BREAST-CAD: A Computer-Aided Diagnosis System for Breast Cancer Detection Using Machine Learning

Abstract

This research presents a novel Computer-Aided Diagnosis (CAD) system called BREAST-CAD, developed to support clinicians in breast cancer detection. Our approach follows a three-phase methodology: Initially, a comprehensive literature review between 2000 and 2024 informed the choice of a suitable dataset and the selection of Naive Bayes (NB), K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Decision Trees (DT) Machine Learning (ML) algorithms. Subsequently, the dataset was preprocessed and the four ML models were trained and validated, with the DT model achieving superior accuracy. We developed a novel, integrated client–server architecture for real-time diagnostic support, an aspect often underexplored in the current CAD literature. In the final phase, the DT model was embedded within a user-friendly client application, empowering clinicians to input patient diagnostic data directly and receive immediate, AI-driven predictions of cancer probability, with results securely transmitted and managed by a dedicated server, facilitating remote access and centralized data storage and ensuring data integrity.

1. Introduction

Fine Needle Aspiration (FNA) is a diagnostic approach commonly used for the cytological evaluation of breast lesions. Usually, FNA is carried out by extracting part of the infected breast using a fine needle, followed by analysis to determine the presence of malignant cells. Among its primary advantages are its simplicity, low risk, quick turnaround time, and suitability for outpatient settings. However, limitations include operator dependency, potential sampling errors, and lower sensitivity compared to core biopsy in certain cases, especially in non-palpable or cystic lesions. Recent developments have focused on enhancing FNA accuracy and clinical utility by integrating IoT technologies and ML algorithms into diagnostic workflows [1]. This integration allows digitized cytological slides and real-time procedural data to be transmitted via IoT-enabled microscopes or mobile devices to cloud-based ML models trained to classify benign and malignant samples, thus supporting remote analysis and automated decision-making.

The Wisconsin Diagnostic Breast Cancer (WDBC) is a publicly available dataset, and it serves as a significant enabler for these models. It contains digitized features derived from FNA tests of breast masses. The dataset encompasses 569 samples, with each sample defined by 30 numerical features related to cell nucleus properties, and each sample labeled to indicate whether the mass is benign or malignant. Its standardized format and balance of diagnostic classes make it highly suitable for training and benchmarking ML models [2]. These advancements highlight the potential of combining IoT and ML with traditional diagnostic techniques like FNA to deliver scalable, accurate, and accessible breast cancer diagnostic systems.

Despite the growing interest in leveraging ML and IoT technologies for medical diagnostics, their integrated application with FNA for breast cancer screening remains significantly underexplored. A critical review of the current literature indicates a clear research gap: while ML and IoT have individually demonstrated potential in improving diagnostic accuracy and accessibility, few studies have effectively combined these technologies within FNA-based diagnostic workflows to exploit their full synergistic benefits. In response to this identified gap, this study introduces BREAST-CAD, an IoT-enabled CAD system specifically engineered to detect breast cancer via FNA analysis [3].

The proposed system employs IoT-based data acquisition and ML-powered classification to support clinicians in informed and prompt decision-making. The development of this type of system demands careful consideration of several crucial components: the diagnostic procedure (FNA), the choice and preprocessing of high-quality breast cancer datasets, and the application of robust and interpretable classification algorithms [4]. This research advances the field by designing a modular, adaptable diagnostic model that integrates real-time IoT connectivity with ML analytics, effectively bridging the divide between clinical application and intelligent automation. The study aims to construct a comprehensive diagnostic prototype, BREAST-CAD, that not only facilitates remote screening and risk prediction but also improves diagnostic precision and operational efficiency in breast cancer care.

This research is guided by the following specific objectives:

• Comprehensive Literature Review: A systematic review of academic publications from 2000 to 2024 will be conducted to identify benchmark FNA-based breast cancer datasets and ML classification techniques.
• Dataset Selection: An appropriate breast cancer dataset will be identified. To ensure data quality, normalization, and suitability for ML analysis, the dataset will undergo a rigorous preprocessing phase.
• Development and Evaluation of ML Models: Four separate ML classification models including SVM, KNN, DT, and NB will be trained and their performance evaluated.
• Client–Server System Integration: The optimal ML model will be deployed within a client–server framework. The chosen model will be integrated into a client framework for real-time classification, connected with a centralized server for data management and sharing within healthcare authority.

