Improvement of Breast Tumor Classification Based on Machine Learning

Enhancing the accuracy of breast tumor diagnosis will result in a global positive impact on women's health by improving prognosis, treatment efficacy, and mortality rates. To achieve superior diagnostic precision, this study introduces an innovative Computer-Aided Diagnosis (CAD) architecture specifically designed for categorizing breast lesions. The experimental database consists of 2,188 mass images from DDSM, 106 from INbreast, and 53 from MIAS. The proposed system was designed in four stages. In the preprocessing stage, different denoising methods were assessed and compared, including Gaussian and Wiener filters. The segmentation stage was evaluated using Otsu thresholding and Fuzzy C-Means (FCM). We implemented a hybrid feature extraction strategy that fuses first-order statistical metrics with second-order textural descriptors derived from the Gray Level Co-occurrence Matrix (GLCM). For the categorization stage, the study evaluates the comparative performance of Support Vector Machine (SVM) and K-Nearest Neighbors (KNN). Quantitative analysis established the superiority of the Gaussian filter over competing denoising techniques, exhibiting the most favorable error and signal-to-noise ratios. In the delineation phase, the proposed framework registered a DSC of 0.93 ± 0.02 and a Jaccard score of 0.88 ± 0.06 relative to manual annotations. The classification accuracy of the features extracted from the Otsu threshold combined with SVM classifier reached 98%, 97%, and 96% for DDSM, INbreast, and MIAS, respectively.