Breast Tumor Classification Using Machine Learning (KNN) and Deep Learning (CNN)

Abstract

Breast cancer is one of the most common cancers affecting women worldwide, and early detection is vital for improving patient outcomes. Ultrasound imaging is frequently used as a diagnostic tool because it is non-invasive, cost-effective, and safer compared to modalities such as mammography, particularly for younger patients with dense breast tissue. However, interpretation of ultrasound images relies heavily on radiologist expertise, which can be subjective and time-consuming. Artificial intelligence offers promising solutions to address these challenges by providing automated, reliable, and efficient tumor classification. This study focuses on the classification of breast tumors in ultrasound images into benign and malignant categories using two approaches: K-Nearest Neighbor (KNN) and Convolutional Neural Networks (CNN). KNN, a classical machine learning algorithm, is applied on handcrafted texture features extracted from Regions of Interest (ROIs) using the Gray-Level Co-occurrence Matrix (GLCM). Features such as contrast, correlation, energy, and homogeneity capture textural variations within the tumor region. CNN, on the other hand, is a deep learning model that learns discriminative spatial and structural patterns directly from the ultrasound images, reducing the need for manual feature engineering. Both models are evaluated on breast ultrasound images with corresponding masks that precisely define the tumor boundaries. Performance is assessed using widely accepted metrics such as accuracy, sensitivity, specificity, and F1-score. KNN provides a strong baseline with interpretable results, while CNN demonstrates the ability to automatically learn complex features that may be difficult to handcraft. The work emphasizes the potential of combining machine learning and deep learning approaches to develop robust computer-aided diagnosis systems based on ultrasound imaging. Such systems can assist clinicians in achieving faster, more consistent, and more accurate breast cancer diagnosis, ultimately contributing to better patient care

Introduction

• Breast cancer is the most common cancer among women and a major cause of cancer-related deaths.

• Early and accurate diagnosis is crucial for improving survival rates.

• Ultrasound imaging complements mammography, especially for women with dense breast tissue, due to its non-invasive, real-time, and radiation-free nature.

• However, ultrasound interpretation is subjective, depending heavily on radiologist experience.

• To improve objectivity and accuracy, computer-aided diagnosis (CAD) systems are being developed using machine learning (ML) and deep learning (DL).

2. Research Objective

• This study compares K-Nearest Neighbor (KNN) and Convolutional Neural Network (CNN) models for breast tumor classification using ultrasound images.

• KNN uses GLCM-based texture features; CNN directly processes raw images.

• Aim: Evaluate strengths, weaknesses, and clinical applicability of both models.

3. Literature Review

• Early CAD relied on handcrafted features (e.g., GLCM) and simple classifiers like KNN, SVM, and Naïve Bayes.

• KNN is interpretable, simple, and effective for small datasets.

• Deep learning (CNNs) now dominates due to its ability to learn features automatically.

• Transfer learning using pretrained networks (e.g., VGG, ResNet) improves accuracy with small datasets.

• Segmentation masks, attention mechanisms, and hybrid models improve robustness.

• Explainable AI (XAI) techniques like Grad-CAM are crucial for clinical trust.

• Despite CNN advances, KNN remains useful in resource-limited settings.

• The BUSI dataset is commonly used and includes labeled ultrasound images with segmentation masks.

4. Methodology

Dataset

• BUSI dataset (Al-Dhabyani et al., 2020): 780 ultrasound images from 600 female patients.

• Classes: Normal, Benign, and Malignant.

• Preprocessing:

  • Resize to 224×224 pixels.

  • Median filtering to reduce speckle noise.

  • Normalization and image masking to isolate tumor regions.

  • Dataset split: 70% training, 15% validation, 15% testing.

KNN Classification

• Features extracted from original and masked images using GLCM (e.g., contrast, correlation, energy).

• Features normalized (z-score).

• Classification using KNN (Euclidean distance); optimal k found via validation.

• Evaluated with confusion matrix, sensitivity, specificity, and F1-score.

CNN Classification

• CNN trained on original images using:

  • Convolutional layers + ReLU.

  • Max pooling + dropout regularization.

  • Softmax for three-class output.

• Training: 50 epochs, batch size 32, Adam optimizer.

• Data augmentation used for generalization (rotation, flipping, zoom).

5. Results and Discussion

KNN Results

• Achieved 100% accuracy across all three classes.

• Confusion matrix, sensitivity, specificity, and F1-score all reported as 100%.

• Masked and ROI images improved classification by focusing on tumor areas.

• Visual validation confirmed perfect alignment between predictions and ground truth.