The organization of the paper is as follows: Section 2 covers the research methodology, Section 3 examines Fine Needle Aspiration, Section 4 presents the results, and, finally, Section 5 provides a summary and possible future research directions.

2. Methodology

This study introduces a structured, three-phase methodology to design, develop, and evaluate the BREAST-CAD system that integrates FNA cytological data, ML classification techniques, and IoT infrastructure for real-time breast cancer diagnosis. The methodology and the workflow diagram illustrated in Figure 1 consist of the following key phases:

A. 

Phase 1: Dataset and Model Selection

The first phase focused on identifying a suitable breast cancer dataset and selecting high-performing ML classifiers. A systematic literature review used Google Scholar as the primary search engine. Boolean logic was applied with search terms including “Fine Needle Aspiration”, “Breast Cancer Dataset”, “Machine Learning”, and “2000–2024”. This process initially retrieved “1532” peer-reviewed publications, which were screened through a multi-stage filtering process based on relevance to classification models, diagnostic context, and methodological rigor. Our inclusion criteria were met by 34 studies, providing valuable perspectives on frequently employed datasets and high-performing ML in FNA-based breast cancer diagnosis. This carefully selected body of work revealed the WDBC dataset as the most prevalent and clinically significant resource, owing to its comprehensive features, actual FNA cytological measurements, and standardized format. The four top ML models chosen for implementation—SVM, DT, KNN, and NB—were selected due to their common use and consistently strong classification performance in medical diagnostics.

B. 

Phase 2: Data Preparation, Model Training, and Evaluation

Phase 2 focused on establishing the predictive core of the BREAST-CAD system. The WDBC dataset was selected for model development. The WDBC dataset includes 569 instances, each characterized by 30 FNA features. A comprehensive preprocessing pipeline was implemented. Data quality and feature consistency were addressed through a series of steps: data cleaning to handle missing values, identification of outliers, and analysis of the distribution of each feature. Feature selection was performed using correlation-based filtering to retain features most strongly associated with malignancy while reducing multicollinearity, thereby improving model interpretability and preventing overfitting. The WDBC was partitioned into training (70%), testing (20%), and IoT-based evaluation (10%). This ensured class balance and the fair representation of malignant and benign cases across each subset. The optimization of learning performance was achieved via hyperparameter tuning. A detailed evaluation was performed and the DT model was identified as the most effective through comparative analysis.

C. 

Phase 3: Real-Time Diagnosis

The final phase centered on integrating the optimal ML model within a networked diagnostic framework, resulting in the implementation of the BREAST-CAD system. This system consists of (1) a client application and (2) a centralized server. The client interface allows clinicians to input cytological FNA data, which are processed locally by the embedded DT model to generate a diagnostic prediction (benign or malignant). This output is then transmitted using the MQTT protocol to the cloud-hosted BREAST-CAD server. The server acts as a centralized node, receiving diagnostic results from multiple client nodes, enabling remote monitoring, centralized data storage, and real-time epidemiological analysis. The server’s architecture supports dynamic dashboards for healthcare administrators, offering trends in diagnostic patterns, flagging anomalies, and identifying potential geographic or demographic clusters. Furthermore, the system architecture is designed to support continuous model improvement. Aggregated data from real-world usage are anonymized and stored, forming a feedback loop that enables the ongoing retraining of ML models to enhance accuracy over time. This capability facilitates collaborative research, allowing institutions to build larger diagnostic datasets and contribute to a shared learning model.

