A Dual-Branch Feature-Fusion Model of ResNet152V2 and DenseNet201 for Breast Abnormality Detection in Digitized Mammograms

Abstract

Screening mammography remains the principal population-level tool for early breast cancer detection, and convolutional neural networks (CNNs) have been widely proposed as second readers. However, published comparisons of CNN backbones on mammographic data are difficult to interpret, because architectures are frequently trained under different hyperparameters, preprocessing pipelines and evaluation protocols, so that reported rankings confound architecture with tuning. This study reports a controlled comparison of six configurations for binary normal versus abnormal classification of digitized mammographic tiles from the Digital Database for Screening Mammography (DDSM): five ImageNet-pretrained backbones — VGG16, ResNet152V2, DenseNet201, Xception and MobileNetV2 — and a dual-branch model that concatenates the globally pooled feature vectors of ResNet152V2 and DenseNet201. A single preprocessing pipeline produced a class-balanced set of 10,000 grayscale tiles at 224x224 resolution, split 80/20 with stratification. All configurations shared a two-phase transfer-learning protocol: a frozen feature-extraction phase followed by partial fine-tuning of the final backbone layers, with identical augmentation and classification head. The dual-branch fusion model achieved the highest validation accuracy (0.9415) with an AUC of 0.9845, exceeding its stronger constituent backbone, ResNet152V2, by 0.55 percentage points and its weaker constituent, DenseNet201, by 1.20 points, confirming that the fusion contributed measurable gain. With only 15.9 million parameters, VGG16 reached the best performance for AUC (0.9852), sensitivity (0.9510), and F1-score (F1 0.9582) – clinical outcome metrics of great importance from a screening perspective. MobileNetV2 achieved 0.8980 accuracy with 4.24 million parameters, the best accuracy/per-parameter ratio of any model tested and 3.6 times higher than the second best (LeNet) The results show that dualbranch feature gives marginally and consistently better accuracy than either of its constituent backbones on average, and that the choice of ranking metric significantly affects preference between architectures. The full pipeline is released to allow for exact replication.