Sensitivity-Based Breast Cancer Characterization on CBIS-DDSM Mammography Dataset Using Fine-Tuned EfficientNet-B2 Model

Abstract

Mammography is crucial in detecting breast cancer early and reliably to reduce the rate of mortality. Models based on deep learning have shown encouraging results, but most of the current methods have poor sensitivity to malignant cases because the class imbalance is extreme and too much emphasis is put on accuracy-based evaluation. This paper presents a sensitivity-based classification framework of breast cancer based on a fine-tuned EfficientNet-B2 model of the CBIS-DDSM dataset. The suggested method integrates the use of class-weighted binary cross-entropy, partial fine-tuning, and higher convolutional layer, and the use of sigmoid to enable binary classification to counter the imbalance and enhance malignant detection. A large scale of experiments prove that the improved model is much higher than the fully frozen baseline and has a test sensitivity of 55.77% on malignant cases, with overall accuracy of 64.96% and ROC-AUC of 0.679. Confusion matrix, ROC curves, and learning dynamics further validate the robustness and clinical relevance of the proposed framework. The results highlight the necessity of sensitivity-oriented learning strategies for real-world breast cancer screening systems.