Breast cancer detection using hybrid convolutional neural network

Abstract

This thesis evaluates frozen-backbone transfer learning for automated Normal-versus-Disease triage in Digital Breast Tomosynthesis (DBT). Using the BCS-DBT dataset, 17 architectures were tested with different slice-window depths and classifier widths, followed by hyperparameter optimization and spatial/probability fusion. Compact CNNs performed better than larger or attention-based models under the frozen constraint. While fusion improved sensitivity and F1 score, overall AUROC remained limited, demonstrating the performance ceiling of frozen transfer learning for DBT triage.