3. Fine Needle Aspiration

FNA involves guiding a thin needle (typically of 22–27 gauge) into a breast lump that is either palpable or detected by imaging to extract cellular material for analysis. The aspirated sample is smeared on slides, stained, and evaluated under a microscope. Based on cellular features, the diagnostic output typically categorizes lesions as benign, malignant, suspicious, or non-diagnostic [5]. These include nuclear morphology (size, shape, chromatin texture), mitotic activity, nuclear membrane irregularity, cell uniformity, and background necrosis. In automated or AI-assisted systems such as those using the WDBC dataset, these features are transformed into numerical attributes used for computational classification models (e.g., SVM, KNN, DT, NB). FNA offers several advantages in breast cancer diagnosis. It is quick, cost-effective, and well tolerated by patients, making it ideal for high-volume screening or outpatient settings. It enables rapid diagnosis, often within 24 h, allowing timely clinical decisions. FNA facilitates repeated sampling and is suitable for the cytological evaluation of axillary lymph nodes or metastatic lesions. Additionally, when adequate, the material collected can be used for immunocytochemistry or molecular testing.

4. Results

4.1. Dataset and Model Selection Results

As depicted in

Table 1. Breast cancer dataset and model selection process analysis.

Reference	Datasets				Classification Techniques
	WBCO	WDBC	WPBC	MUWP	SVM	RF	ANN	DT	BT	KNN	MLPs	LR	BS	NB	GA	FL	XGboost
[6]	√	√	√		√		√	√	√			√	√
[7]				√	√						√
[8]		√								√				√
[9]		√						√
[10]		√			√					√				√
[11]		√			√	√	√	√	√	√		√
[12]		√					√
[13]		√			√										√
[14]		√					√
[15]		√					√
[16]		√			√					√
[17]		√														√
[18]		√				√		√				√
[19]				√	√			√
[20]		√			√	√		√		√		√		√
[21]	√				√				√	√				√
[22]		√												√
[23]		√			√	√				√	√	√					√
[24]		√				√				√	√				√
[25]		√				√				√		√					√
[26]		√			√			√		√		√
[27]		√			√	√										√	√
[28]		√					√								√	√
[29]		√			√					√				√
[30]		√				√			√			√
[31]		√			√	√		√		√	√	√		√
[32]		√			√	√		√									√
[33]		√			√		√							√
[34]		√						√		√		√		√
[35]	√	√	√					√		√	√			√
[36]		√			√			√		√				√
[37]		√			√
[38]		√						√		√				√
Frequent	3	27	2	2	19	11	7	13	4	17	6	11	1	12	3	3	5

and Figure 2, from the comprehensive literature review, we identified four datasets that are frequently used: the Original Wisconsin Breast Cancer dataset (WBCO), the WDBC, the Wisconsin Prognosis Breast Cancer dataset (WPBC), and the dataset from the Medical University of Wroclaw Poland (MUWP). The WDBC dataset appeared most often (in 27 studies), likely because it has a lot of different features and is a well-known standard for comparison. This widespread adoption of the WDBC facilitates comparability across different modeling approaches.

Our model selection process prioritized four key ML models, SVM, DT, KNN, and NB, based on their frequency of use in the literature, as described in Figure 3. While a broader range of models like Artificial Neural Networks (ANNs), Bayesian Statistics (BSs), Logistic Regression (LR), Feed Forward Neural Networks (FFNNs), Multilayer Perceptrons (MLPs), genetic algorithms (GAs), fuzzy logic (FL), and gradient boosting (XGboost) were identified, SVM (used in 19 studies), KNN (used in 17 studies), NB (used in 12 studies), and DT (used in 13 studies) significantly dominated in the literature, with Naive Bayes also showing notable prior usage and strong performance. These selected models offer distinct advantages: SVMs’ effect in high-dimensional spaces and in handling non-linear relationships; DT’s robustness and capacity to manage numerous predictive features; KNN’s potential to discern complex patterns; and Naive Bayes’ computational efficiency and effectiveness with independent features. Critically, these four models have consistently demonstrated strong performance metrics in previous WDBC-based breast cancer classification studies, justifying their selection for this research.

4.2. Model Training and Evaluation Results, Phase 2

4.2.1. Data Collection and Preparation Results

As in

Table 2. The WDBC dataset feature description.

