Optimizing Transfer Learning for Low-Resource Breast Imaging: A Comparative Study of Progressive Layer Unfreezing and Pre-trained Backbones

Abstract

Objective: To find the optimal combination of pre-trained convolutional neural network (CNN) architectures and depth of progressive layer unfreezing for breast cancer classification in a low-resource mammography scenario through transfer-learning. Methods: Five pre-trained CNN backbones (VGG-19, ResNet50, DenseNet121, Xception and EfficientNetB0) were tested on the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) containing 3,265 labelled cases with a progressive layer-unfreezing protocol: 3, 5, 7, 10 and 20 layers. The accuracy, area under the receiver operating characteristic curve (AUC), precision, recall, and F1-score were used for the evaluation of the models and the statistical significance was evaluated using DeLong's test, bootstrap confidence intervals and McNemar's test. Results: VGG-19 with the last five convolutional layers unfrozen had the highest performance (89.5% validation accuracy, AUC 0.890), significantly better than DenseNet121 (85.3% accuracy and AUC 0.860) and EfficientNetB0 (72.1% accuracy, AUC 0.670) (DeLong's test p < 0.01 for most pairwise comparisons). The unfreezing depth was found to have a non-monotonic relationship with performance, with the best generalisation at moderate unfreezing (5-7 layers), and unfreezing more than 10 layers consistently resulting in overfitting, whilst unfreezing fewer than 3 layers resulted in limited domain adaptation. Conclusions: For low-resource breast imaging, the simple, older architectures (VGG-19) can be superior to more recent, parameter-efficient architectures, and modest progressive unfreezing (5 layers, 20-25% unfreezing of convolutional depth) provides the best compromise between domain adaptation and overfitting. The results guide the optimization of transfer learning protocols in breast imaging applications that are data-constrained.