4911. Can Contrast-Enhanced Mammography-Based Radiomics Analysis Provide Robust Results Within a Multivendor, Multisite Study?
Authors * Denotes Presenting Author
  1. Bino Varghese *; Keck School of Medicine
  2. Mariam Thomas; Olive View Medical Center
  3. Linda Larsen; Keck School of Medicine
  4. Guita Rahbar; Olive View Medical Center
  5. Sandy Lee; Keck School of Medicine
  6. Steven Cen; Keck School of Medicine
  7. Mary Yamashita; Keck School of Medicine
Despite reports on the additional value of radiomics models for patient management, poor reproducibility stemming from mainly inter-site variations in image acquisition and processing have hampered its clinical translation. Here, we rigorously evaluate the extent to which differences in scanners, imaging protocols, and patient cohort affect contrast-enhanced mammography (CEM)-based radiomics analysis using multicenter data.

Materials and Methods:
In this multicenter retrospective study, 99 biopsy-proven invasive breast cancers were imaged using CEM prior to any treatment. Of these, 48 were acquired using the Hologic Selenia Dimensions 3D full-field digital mammography system at Center A, and 51 were obtained using the SenoBright, GE Healthcare system at Center B using comparable image acquisition protocols. Radiomic analysis was performed using opensource LIFEx software. For each patient, the craniocaudal view of the breast was used, and the largest lesion was analyzed. Using LIFEx software, in both cohorts, size-matched regions of interest (ROI) were manually contoured by a fellowship-trained breast radiologist on the CEM, capturing the lesion, background parenchymal enhancement (BPE), and fat, respectively. The radiomics panel was comprised of 144 histogram-based indices, which quantified distribution characteristics of the signal intensities comprising the ROIs. The effect of fat on the extracted radiomic metrics was removed by subtracting the metrics extracted from the fat ROI from both the BPE ROI and the lesion ROI, respectively. Site differences were assessed before and after adjustments of population-based confounders (BI-RADS, age, and breast density). Random forest (RF) was used to create prediction models to distinguish lesions from BPE. To ensure the robustness of our findings, the models were trained with data from one site and then tested using data from the second site.

Univariate analyses showed that > 85% of radiomic features were different (<em>p</em> < 0.05) between the two mammography sites in both the BPE and the lesion assessments, respectively. When the effects of fat were removed, the difference between the mammography sites in both the BPE and the lesion assessments reduced to 20% and 44%, respectively. When the effects of population-based confounders were adjusted, the difference between the mammography sites in both the BPE and the lesion assessments reduced to 11.8% (BPE) and 5.5% (lesion). The RF model achieved an area under the curve (AUC) of 0.99 in distinguishing lesions from BPE using independent cross-site validation.

When discriminating between BPE and lesion, CEM-based radiomics analysis was not confounded by scanner models and imaging protocols. However, the heterogeneity of patient characteristics and confounding signals such as fat should be adjusted when analyzing multivendor multisite CEM-based radiomics data.