Title: MGRegBench: A New Benchmark Dataset Incorporating Anatomical Landmarks for Mammography Image Registration

Abstract:

Accurate registration of mammography images is a critical component for clinical tasks such as monitoring the progression of breast tissue diseases. Nevertheless, advancements in this field have been hindered by a lack of standardized, reproducible benchmarks and publicly available datasets. Because prior research frequently relies on proprietary data and varying evaluation metrics, direct comparisons between studies are often impossible. To resolve these issues, we introduce MGRegBench, a comprehensive evaluation framework designed to prevent data leakage and ensure patient-disjoint testing. This resource includes over 5,000 image pairs accompanied by breast segmentation masks, along with 100 pairs featuring manually annotated anatomical landmarks. Additionally, it provides standardized splits for training and evaluation, as well as ready-to-execute baseline models.

Leveraging this dataset, we conducted benchmarks on a variety of registration techniques. These included classical methods like ANTs, learning-based approaches such as VoxelMorph and TransMorph, implicit neural representations via IDIR, a specialized mammography method, and the recent deep learning architecture MammoRegNet, with implementations tailored to this imaging modality. We further assessed generalization capabilities using the independent SDM-MCs dataset.

Our key contributions are as follows: (1) We present the first large-scale public dataset for mammography registration that includes both manual landmarks and segmentation masks; (2) We establish a transparent, leakage-controlled benchmark that allows for the first direct comparison of classical and machine learning-based methods; (3) We perform external validation on the SDM-MCs dataset to determine if the observed trends hold true outside of MGRegBench; and (4) We provide a thorough analysis of deep learning-based registration techniques. By publicly releasing both the code and data, we aim to create a foundational resource that promotes fair, reproducible, and clinically significant comparisons, thereby stimulating further innovation in AI-driven medical imaging.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC