Latency-Aware Breast Cancer Detection in Mammography Images Using a Fog-Cloud Framework Based on Stacked Transfer Learning and PSO-Optimized XGBoost

Abstract

Breast cancer is one of the most important public health challenges worldwide, and the quality of its early detection is directly associated with reduced mortality, lower treatment costs, and improved care pathways. Mammography remains one of the principal screening tools; however, the interpretation of mammographic images is affected by human error and infrastructural limitations in the presence of tissue density, noise, similarities between benign and malignant lesions, and variations in imaging protocols. This paper presents an automated breast cancer detection framework focused on reducing processing latency and increasing diagnostic reliability within a fog-cloud architecture. At the fog layer, mammographic images are standardized through resizing, intensity normalization, noise reduction, and lightweight preparation to reduce data-transmission volume and initial response time. At the cloud layer, three transfer-learning models—ResNet50, DenseNet121, and EfficientNetB7—are employed to extract deep features. The probabilistic outputs of these models are integrated in a three-level stacked ensemble architecture, with XGBoost used as the final meta-classifier. To reduce dependence on manual tuning and control overfitting, the key hyperparameters of the base models and the meta-classifier are optimized using particle swarm optimization. Evaluation was conducted on the public CBIS-DDSM dataset using stratified cross-validation. The results showed that the proposed model achieved an accuracy of 97.5% on the test subset, balanced performance across the two classes, a recall of 98.0% for the benign class, and 96.9% for the malignant class. The findings indicate that combining near-source preprocessing, transfer-based feature extraction, stacked ensembling, and metaheuristic optimization can provide a deployable framework for intelligent screening in distributed healthcare environments.

Breast cancer is one of the most important public health challenges worldwide, and the quality of its early detection is directly associated with reduced mortality, lower treatment costs, and improved care pathways. Mammography remains one of the principal screening tools; however, the interpretation of mammographic images is affected by human error and infrastructural limitations in the presence of tissue density, noise, similarities between benign and malignant lesions, and variations in imaging protocols. This paper presents an automated breast cancer detection framework focused on reducing processing latency and increasing diagnostic reliability within a fog-cloud architecture. At the fog layer, mammographic images are standardized through resizing, intensity normalization, noise reduction, and lightweight preparation to reduce data-transmission volume and initial response time. At the cloud layer, three transfer-learning models—ResNet50, DenseNet121, and EfficientNetB7—are employed to extract deep features. The probabilistic outputs of these models are integrated in a three-level stacked ensemble architecture, with XGBoost used as the final meta-classifier. To reduce dependence on manual tuning and control overfitting, the key hyperparameters of the base models and the meta-classifier are optimized using particle swarm optimization. Evaluation was conducted on the public CBIS-DDSM dataset using stratified cross-validation. The results showed that the proposed model achieved an accuracy of 97.5% on the test subset, balanced performance across the two classes, a recall of 98.0% for the benign class, and 96.9% for the malignant class. The findings indicate that combining near-source preprocessing, transfer-based feature extraction, stacked ensembling, and metaheuristic optimization can provide a deployable framework for intelligent screening in distributed healthcare environments.