BREAST CANCER MULTI-CLASS CLASSIFICATION THROUGH MACHINE LEARNING TECHNIQUES USING MAMMOGRAPHIC IMAGES

Abstract

This study presents a research article on breast cancer multi-class classification using mammographic images and a BI-RADS-oriented analytical framework. The analytical file contained 1,200 mammographic image records distributed across training, validation, and test partitions, with linked demographic, imaging, lesion, and pathology descriptors. The methodological design combined structured preprocessing, feature normalization, feature selection, comparative machine learning experiments, and transfer-oriented deep learning evaluation. The Results section was built around class distribution, image and lesion profile, category gradients, multiclass model performance, class-wise behavior, and robustness across density, image quality, and cross-dataset holdout conditions. The dataset showed a broad spread across BI-RADS 0 to 6, with category 2 representing the largest share and category 6 the smallest. Age, lesion size, texture, intensity, contrast, and spiculation rose in a clear direction as the diagnostic category moved from negative or probably benign observations toward highly suspicious and biopsy-proven malignant groups. Among conventional models, ensemble learning produced the strongest category-level performance, while the transfer-enhanced Xception configuration achieved the best overall multiclass outcome with an accuracy of 95.4%, weighted precision of 94.9%, weighted recall of 95.4%, and weighted F1-score of 94.8. Class-wise analysis showed that the most stable recognition occurred in BI-RADS 2, BI-RADS 3, BI-RADS 4, and BI-RADS 5, whereas the hardest distinctions appeared around BI-RADS 0 and BI-RADS 6 because of assessment incompleteness in one case and smaller sample size in the other. Stratified analysis indicated that dense breasts and low image quality reduced model performance, yet the final framework remained strong across all subgroups and retained acceptable cross-dataset robustness. The findings show that BI-RADS-aligned machine learning and deep learning can provide clinically meaningful multi-class support for mammographic interpretation, with strong potential for decision assistance, triage, and diagnostic standardization in breast imaging practice.

Keywords : breast cancer, mammography, BI-RADS, machine learning, deep learning, multi-class classification, transfer learning.