Comparative Evaluation of ResNet50 and DenseNet121 with SE-Block for Breast Cancer Classification

Abstrak

Breast cancer classification using mammography remains a challenging task due to the similarity between tumor masses and healthy tissue. Convolutional Neural Networks (CNNs) have been widely applied in medical image analysis; however, their performance strongly depends on architecture design and hyperparameter optimization. This study presents a comparative evaluation of baseline and Squeeze-and-Excitation (SE) Block-integrated ResNet50 and DenseNet121 architectures using the CBIS-DDSM dataset for benign and malignant mass classification, with emphasis on model stability and false negative reduction. Contrast Limited Adaptive Histogram Equalization (CLAHE) preprocessing was applied to enhance the Region of Interest (ROI). Experiments were conducted through 64 scenarios involving variations of optimizers (Adam, SGD with Momentum, RMSProp, and Nadam), learning rates (0.001 and 0.0001), and batch sizes (16 and 32). Experimental results show that ResNet50 tends to be unstable under hyperparameter variations, while DenseNet121 combined with SE-Block improves model stability and classification performance. The best result was achieved by SE-DenseNet121 using the Nadam optimizer with a learning rate of 0.0001 and batch size of 32, obtaining 69% accuracy, 68% precision, 76% recall, and 72% F1-score. These findings indicate that SE-Block integration can improve feature attention and produce a more balanced mammography classification model.

Identifier

• DOI: https://doi.org/10.31328/jointecs.v11i1.7676
• OpenAlex: https://openalex.org/W7167664381