Intelligent Computer-Aided Systems for Breast Cancer Classification in Digital Breast Tomosynthesis Scans

Breast cancer is a significant worldwide health concern, with high incidence and fatality rates. Early identification is critical to improving patient outcomes and reducing the overall burden of the disease. This thesis contributes to knowledge by developing five novel systems for multi-class classification of Digital Breast Tomosynthesis (DBT) scans as normal, benign, or malignant.

This thesis presents unique methodologies and combines them to create five systems. The first system, DeepEval System (DE System), compares six cuttingedge DL models for feature extraction prior to classification using a Support Vector Machine (SVM), serving as a benchmarking tool.

The Mod_AlexNet System (MA System) is presented thereafter. It is a novel system that modifies the traditional AlexNet by incorporating max-pooling layers and batch normalisation layers. These modifications are designed to improve the classification performance. Various optimizers were tested and compared while training Mod_AlexNet with different batch sizes to optimize the performance. In the Feature Fusion and Selection with Ensemble Classifier (FFS-EC) System, feature fusion is integrated, followed by several feature selection models and a majority voting ensemble classification model.

The Multi-Head Mod_AlexNet Attention (MHMA) system introduces a novel attention model to the previously developed Mod_AlexNet. This attention framework focuses on the most relevant parts of the input images, improving the ability of the system to represent important features and thereby enhancing the overall classification performance. Finally, the Hybrid Multi-Head Self-Attention Model with Feature Fusion, Selection, and IVECM for Enhanced DBT Classification (HMSA-FFS-IVECM) System integrates Mod_AlexNet with the attention model, feature fusion, and selection models, as well as a newly developed Integrated Voting Ensemble Classification Model, IVECM, that incorporates class and classifier weights. This comprehensive integration maximizes the classification performance, particularly for the abnormal classes, benign and malignant, which are minorities in the dataset.

The systems were tested using a publicly available dataset, called Breast Cancer Screening – Digital Breast Tomosynthesis (BCS-DBT) dataset (Buda et al., 2020). The HMSA-IVECM System achieved a remarkable specificity of 62.20%, significantly outperforming: DE System (21.43%), MA System (23.91%), FFS-EC (43.07%), and MHMA System (51.99%). The proposed HMSA-IVECM system consistently outperforms existing methods across all scenarios. For benign versus malignant classification, it achieved the highest accuracy of 91.24% and 91.09% in both scenarios, surpassing the closest competitor (Farangis Sajadi Moghadam and Rashidi, 2023) by over 2.5% and Hassan et al. (2024) by over 6%. For normal versus abnormal and cancerous versus non-cancerous cases, HMSA-IVECM demonstrated superior accuracy (94.53% and 93.81%, respectively), showing substantial improvement in sensitivity, precision, and F1-score compared to prior models. These systems contribute to the development of automated breast cancer classification technologies, which promises to greatly enhance the early diagnosis and classification of breast abnormalities.