No	Attribute/Feature	Description	Range
1	ID	Identification number for patients	Unique value
2	Diagnosis	Cancer status Malignant (M), Benign (B)	M or B
3	Radius	mean of distances from the center to points on the perimeter	11–27
4	Area	π(Radius)2	360–2300
5	Perimeter	2π(Radius)	71–82
6	Texture	standard deviation of gray-scale values	11–40
7	Smoothness	local variation in radius lengths	0.05–0.2
8	Compactness	(primeter)2/area	0.04–0.45
9	Concavity	severity of concave portions of the contour	0.02–0.5
10	Concave Points	number of concave portions of the contour	0.02–0.5
11	Symmetry	The length difference between lines perpendicular to the major axis to the cell boundaries in both directions	0.1–0.3
12	Fractal dimension	coastline approximation—1	0.05–0.1

, the WDBC include a specific ID, a diagnosis (benign or malignant), and 30 numerical features. These 30 features come from ten different measurements taken from cell nuclei: “radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, and fractal dimension”. The dataset provides the “mean, standard error, and worst” values for each measurement. The dataset is relatively balanced, with 357 benign (62.6%) and 212 malignant (37.4%) samples. After removing the irrelevant “Unnamed” and “ID” columns, the categorical “diagnosis” column, indicating Benign (B) or Malignant (M), was numerically encoded, with Malignant assigned 1 and Benign assigned 0. The dataset contained no missing values, leaving 30 features for model training. To assess inter-variable relationships, we employed correlation matrices, visualizing pairwise correlation coefficients. We quantified these relationships using the Pearson correlation coefficient, ranging from −1 to +1. Feature selection was performed based on a 0.6 correlation threshold with the diagnostic column, resulting in 12 selected features. Critically, we then examined correlations among these 9 chosen features to detect any remaining multicollinearity. Figure 4 displays this correlation matrix, enabling us to identify and address any high inter-predictor correlations. The pair plot shown in Figure 5 was employed to explore the distributional characteristics and interrelationships among seven features, utilizing the mean values of previously identified strongly correlated feature subsets. The diagonal elements show the marginal distributions of each feature, visualized as kernel density estimations. The off-diagonal elements, representing pairwise relationships, are mirrored across the diagonal due to the commutative nature of correlation. Outliers were addressed using the Local Outlier Factor (LOF) algorithm. Applying LOF to the nine feature sets identified outliers in three features, which were subsequently removed.

4.2.2. Machine Learning Models Training Results

The WDBC was partitioned into training (70%), testing (20%), and evaluation (10%). This division allowed us to check how well k-NN, SVM, DT, and NB performed, as well as the BREAST-CAD system itself (as shown in

Table 3. Dataset partitioning and usage for training, testing, and BREAST-CAD integration validation.

Samples	Data Section	Percentage	Usage
569	Training	70%	Training the DT, SV, and KN models
	Testing	20%	Testing the performance of the DT, SV, and KNN models
	BREAST-CAD validation	10%	Testing the integration of the DT model and the IoT BREAST-CAD client

). For k-NN, which does not have a traditional training step, we found the optimal ‘k’ value using cross-validation before testing its generalization on the separate testing set. For SVM, we scaled all features and used a Radial Basis Function (RBF) kernel, fine-tuning ‘C’ and ‘gamma’ parameters with Sequential Minimal Optimization (SMO) before training on the entire training set and evaluating on the test set. For the DT model, we used Gini impurity to recursively split data based on features, applying a maximum depth criterion to prevent overfitting. Finally, for the NB model, we employed a Gaussian NB classifier, leveraging its assumption of normally distributed features within each class, calculating means and standard deviations from the training set to determine class probabilities for new data points.

We assessed how well the k-NN, SVM, NB and DT classification models performed after being trained on the WDBC dataset by using a separate test set that the models had not seen before. To measure their performance, we looked at several metrics. The results for each of these metrics for the k-NN, SVM, NB and DT models can be seen in Figure 6, Figure 7, Figure 8, and Figure 9, respectively.

4.2.3. Comparative Analysis Results

A comparative analysis of DT, SVM, and KNN classifiers for a BREAST-CAD system revealed nuanced performance differences (Figure 10 and

Table 4. Performance results of DT, KNN, NB and SVM models.

