# Direct Segmentation of Mammography and Tomosynthesis Sinograms for Lesion Localization

## TL;DR

This paper introduces a new method for detecting breast lesions using raw scan data directly, improving accuracy and reliability in mammography and tomosynthesis.

## Contribution

The novel approach uses deep learning to segment sinograms directly, avoiding reconstruction and improving lesion localization.

## Key findings

- Mammography sinogram segmentation achieved a Dice score of 0.90 on the test set.
- Combining mammography and tomosynthesis sinograms improved tomosynthesis performance to a Dice score of 0.84.
- The method outperformed YOLOv5x in localizing small or multiple lesions.

## Abstract

Breast cancer detection commonly relies on mammography and digital breast tomosynthesis, but lesion localization is often hindered by tissue overlap and information loss during image reconstruction. This study investigates a direct analysis of sinograms, which preserve raw projection data, without prior reconstruction. A deep learning-based segmentation approach is proposed to localize breast lesions directly from sinograms. Mammography sinograms achieved the most accurate localization, while combining mammography and tomosynthesis sinograms improved tomosynthesis performance, supporting more reliable lesion localization.

Background: The Detection and localization of breast lesions remain challenging in mammography and digital breast tomosynthesis (DBT) due to tissue overlap and information loss during volumetric reconstruction. Sinograms preserve the full angular projection data acquired during scanning, enabling analysis of tissue structure without reconstruction. Methods: This study proposes a direct segmentation approach for mammography and DBT sinograms using a U-Net architecture. Experiments were conducted on 1082 annotated mammography mass images from the CBIS-DDSM dataset (521 benign, 561 malignant) and 272 annotated DBT images from the Breast Cancer Screening DBT dataset (136 benign, 136 malignant). Dataset splitting was performed at the patient level to prevent data leakage, and all reported quantitative results correspond to the independent test set, with the validation set used solely for model selection and early stopping. Three input configurations were evaluated: mammography sinograms, DBT sinograms, and a combined model. Results: The mammography model achieved the highest performance (Dice: 0.94 training, 0.90 test), outperforming DBT alone (0.77 training, 0.70 test). Multimodal fusion improved DBT results (Dice: 0.84 test). Centroid analysis showed 99.11% correspondence with reference annotations (average distance: 2.83 pixels), and partial back-projection reconstructions confirmed anatomical consistency. Compared with YOLOv5x, the proposed approach provided superior lesion localization, particularly for small or multiple lesions. Conclusions: Direct sinogram segmentation is an efficient, clinically viable strategy for breast lesion detection and localization.