion, the integration of edge computing into breast cancer diagnostics opens exciting opportunities for fast, cost-effective and secure diagnostic solutions that can be deployed across both clinical and remote settings, contributing to more accessible and efficient healthcare.

Footnotes

Acknowledgment

Work presented in this article was supported by projects No. 451-03-34/2026-03/ 200122 and No. 451-03-33/2026-03/ 200132, financed by the Ministry of Science, Technological Development and Innovation of the Republic of Serbia.

Ethical considerations

This study used publicly available anonymised mammographic images from the MIAS database. No additional patient data were collected. Therefore, ethical approval was not required in accordance with institutional and national guidelines.

Consent to participate

Not applicable. The study is based on publicly available anonymised data.

Author contribution

All authors contributed equally to this work and approved the final version of the manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The datasets analysed during the current study are publicly available in the Mammographic Image Analysis Society (MIAS) repository. The processed data and implementation details of the SAAR algorithm are available from the corresponding author upon reasonable request.

ORCID iDs

Aleksandar Peulic

Zeljko Jovanovic

Marina Milosevic

References

Muduli

Dash

Majhi

. Automated breast cancer detection in digital mammograms: a moth flame optimization based ELM approach. Biomed Signal Process Control 2020; 59: 101912.

Shen

, et al. Optimal breast tumor diagnosis using discrete wavelet transform and deep belief network based on improved sunﬂower optimization method. Biomed Signal Process Control 2020; 60: 101953.

Casti

Mencattini

Salmeri

, et al. Contour-independent detection and classification of mammographic lesions. Biomed Signal Process Control 2016; 25: 165–177.

Milosevic

Jankovic

Peulic

. Segmentation for the enhancement of microcalcifications in digital mammograms. Technol Health Care 2014; 22: 701–715.

Zhao

Hua

Geng

, et al. A shortcut weighted fusion pyramid network for microcalcification detection in breast mammograms. Technol Health Care 2023; 31: 841–853.

Wang

Feng

Wang

. Grouped fuzzy SVM with EM-based partition of sample space for clustered microcalcification detection. Technol Health Care 2017; 25: 325–336.

De Santo

Molinara

Tortorella

, et al. Automatic classification of clustered microcalcifications by a multipleexpert system. Pattern Recognit 2003; 36: 1467–1477.

Rehman

Pei

, et al. Computer vision-based microcalcification detection in digital mammograms using fully connected depthwise separable convolutional neural network. Sensors 2021; 21: 4854.

Lee

Park

Jeon

, et al. Noise removal in medical mammography images using fast non-local means denoising algorithm for early breast cancer detection: a phantom study. Optik (Stuttg) 2019; 180: 569–575.

10.

Nasser

Yusof

. Deep learning based methods for breast cancer diagnosis: a systematic review and future direction. Diagnostics (Basel) 2023; 13: 161.

11.

Shi

Cao

Zhang

, et al. Edge computing: vision and challenges. IEEE Internet Things J 2016; 3: 637–646.

12.

Yoshida

Doi

Nishikawa

, et al. Application of the wavelet transform to automated detection of clustered microcalcifications in digital mammograms. Acad Rep Tokyo Inst Polytech 1994; 16: 24–37.

13.

Yoshida

Nishikawa

Maryellen

, et al. Computer-aided diagnosis in mammography: Detection of clustered microcalcifications based on multiscale edge representation. Computer Assisted Radiology 1996; 1091: 390–395.

14.

Boccignonea

Chianese

Picariello

. Computer aided detection of microcalcifications in digital mammograms. Comput Biol Med 2000; 30: 267–286.

15.

Zhang

Homma

Goto

, et al. A hybrid image filtering method for computer-aided detection of microcalcification clusters in mammograms. J Med Eng 2013; 2013: 615254.

16.

Lahmiri

. A wavelet-wavelet based processing approach for microcalcifications detection in mammograms. J Adv Inf Technol 2012; 3: 162–167.

17.

Subash Chandra Bose

Shankar Kumar

Karnan

. Detection of microcalcification in mammograms using soft computing techniques. Eur J Sci Res 2012; 86: 103–122.

