A Comprehensive Benchmarking and Systematic Analysis of Deep Learning Models for Sonomammogram Segmentation

Malitha Gunawardhana, Norbert Zolek

Keywords: Segmentation, Sonomammogram, Deep Learning, Benchmarking

Abstract: Accurate segmentation of breast lesions in sonomammograms supports computer assisted diagnosis and early breast cancer detection. Existing public ultrasound datasets contain duplicates, mislabeled cases, and non-breast images, which leads to unreliable model evaluation. To address this, we construct a curated multi-centre dataset of 3,494 images with expert-verified annotations and patient-level splits. Using this dataset, we define a unified benchmarking protocol and evaluate eleven representative architectures, including nnU Net variants, SegResNet, SwinUNETR, U Mamba, and SAMed. All models are trained and assessed under identical preprocessing, training, and evaluation settings. Performance is measured with Dice, Sensitivity, Specificity, Accuracy, and Hausdorff Distance metrics. We also analyse how loss function choice and training data volume influence performance. SAMed p512 obtains the best Dice score at 0.860 ± 0.141 and the lowest Hausdorff Distance at 3.896 ± 5.472. The benchmark provides a reproducible reference for breast ultrasound segmentation and clarifies how architecture design and data-related factors shape performance in this setting.

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Other

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 25