COMPARATIVE ANALYSIS OF TRANSFER LEARNING PERFORMANCE USING VGG16 AND XCEPTION ARCHITECTURES FOR BREAST CANCER HISTOPATHOLOGICAL IMAGE CLASSIFICATION

Abstract

Breast cancer ranks among the most prevalent malignancies in women globally and demands reliable image-based diagnostic support systems for accurate clinical decision-making. This study compares two transfer learning architectures, VGG16 and Xception, for breast cancer histopathological image classification using the BreakHis dataset, partitioned into training, validation, and testing subsets through stratified splitting. The pipeline included preprocessing, data augmentation, and a two-phase training strategy (freeze then fine-tuning), with performance assessed using eight metrics: accuracy, precision, recall, specificity, F1-score, AUC-ROC, Matthews Correlation Coefficient (MCC), and Cohen Kappa. Testing on 747 images demonstrated that Xception attained an accuracy of 95.72%, F1-score of 95.72%, AUC-ROC of 0.9926, MCC of 0.9144, and Cohen Kappa of 0.9143, compared to VGG16's accuracy of 90.90%, F1-score of 90.90%, AUC-ROC of 0.9686, MCC of 0.8181, and Cohen Kappa of 0.8179. Confusion matrix analysis revealed that Xception produced only 18 false negatives in the malignant class, versus 37 for VGG16, a clinically meaningful reduction. VGG16 required fewer parameters and shorter training duration, whereas Xception delivered superior classification performance. These findings suggest that Xception is the more suitable architecture for breast cancer Computer-Aided Diagnosis (CAD) system development, particularly where diagnostic sensitivity and overall accuracy outweigh computational cost.