CNN Results

• Not explicitly detailed in the provided text, but implied to be included in comparative evaluation.

Conclusion

This study demonstrated the effectiveness of machine learning and deep learning techniques for the classification of breast ultrasound images into benign, malignant, and normal categories. Two approaches, K-Nearest Neighbor (KNN) and Convolutional Neural Network (CNN), were employed and their performance evaluated using sensitivity, specificity, F1-score, and confusion matrix analysis. The results showed that both KNN and CNN achieved 100% classification accuracy, with all samples correctly classified. The KNN model, despite its algorithmic simplicity, performed remarkably well when applied to both masked and original images, highlighting its suitability for problems with well-defined feature spaces. The CNN model further reinforced its superiority in automated feature learning, capturing spatial and textural patterns in ultrasound images and producing highly reliable classification outputs. Additionally, tumor mask overlays and prediction visualizations confirmed the precision of CNN in identifying and localizing abnormal regions. These findings indicate that both classical machine learning and modern deep learning techniques have significant potential in assisting radiologists with breast cancer diagnosis. KNN offers a lightweight solution with low computational demands, whereas CNN provides greater scalability and robustness for clinical applications involving large and diverse image sets. Although the current evaluation demonstrated exceptional performance, further research is necessary to validate these results on broader datasets with varying imaging conditions to ensure generalizability. In conclusion, the integration of KNN and CNN models in ultrasound-based breast cancer diagnosis highlights the transformative potential of artificial intelligence in medical imaging. By reducing diagnostic errors and providing consistent decision support, such systems can greatly enhance clinical workflows and improve patient outcomes. Future work will focus on expanding model validation, incorporating cross-dataset testing, and exploring hybrid feature-based and deep learning frameworks for improved diagnostic performance.

References

[1] M. Yap, et al., “Breast ultrasound lesion detection using deep learning: A review,” Medical Image Analysis, vol. 73, pp. 102–134, 2021. [2] M. Al-Dhabyani, et al., “Deep learning approaches for breast cancer detection and diagnosis using ultrasound: A review,” Journal of Imaging, vol. 6, no. 11, pp. 1–19, 2020. [3] M. Al-Dhabyani, et al., “Breast ultrasound images dataset (BUSI),” Mendeley Data, V1, 2019. [4] [4] H. Abdelrahman, et al., “An annotated dataset of breast ultrasound images,” Data in Brief, vol. 28, pp. 104–106, 2020. [5] Y. Cao, et al., “Mask-guided classification for breast ultrasound images,” IEEE Transactions on Medical Imaging, vol. 40, no. 9, pp. 2436–2447, 2021. [6] S. Ganesan, et al., “Texture analysis of breast lesions in ultrasound images using GLCM features,” Pattern Recognition Letters, vol. 34, no. 9, pp. 103–110, 2013. [7] D. K. Iakovidis, et al., “Combining classifiers with KNN for breast tumor classification in ultrasound,” Computer Methods and Programs in Biomedicine, vol. 122, no. 3, pp. 347–356, 2015. [8] Y. Liu, et al., “Deep learning in breast ultrasound imaging for cancer detection,” Ultrasound in Medicine & Biology, vol. 45, no. 9, pp. 2673–2684, 2019. [9] X. Zhang, et al., “Breast lesion classification with deep CNNs on ultrasound images,” Computerized Medical Imaging and Graphics, vol. 79, pp. 101–112, 2020. [10] J. Shen, et al., “Transfer learning in medical image analysis: Breast ultrasound classification case study,” IEEE Access, vol. 8, pp. 123–134, 2020. [11] Y. Hu, et al., “Attention-guided CNN for breast ultrasound classification,” Artificial Intelligence in Medicine, vol. 107, pp. 101–118, 2020. [12] J. Xu, et al., “Hybrid CNN-transformer models for breast ultrasound tumor classification,” Medical Image Analysis, vol. 76, pp. 102–145, 2022. [13] P. K. Saha, et al., “Computer-aided diagnosis in automated breast ultrasound: Challenges and opportunities,” Frontiers in Oncology, vol. 11, pp. 1–15, 2021. [14] S. M. Lundberg and S. Lee, “Explainable AI for breast ultrasound diagnosis using CNN visualization,” Nature Machine Intelligence, vol. 3, pp. 1–10, 2021. [15] D. B. Nguyen, et al., “Hybrid machine learning frameworks for breast ultrasound tumor detection,” Biomedical Signal Processing and Control, vol. 68, pp. 102–139, 2021.

Copyright

Copyright © 2025 Abinaya R. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.