Effect of Various DCE-MRI Image Parameters On AI Assessment of Breast Parenchymal Enhancement
Presentations
TH-A-206-2 (Thursday, 7/14/2022) 7:30 AM - 8:30 AM [Eastern Time (GMT-4)]
Room 206
Purpose: Current clinical assessment qualitatively describes background parenchymal enhancement (BPE) as Minimal, Mild, Moderate, or Marked BPE based on the visually perceived volume of enhancement in dynamic contrasted-enhanced (DCE)-MRI. This is prone to interobserver variability, thus necessitating an objective method for quantifying BPE, a significant predictor of breast cancer risk. This study investigates the effect of various image parameters on the AI assessment of BPE.
Methods: Our dataset consisted of 350 breast DCE-MRI cases retrospectively collected from 2005-2017 under IRB protocol. Maximum- and average-intensity-projections (MIP, AIP) were created from 1st- and 2nd-post-contrast subtraction images after electronic lesion removal. Lesions and breasts were segmented using fuzzy c-means clustering and U-Net, respectively. Within the affected breast, mean pixel intensities of original and rescaled images were calculated. ROC analysis was performed using the clinical radiologist BPE ratings as the reference standard to assess the predictive values of the computer-extracted BPE in the classification tasks of Minimal vs. Marked BPE and of Low (Minimal, Mild) vs. High (Moderate, Marked) BPE. Statistical significance of the ROCAUC was determined (z-test). Kendall’s tau-b was used in rank correlation of the quantitative BPEs with radiologist BPEs (t-test). P-values were evaluated using Bonferroni correction for multiple comparisons.
Results: Quantitative BPEs were statistically superior to random guessing, except for mean of original MIPs. ROC curves showed quantitative BPEs from 2nd-post-contrast-projections were superior to 1st-post-contrast-projections, and MIP BPEs were superior to AIP BPEs. BPE values from rescaled images were superior predictors than from original images. Statistically significant trends were found between the radiologist BPE ratings and the computer-extracted BPE values (rescaled images).
Conclusion: Results demonstrate the potential for BPE AI (including electronic lesion removal), which yields quantitative values in classifying Marked vs. Minimal BPE across various image viewing projections and DCE timepoints.
Funding Support, Disclosures, and Conflict of Interest: NIH T32 EB002103, S10 OD025081, U01CA195564, P30 CA014599, UCCCC grant. MLG receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba, and was a cofounder in Qlarity Imaging.
Keywords
Taxonomy
Contact Email