	Accuracy	Precision	Recall	AUC	F1-Score
DT	0.97	0.97	0.95	0.97	0.96
SVM	0.96	1	0.89	0.94	0.93
KNN	0.95	1	0.89	0.94	0.93
NB	0.94	0.94	0.92	0.94	0.93

). While all models demonstrated comparable accuracy (DT: 0.97, SVM: 0.96, KNN: 0.95, NB: 0.94), indicating general proficiency, notable variations emerged in precision and recall. SVM and KNN exhibited perfect precision (1.0). The DT classifier achieved an F1-score of 0.96, which is the highest. This suggests that the DT classifier likely identified a greater proportion of true positives, even with a slight increase in false positives. Consequently, while SVM and KNN may be preferable in contexts prioritizing the minimization of false positives, the DT classifier appears to offer a superior compromise between precision and recall, potentially making it a more robust choice for the BREAST-CAD system design, particularly given the importance of maximizing true identifications in medical diagnostic applications.

4.3. The BREAST-CAD Design Results

This phase of the research integrates a comprehensive three-stage methodology, as depicted in Figure 11. Initially, a GUI-based program was developed to function as the BREAST-CAD client, as illustrated in Figure 12. This client consolidates a pre-trained DT model, forming a unified breast cancer diagnostic framework.

The WDBC dataset, though integral to this phase, contains only FNA diagnostic results. To enhance the system’s utility beyond the dataset’s scope, the design incorporates additional variables that influence the risk of breast cancer. These factors include both modifiable (e.g., “alcohol consumption, smoking, physical inactivity, obesity, dietary habits”) and non-modifiable (e.g., “genetic mutations like BRCA1/BRCA2, family history, personal history of benign breast conditions”) determinants. In this context, clinicians are prompted to input both the FNA diagnostic results and these relevant modifiable and non-modifiable risk factors into the BREAST-CAD client.

In the subsequent stage, the second phase of the system architecture entails the establishment of a secure client connection, facilitating the transmission of patient data to the BREAST-CAD server. As depicted in Figure 13, this connection is established by the BREAST-CAD client, initiating a connection request to the designated BREAST-CAD server, commonly referred to as the MQTT broker [39]. The client transmits a CONNECT message containing essential credentials and connection parameters, such as client ID and keep-alive intervals, to the broker. Upon receipt of this request, the broker authenticates the client’s credentials and connection settings, responding with a CONNACK message that indicates the success or failure of the connection. Once the connection is established, the client maintains a persistent session, subscribing to relevant topics, and the broker facilitates message publishing and reception as per the subscription protocols. The connection remains alive through periodic keep-alive signals, ensuring continuous communication between the client and server.

Upon the establishment of the connection, the BREAST-CAD client program formulates a JSON (JavaScript Object Notation) data packet, encapsulating the patient’s information and FNA analysis results, as shown in Figure 14. This packet is subsequently transmitted to the BREAST-CAD server. Upon receipt at the server, the JSON packet is unpacked, with patient data extracted and stored in the corresponding hospital’s dataset. Figure 15 illustrates the successful publication of data for 10% of the WDBC dataset, showcasing the integration of the BREAST-CAD client with the server and providing statistical analysis of FNA features about breast cancer malignancy. This multi-stage process facilitates real-time data transmission and processing, demonstrating the potential of networked diagnostic systems in breast cancer detection.

5. Conclusions and Future Work

This research introduces BREAST-CAD, a novel CAD system designed to aid in breast cancer detection through ML and real-time remote support. The study followed a clear three-step process. First, we conducted a comprehensive literature review between 2000 and 2024 to select the WDBC dataset and ML models: DT, SVM, KNN, and NB. After carefully preparing and testing these models, the DT model performed the best, achieving an accuracy of 97% and an F1-score of 0.96. This top-performing model was then built into a user-friendly client application.

This client application allows clinicians to input cell characteristics and relevant risk factors, with the diagnoses securely transmitted to a central server using MQTT, enabling real-time diagnostic support from anywhere. While BREAST-CAD presents an innovative approach to enhancing diagnostic capabilities, it does have some limitations. It currently relies on a single dataset, has not yet undergone validation in a real clinical setting, and requires manual input of risk factors; furthermore, the transmission of sensitive data necessitates the further enhancement of security measures. Future work will focus on extensive clinical validation in hospital environments, incorporating more diverse datasets (including imaging data), automating feature extraction, bolstering system security, and integrating adaptive learning techniques to improve scalability and diagnostic accuracy over time.