Benchmarking Transfer Learning for Dense Breast Tissue Segmentation on Small Mammogram Datasets

Dense breast tissue diminishes the sensitivity of mammographic screening and is a key cancer risk factor, which motivates accurate segmentation under scarce and expensive expert annotations in the medical imaging domain. Here, we benchmark the effect of backbone architecture, self-supervised pre-training (SSL), fine-tuning strategy, and loss design for dense-tissue segmentation on a small expert-labeled dataset (596 images) and an in-domain unlabeled corpus (20, 000 images), reflecting the lack of large public pixel-level density datasets. CNNs (EfficientNet, Xception, nnUNet) clearly outperform transformer and Medical-SAM2 models, and full or layer-wise fine-tuning reliably exceeds parameter-efficient updates. Generic image-only SSL (MIM, SimCLR, Barlow Twins) often yields negligible or negative gains over ImageNet initialization, whereas a simple multi-view contrastive SSL and a hybrid segmentation-density loss provide the best accuracy and calibration (e.g., MAE from 14.8% to 11.8%, Spearman with the four BI-RADS breast density categories from 0.42 to 0.51 on VinDr). We also quantify GPU hours for different SSL and fine-tuning choices, showing that only a small set of protocols, such as EfficientNet with multi-view SSL, hybrid loss, and full fine-tuning, offers favorable accuracy-efficiency trade-offs. These findings provide practical defaults for annotation-limited mammography studies and support compute-conscious deployment of automatic breast density assessment in web-based screening workflows.