An MRI-based nomogram helps predict programmed death-ligand (PD-L1) status in breast cancer with 86.3% accuracy, according to a poster to be presented May 12 at the ISMRM meeting in Cape Town, South Africa.
The findings are promising for reducing invasive tumor biopsy procedures and improving patient selection for immune checkpoint inhibitors, according to lead author Yiwen Wang of the Affiliated Hospital of Jining Medical University in Jining, China, and colleagues.
“PD-L1 expression correlates with aggressive tumor features,” the group wrote, adding that “current PD-L1 evaluation through immunohistochemical biopsy is limited by tumor heterogeneity.”
The team’s retrospective study focused on postoperative PD-L1 assessment, enrolling 146 female patients with pathologically confirmed primary invasive breast carcinoma between August 2019 and January 2024.
All of the women underwent dynamic contrast-enhanced MRI (DCE-MRI) and intravoxel incoherent motion (IVIM) examinations and had definitive immunohistochemical (IHC) results. Patients who had received neoadjuvant therapy or had inadequate image quality were excluded. Participants were randomly allocated to training (70%, n = 102) and test (30%, n = 44) sets.
MRI scans were performed on performed on a 3-tesla scanner (Discovery MR 750W, GE HealthCare) using a dedicated breast coil. The protocol included FSE-T1WI, FRFSE-T2WI, IVIM with multiple b-values (20-2000 s/mm²), and DCE-MRI following gadopentetate dimeglumine administration (45 dynamic phases), Wang and colleagues noted.
IVIM parameters (diffusion coefficients D and D*, and f) were calculated using bi-exponential analysis. DCE-MRI parameters (volume transfer constant [Ktrans], mean return flow [Kep], extravascular, extracellular volume [Ve]) were derived through pharmacokinetic modeling. In addition, two experienced radiologists independently analyzed morphological features and placed regions of interest on the most solid tumor components, avoiding necrotic areas, the group explained.
Pathological assessment relied on the SP142 assay to evaluate PD-L1 expression and standard IHC protocols.
The group used multiple logistic regression models to identify the most predictive feature set for PD-L1 status, evaluating tumors based on margin, D and D* values, Ktrans, and Ki-67.
Collectively, DCE-MRI, IVIM, and pathology features captured proliferative capacity, perfusion dynamics, and morphological invasiveness, the group reported. Their integrated approach demonstrated an area under the curve (AUC) of 0.95, with an accuracy of 86.3%, sensitivity of 84.9%, and specificity of 98% in the training set, they noted, suggesting it outperforms any single-modality model.
Analysis of a 49-year-old woman who was diagnosed with triple-negative breast cancer depicts (left to right) IVIM maps, IVIM pseudo-colored maps, DCE-MRI maps of the patient, along with DCE-MRI pseudo-colored maps for Ktrans, Kep, Ve, Ktrans, Kep, and Ve, and PD-L1 IHC testing for the patient (×400).ISMRM
The researchers reported that lower D values indicated higher cellularity in PD-L1+ tumors, while elevated D* and Ktrans reflected increased perfusion and vascular permeability.
Additionally, D values and Ktrans parameters complemented Ki-67 proliferation index and unclear tumor margins, Wang and colleagues noted. They added that there was no statistically significant difference between the IVIM and the DCE models.
“The developed nomogram effectively translated these multidimensional features into a clinically applicable tool, showing excellent calibration and clinical utility across decision curve analysis,” the group wrote, noting that the study was limited by sample size and technical variability in IVIM parameters.
While future multicenter validation is required, “this multimodal approach shows potential as a noninvasive alternative to biopsy for identifying candidates for PD-L1 blockade immunotherapy,” the researchers concluded.