EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-SA 4.0

3D Mammogram Reconstruction and Tumor Localization via 2D U-Net–GAN

Samar S. Mamand, Abdulqadir I. Abdulla

ARO-THE SCIENTIFIC JOURNAL OF KOYA UNIVERSITY · 17.07.2026 · 14 (2) · 1-7

DOI: 10.14500/aro.12766 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ARO-THE SCIENTIFIC JOURNAL OF KOYA UNIVERSITY, online erschienen am 17.07.2026. Verfasst von Samar S. Mamand und 1 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-SA 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Conventional two-dimensional (2D) mammography compresses complex breast anatomy into a single projection, obscuring lesions through tissue overlap and limiting reliable tumor sizing. Although tomosynthesis and computed tomography recover depth information, these modalities are not universally available and add dose, time, and cost. We propose a single-view 2D-to-three-dimensional (3D) reconstruction framework that couples a 3D U-Net with a Generative Adversarial Network (GAN) to synthesize anatomically coherent 3D mammograms from a single 2D input. The end-to-end pipeline performs (i) seed stack formation and normalization, (ii) adversarial 3D reconstruction, and (iii) lightweight post-processing—Gaussian smoothing, global thresholding, and 26-connected component analysis—to localize regions of interest (ROI) and compute voxel-accurate lesion volumes in mm³/cm³ from pixel and slice spacings. Training combines weighted GAN and reconstruction losses to balance perceptual realism with structural fidelity. On a held-out test set, the method achieves a mean squared error of 0.0059, a mean peak signal-to-noise ratio of 24.13 dB, a mean absolute error of 0.0283, and a structural similarity index of 0.9296, outperforming interpolation and non-adversarial 3D U-Net baselines. Qualitative renderings show preserved parenchymal textures, smooth interslice transitions, and precise ROI overlays, while a red isosurface visualization highlights the 3D lesion extent for volumetry. The approach is simple to train, reproducible, and compatible with standard imaging toolchains, delivering clinically actionable volumetric measurements without multiview acquisition. Future work includes multiview supervision (craniocaudal and mediolateral oblique) to strengthen depth consistency, uncertainty maps to qualify reconstructions, and learned ROI segmentation with topology-aware objectives, advancing low-overhead 3D support in routine mammography.

Zitieren

APA

Mamand, S. S., & Abdulla, A. I. (2026). 3D Mammogram Reconstruction and Tumor Localization via 2D U-Net–GAN. ARO-THE SCIENTIFIC JOURNAL OF KOYA UNIVERSITY, 14(2), 1-7. https://doi.org/10.14500/aro.12766

Vancouver

Mamand SS, Abdulla AI. 3D Mammogram Reconstruction and Tumor Localization via 2D U-Net–GAN. ARO-THE SCIENTIFIC JOURNAL OF KOYA UNIVERSITY. 2026;14(2):1-7. doi:10.14500/aro.12766

BibTeX
@article{Mamand2026,
  author = {Mamand, Samar S. and Abdulla, Abdulqadir I.},
  title = {3D Mammogram Reconstruction and Tumor Localization via 2D U-Net–GAN},
  journal = {ARO-THE SCIENTIFIC JOURNAL OF KOYA UNIVERSITY},
  year = {2026},
  volume = {14},
  number = {2},
  pages = {1--7},
  doi = {10.14500/aro.12766}
}
Kostenlos starten →