18.

Wang

Xin

Zhang

, et al. Breast microcalcifications detection based on fusing features with DTCWT. J X-Ray Sci Technol 2020; 28: 197–218.

19.

Woods

Doss

Bowyer

, et al. Comparative evaluation of pattern recognition techniques for detection of microcalcifications in mammography. Int J Pattern Recognit Artif Intell 1994; 7: 1417–1436.

20.

El-Naqa

Yang

Wernick

, et al. A support vector machine approach for detection of microcalcifications. IEEE Trans Med Imaging 2002; 21: 1552–1563.

21.

Rízzí

D’Aloia

Castagnolo

. A supervised method for microcalcification cluster diagnosis. Integr Comput Aided Eng 2013; 20: 157–167.

22.

Verma

Zakos

. A computer-aided diagnosis system for digital mammograms based on fuzzy-neural and feature extraction techniques. IEEE Trans Inf Technol Biomed 2001; 5: 46–54.

23.

Cheng

Lui

Feiimanis

. A novel approach to microcalcification detection using fuzzy logic techniques. IEEE Trans Med Imaging 1998; 17: 442–450.

24.

Nagel

Nishikawa

Papaioannou

, et al. Analysis of methods for reducing false positives in automated detection of clustered microcalcifications in mammograms. Med Phys 1998; 25: 1502–1506.

25.

Papadopoulosa

Fotiadis

Likas

. An automatic microcalcification detection system based on a hybrid neural network classifier. Artif Intell Med 2002; 25: 149–167.

26.

Wang

Nishikawa

Yang

. Global detection approach for clustered microcalcifications in mammograms using a deep learning network. J Med Imaging 2017; 4: 024501.

27.

Wang

Yang

. A context-sensitive deep learning approach for microcalcification detection in mammograms. Pattern Recognit 2018; 78: 12–22.

28.

Valvano

Santini

Martini

, et al. Convolutional neural networks for the segmentation of microcalcification in mammography imaging. J Healthc Eng 2019; 9360941: 1–9.

29.

Cai

Huang

Rong

, et al. Breast microcalcification diagnosis using deep convolutional neural network from digital mammograms. Comput Math Methods Med 2019; 2717454: 1–10.

30.

Wang

Yang

Cai

, et al. Discrimination of breast cancer with microcalcifications on mammography by deep learning. Sci Rep 2016; 6: 27327.

31.

Kim

Park

. Statistical textural features for detection of microcalcifications in digitized mammograms. IEEE Trans Med Imaging 1999; 18: 231–238.

32.

Pesapane

Trentin

Ferrari

, et al. Deep learning performance for detection and classification of microcalcifications on mammography. Eur Radiol Exp 2023; 7: 69.

33.

Makeev

Glick

. Task-based assessment of digital mammography microcalcification detection with deep learning denoising algorithms using in silico and physical phantom studies. J Med Imaging 2023; 10: 053502.

34.

Doma

Padmanandam

Tambvekar

, et al. Artificial intelligence-based breast cancer detection using WPSO. Int J Oper Res Inf Syst 2022; 13: 1–16.

35.

Schonenberger

Hejduk

Ciritsis

, et al. Classification of mammographic breast microcalcifications using a deep convolutional neural network: a BI-RADS-based approach. Invest Radiol 2021; 56: 224–231.

36.

Liu

Chen

Zhang

, et al. A deep learning model integrating mammography and clinical factors facilitates the malignancy prediction of BI-RADS 4 microcalcifications in breast cancer screening. Eur Radiol 2021; 31: 5902–5912.

37.

Stelzer

Steding

Raudner

, et al. Combined texture analysis and machine learning in suspicious calcifications detected by mammography: potential to avoid unnecessary stereotactical biopsies. Eur J Radiol 2020; 132: 109309.

38.

Mansour

Mokhtar

El Galil

MASMA

, et al. Artificial intelligence reading digital mammogram: enhancing detection and differentiation of suspicious microcalcifications. Br J Radiol 2025; 98: 246–253.

39.

