Breast Cancer Area Identification in Mammograms Using Expectation Maximization Gaussian Mixture Model

Breast cancer accounts for 25% of all cancer diagnoses and 16% of cancer-related deaths among women globally, with high mortality rates due to late diagnosis. Early detection relies on imaging techniques such as mammography, histopathology, and breast ultrasound, with mammography being the gold standard due to its proven to detect breast cancer, thus it is effective for breast cancer treatment. However, mammogram images often produce noise and artefacts, complicating early-stage cancer detection and emphasizing the need for advanced image processing. Clustering algorithms such as K-means and Expectation Maximization - Gaussian Mixture Model (EM-GMM) have shown potential in image segmentation. This study investigates the application of EM-GMM for mammogram image segmentation, a relatively underexplored area compared to K-means, focusing on identifying potentially cancerous regions. EM-GMM, which models data based on Gaussian distribution through iterative EM algorithm application, was used on grayscale images of malignant tissue. The optimal cluster number for detecting cancerous areas was determined to be nine using the Bayesian Information Criterion (BIC). This was tested and compared with segmentations using seven and eight clusters. Results demonstrate that the segmentation with nine-cluster achieves the highest accuracy for segmenting and identifying cancerous regions. These findings support early detection and thorough breast health assessment, potentially enhancing diagnostic precision and lowering breast cancer mortality and morbidity.