MammoNeRF: Neural Radiance Fields for Mammography Imaging

Yassine Ameskine, Youssef Alj

Keywords: Mammography, NeRF, 3D Reconstruction, Image Synthesis

TL;DR: We propose MammoNeRF, a novel neural radiance field-based framework for 3D volumetric reconstruction of breast tissue from mammogram images. MammoNeRF is trained using CC and MLO views annotated with the corresponding binary lesion masks.

Abstract: Reconstructing a 3D volumetric representation of breast tissue from 2D mammographic views remains a challenging problem. This is primarily due to extremely sparse imaging, the absence of geometric reasoning, and the lack of volumetric imaging in routine screening. To address these limitations, we propose MammoNeRF, a novel neural radiance field-based framework for 3D volumetric reconstruction of breast tissue from mammogram images. MammoNeRF is trained using craniocaudal (CC) and mediolateral oblique (MLO) views annotated with the corresponding binary lesion masks. To enforce cross-view alignment, we introduce a lesion consistency loss, encouraging the model to estimate consistent poses across both views. This enables the reconstruction of pseudo ground-truth 3D lesion bounding boxes (3DPB) and 3D volumetric breast representations. Without this constraint, traditional NeRF models tend to reconstruct the same lesion at inconsistent spatial locations across views, resulting in incoherent lesion geometry. MammoNeRF jointly optimizes NeRF parameters, camera poses, and lesion centroids within a unified framework. We compare MammoNeRF with existing reconstruction models using the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics, demonstrating that MammoNeRF achieves superior quantitative performance while also improving anatomical fidelity. By transforming standard mammograms into a meaningful 3D representation, MammoNeRF provides a new pathway for enhanced structural understanding in breast imaging and establishes a robust foundation for future work in visualization, and downstream clinical applications.

Primary Subject Area: Image Synthesis

Secondary Subject Area: Image Acquisition and Reconstruction

Registration Requirement: Yes

Reproducibility: https://github.com/yass123am/MammoNeRF

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 105