3D-Printed Breast Phantom: How Close Was the Mammogram?

Mammography systems need to reveal fine breast details, but testing image quality should not require repeated X-ray exposure of patients. Physical phantoms offer a way to examine imaging performance using a model that mimics anatomy and X-ray attenuation. This study asks whether a breast phantom can be produced by 3D printing from an actual mammogram—and how closely its resulting image resembles the clinical original.

Key findings

• The breast phantom was made from Grey V4 resin using stereolithography (SLA), based on a digital mammogram.
• For the full image area compared, SSIM was 0.9893, PSNR was 46.2495, and the Pearson correlation coefficient was 0.9598.
• For cropped areas, SSIM ranged from 0.9947 to 0.9949, while PSNR ranged from 59.4286 to 59.6697.
• Spatial frequency at MTF = 0.1 was 3.800 mm⁻¹ for the phantom and 3.842 mm⁻¹ for the patient image.
• The noise power spectrum curves were similar, with differences at higher frequencies that the paper associates with differences in detector exposure.

Why build a mammography phantom?

Mammography is an important imaging method for detecting breast abnormalities. Physical phantoms provide breast-tissue surrogates for evaluating mammography systems, image quality, and visible detail. They make it possible to repeat imaging tests without repeatedly exposing patients to ionizing radiation.

A challenge is to make a model that does more than approximate breast shape. It should also reproduce how X-rays are attenuated across the image. The paper reports converting each point in a source mammogram into a corresponding thickness of printing resin. Changes in model thickness change the amount of material crossed by the X-ray beam, which in turn affects the image.

The paper was authored by Adi Pandžić, PhD. The study appeared in 2024 in IFMBE Proceedings 93, as part of MEDICON 2023/CMBEBIH 2023.

From a mammogram to a printable model

The input was a raw DICOM craniocaudal mammogram. DICOM is a standard format for medical images and associated metadata. The paper reports that custom MATLAB software was developed to convert a two-dimensional image into a three-dimensional mesh, which could then be saved as an STL file for 3D printing.

Before constructing the phantom, the study reports measuring how Grey V4 resin attenuates X-rays. A step wedge, a sample with sections of different thicknesses, was printed and imaged. Pixel values in the raw mammography image were associated with the corresponding resin thickness, and a regression curve was fitted. The reported fit had R² = 0.999.

The source mammogram’s pixel values were then converted into Grey V4 thicknesses. This calibration matters because breast tissue and resin do not have identical attenuation properties. Pixel values cannot simply be treated as physical model thickness without first accounting for the printing material.

The paper reports that imaging geometry was also corrected. The X-ray beam diverges from the source towards the detector, so its path through the phantom changes across the image. A perspective and projection correction was applied using a source-to-detector distance of 655 mm, then the surface was triangulated to create the STL mesh.

The phantom was produced using SLA. In stereolithography, an ultraviolet laser cures liquid resin. The material was Grey V4, a resin from Formlabs. This paper investigates SLA, not FDM printing. After manufacture, the phantom was imaged and its mammogram compared with the original patient image.

Imaging conditions and comparison methods

The imaging system was a Hologic Selenia Dimensions digital mammography unit. The patient’s clinical image was acquired at a tube voltage of 29.0 kV and a current-time product of 126.0 mA s, selected using automatic exposure control. Compression force was 58.4 N. The phantom was imaged at 29.0 kV and 120 mA s, selected manually, without compression. The source-to-detector distance was the same for both.

The image comparisons used PSNR, SSIM, and PCC. PSNR relates signal strength to image error; SSIM assesses similarity in image structure; and PCC measures the linear relationship between pixel values. The study also examined the modulation transfer function (MTF), which describes how imaging preserves detail at different spatial frequencies, and the noise power spectrum (NPS), which describes how noise is distributed across frequencies.

For MTF, the analysis used a profile across a microcalcification and fitted a Gaussian line spread function. For NPS, the analysis used 100 relatively homogeneous regions of interest, each 50 × 50 pixels. These regions did not include microcalcifications or high-density tissue.

What did the image-quality results show?

The two largest microcalcifications were clearly visible in the phantom image, with contrast similar to that in the original clinical image. In the physical model, they corresponded to taller columns extending above shorter background columns. This indicates that the workflow transferred important image detail, although reproduction of very small structures depends on printer resolution.

For the full image area, the results were PSNR 46.2495, SSIM 0.9893, and PCC 0.9598. SSIM for the three cropped areas was 0.9948, 0.9947, and 0.9949. Their PSNR values ranged from 59.4286 to 59.6697. These measurements indicate strong similarity in the regions assessed; they do not establish that the model represents every tissue property or clinical case.

For sharpness, the full width at half maximum of the line spread function was 1.330 mm for the phantom and 1.315 mm for the patient image. Spatial frequency at MTF = 0.1 was 3.800 mm⁻¹ for the phantom and 3.842 mm⁻¹ for the patient image. The values are close, with slightly higher measured resolution in the clinical image.

The NPS curves were similar in their general shape and local maxima. The paper associates differences at higher frequencies with different detector exposure, since the patient and phantom were imaged using different exposure settings. The paper notes that the difference could be corrected by introducing a scaling factor.

What the workflow means for engineers and researchers

The method brings together medical-image processing, material characterization, geometric correction, and additive manufacturing. For engineers, a key point is that the model depends on calibration of the actual resin: image pixels are converted into thicknesses intended to produce corresponding X-ray attenuation, rather than directly into an uncalibrated shape.

For students, the project shows how an STL model can be generated from data that did not begin as CAD geometry. For researchers, SSIM and PSNR provide quantitative checks of image similarity, while MTF and NPS add information about sharpness and noise that a single similarity score cannot capture.

The paper also describes potential for patient-specific phantoms and production at a fraction of the cost of similar commercially available phantoms. These are proposed practical benefits of the approach; the paper does not report a cost comparison.

Limitations and conclusion

The results concern the phantom, material, and SLA workflow examined in this study. The patient and phantom were not imaged under identical exposure conditions: their mA s settings differed, and the phantom was not compressed. The paper also points out that the resolution limits of available SLA printers require compromises when reproducing very small objects. Therefore, an SSIM close to 0.99 describes the comparison of these images; it should not be read as proof that the phantom completely replaces a breast for every purpose.

The study demonstrates that calibrated Grey V4 resin and SLA 3D printing can produce a mammography phantom whose image was highly similar to the selected clinical mammogram. Using spatial and frequency-domain measures gave a broader assessment than pixel comparison alone. The work provides a practical foundation for reproducible anthropomorphic models to support mammography-system evaluation and further research.

Frequently asked questions

Author: Adi Pandžić, PhD in Mechanical Engineering

This article summarises a published research paper; please refer to and cite the original paper. About the author

← All articles