CESM@UCBM

A Deep Learning Approach for Virtual Contrast Enhancement in Contrast Enhanced Spectral Mammography

Article: Download

Dataset Request:

Download the pdf, fill it out and send it to:
valerio.guarrasi@unicampus.it
p.soda@unicampus.it
cosbi.dev@gmail.com


SUMMARY OF THE STUDY

Contrast Enhanced Spectral Mammography (CESM) is a dual-energy mammographic imaging technique that first needs intravenously administration of an iodinated contrast medium; then, it collects both a low-energy (LE) image, comparable to standard mammography, and a high-energy (HE) image. The two scans are then combined to get a recombined image, also referred to as dual-energy subtracted (DES) images, showing contrast enhancement.
Despite CESM diagnostic advantages for breast cancer diagnosis, the use of contrast medium can cause side effects, and CESM also beams patients with a higher radiation dose compared to standard mammography.
To address these limitations, this study is the first to propose deep generative models to perform virtual contrast enhancement (VCE) on CESM, which involves generating synthetic DES images form only LE images. This would allow having the advantages of CESM without the need to inject contrast medium and without having to previously acquire the HE image by undergoing double radiation. The deep networks used for this purpose consist of an autoencoder and two Generative Adversarial Networks, the Pix2Pix, and the CycleGAN. We perform an extensive quantitative and qualitative analysis of the model’s performance, also exploiting radiologists’ assessments, on a novel CESM dataset that includes 1138 images that, as a further contribution of this work, we make publicly available.
The results show that CycleGAN is the most promising deep network to generate synthetic recombined images, highlighting the potential of artificial intelligence techniques for virtual contrast enhancement in this field.
A schematic representation of the standard approach and our VCE approach leading to DES images generation is shown in Figure 1.

Figure 1. Schematic representation of the standard approach (a) and the VCE approach (b) leading to DES images generation.

CESM@UCBM DATASET

In this work we collected and utilized an in-house dataset, named as CESM@UCBM in the following. Table 1 summarizes its main characteristics. The dataset consists of 1138 images, divided into 569 LE images and 569 DES images. These images were obtained from 105 patients aged between 31 and 90 years, with an average age of 57.4 years and a standard deviation of 11.5 years. They were enrolled in the study under the Ethical Committee approval PAR 51.23 OSS. All the exams were performed at the Fondazione Policlinico Universitario Campus Bio-Medico in Rome between September 2021 and October 2022, using the GE Healthcare Senographe Pristina full-field digital mammography. The resolution of 984 images is 2850 × 2396 pixels, while the remaining 154 images have a resolution of 2294 × 1916 pixels. Among the 1138 images in the CESM@UCBM dataset:

  • 284 images show craniocaudal projection of right breasts, with 142 LE images and 142 DES images;
  • 282 images show craniocaudal projection of left breasts, with 141 LE images and 141 DES images;
  • 292 images show mediolateral oblique projection of right breasts, with 146 LE images and 146 DES images;
  • 280 images show mediolateral oblique projection of left breasts, with 140 LE images and 140 DES images.

For 86 patients late-phase acquisitions are available, implying that the image of the same breast, in the same projection, was acquired multiple times within a few minutes. Unilateral acquisitions were performed for 7 patients. Furthermore, for each patient, we extract information from the medical report and the biopsy, if available. This includes the classification of breast composition with one of the 4 categories (a, b, c, d) defined by the fifth edition of the Breast Imaging Reporting and Data System (BI-RADS) released by the American College of Radiology (ACR). The category a indicates almost entirely fatty breasts and characterizes 130 images, of which 35% show contrast enhancement. The category b identifies breasts with scattered areas of fibroglandular density and characterizes 360 images, of which 33% show contrast enhancement. The category c indicates heterogeneously dense breast tissue and characterizes 414 images, of which 39% show contrast enhance- ment. Finally, the category d identifies extremely dense breasts and characterizes 190 images, of which 36% show contrast enhancement. Instead, for 44 images the ACR category is not reported. Figure 2 compares 4 pairs of LE and DES images belonging
to the CESM@UCBM dataset, which differ in the ACR category. In addition, for 31 patients, the reports are completed with the BI-RADS descriptors (from 1 to 6) associated with the probability of malignancy of the lesion and the actions to be taken to conclude the diagnostic-therapeutic management process. Based on biopsies, in 258 images, 129 of which were LE and 129 were DES, malignant lesions were identified. In 104 images, 52 LE and 52 DES, benign lesions were identified. In 16 images, 8 LE and 8 DES, borderline lesions were identified. The remaining images did not show any tumor-related abnormalities.

Table 1. Characteristics of the CESM@UCBM dataset.

Figure 2. From left to right, pairs of LE (above) and DES images (below) form the CESM@UCBM dataset with ACR categories a, b, c, d.

Upon completing the appropriate form and sending it to the emails provided at the top of the page, we will make the CESM@UCBM dataset available in DICOM format. You will have access to all DICOM tags, except for those containing sensitive data, which have been properly anonymized.
The LE and DES images are provided in the respective subfolders named “Low energy” and “Recombined,” in which they are named as {patient’s number}_{series number}_{breast side}_{image view}_{image type}.

  • “Patient number” is an identification code for the patient ranging from P001 to P105.
  • “Series number” is a parameter extracted from DICOM tags, useful for recognizing late-phase acquisitions. In fact, for the same patient’s number, breast side, image view, and image type, an image acquired in the late phase will have a higher series number compared to the corresponding image acquired in the early phase.
  • “Breast side” can be L or R, depending on whether the image belongs to the left or right breast, respectively.
  • “Image view” can be CC or MLO, depending on whether the image was acquired in craniocaudal or mediolateral oblique projection.
  • “Image type” can be “LOW_ENERGY” or “RECOMBINED”.

Additionally, the file “CESM@UCBM INFO.xlsx” is included, containing information associated with each patient’s number. This information includes the number of available images, the ACR category, the BIRADS score for the left breast (BIRADS L), the BIRADS score for the right breast (BIRADS R), and the results of the biopsies for both the left (BIOPSY L) and right (BIOPSY R) breasts. In cases where the information is not available, the symbol “\” is used.


Acknowledgements

This work was partially founded by: i) Bando Accordo Innovazione DM 24/5/2017 (Ministero delle Imprese e del Made in Italy), under the project with CUP B89J23000580005; ii) PNRR MUR project PE0000013-FAIR; iii) University Campus Bio-Medico di Roma under the program “University Strategic Projects” within the project “AI-powered Digital Twin for next-generation lung cancEr cAre (IDEA)”.