Aphinives

Nawapan

. Artificial intelligence development for detecting microcalcification within mammography. J Med Assoc Thail 2021; 104: 1.

40.

Huang

. Detection of microcalcifications in digital mammograms using wavelet filter and markov random field model. Comput Med Imaging Graph 2006; 30: 163–173.

41.

Marrocco

Molinara

D’Elia

, et al. A computer-aided detection system for clustered microcalcifications. Artif Intell Med 2010; 50: 23–32.

42.

Quellec

Lamard

Cozic

, et al. Multiple-Instance learning for anomaly detection in digital mammography. IEEE Trans Med Imaging 2016; 35: 1604–1614.

43.

Guo

, et al. Uncertainty ordinal multi-instance learning for breast cancer diagnosis. Healthcare 2022; 10: 2300.

44.

Zhang

Xie

, et al. A multi-instance networks with multiple views for classification of mammograms. Neurocomputing 2021; 443: 320–328.

45.

Elmoufidi

. Deep multiple instance learning for automatic breast cancer assessment using digital mammography. IEEE Trans Instrum Meas 2022; 71: 4503813.

46.

Zhang

Zhu

, et al. Fast vibration characteristics analysis of an underwater shield tunnel using the accelerometer network enhanced by edge computing. Measurement ( Mahwah N J) 2019; 141: 52.

47.

Ray

Dash

. Edge computing for internet of things: a survey, e-healthcare case study and future direction. J Netw Comput Appl 2019; 140: 1.

48.

Satyanarayanan

Simoens

Xiao

, et al. Edge analytics in the internet of things. IEEE Pervasive Comput 2017; 16: 24–31.

49.

Dinh

Lee

Niyato

, et al. A survey of mobile cloud computing: architecture, applications, and approaches. Wirel Commun Mob Comput 2013; 13: 1587–1611.

50.

Mao

You

Zhang

, et al. A survey on Mobile edge computing: the communication perspective. IEEE Commun Surv Tutorials 2017; 19: 2322–2358.

51.

Chen

Lou

, et al. Intelligent edge computing based on machine learning for smart city. Future Gener Comput Syst 2021; 115: 90.

52.

Elhanashi

Dini

Saponara

, et al. Integration of deep learning into the IoT: a survey of techniques and challenges for real-world applications. Electronics (Basel) 2023; 12: 4925.

53.

Rostami

Tarvirdizadeh

Alipour

, et al. Real-time stress detection from raw noisy PPG signals using LSTM model leveraging TinyML. Arab J Sci Eng 2024; 50: 6959–6981.

54.

Suckling

Parker

Dance

, et al. The mammographic images analysis society digital mammogram database. Exerpta Medica 1994; 1069: 375–378.

55.

Milosevic

Jankovic

Peulic

. Comparative analysis of breast cancer detection in mammograms and thermograms. Biomed. Eng. / Biomed. Tech 2015; 60: 49–56.

56.

Milosevic

Jankovic

Peulic

. Thermography based breast cancer detection using texture features and Minimum variance quantization. EXCLI J 2014; 13: 1204–1215.

57.

Ellinas

Mandadelis

Tzortzis

, et al. Image de-noising using wavelets. T.E.I. Piraeus Appl Res Rev 2004; 9: 97–109.

58.

Rangarajan

Venkataramanan

Shah

. Image Denoising Using Wavelets. Technical report. College of Engineering, University of Michigan 2002.

59.

Canny

. A computational approach to edge detection. IEEE Trans Pattern Anal Mach Intell 1986; 8: 679–698.

60.

Zhou

Wang

. An improved canny algorithm for edge detection. J Comput Inf Syst 2011; 7: 1516–1523.

61.

Chen

, et al. AI Computing Systems: An Application Driven Perspective. Amsterdam, Netherlands: Elsevier, 2023, ISBN 978-0-323-95399-3.

62.

Chanussot

Lambert

. Adaptive filtering and change detection in hyperspectral imagery. J Remote Sens 2022; 25: 123–145.

63.

NVIDIA Jetson Nano Developer Kit. 2022. Available from: https://developer.nvidia.com/jetson-nano