A predictive model that combines magnetic resonance imaging radiomics with machine learning can estimate PD-L1 expression in breast cancer patients before immunotherapy begins. The work was showcased at the 2026 meeting of the International Society for Magnetic Resonance in Medicine (ISMRM) and continues a five-year research arc: using the MRI already acquired for diagnosis to extract an immunotherapy biomarker, with no extra biopsy.
Why PD-L1 expression matters
PD-L1 is the ligand for the PD-1 receptor on T cells. When a tumor over-expresses it, the local immune response gets dampened — and the tumor becomes an ideal candidate for immune checkpoint inhibitors such as pembrolizumab and atezolizumab. In triple-negative breast cancer (TNBC) this decision is central: the KEYNOTE-522 trial established pembrolizumab plus chemotherapy as the neoadjuvant standard for many of those patients, and PD-L1 status refines who benefits most. The practical problem is that the immunohistochemistry readout depends on adequate tissue, varies between pathologists, and can shift between pre- and post-treatment samples.
How breast MRI radiomics actually works
The pipeline presented at ISMRM follows the canonical architecture of this kind of study. First, the lesion is segmented in dynamic contrast-enhanced MRI (DCE-MRI), usually in 3D and across every post-contrast phase. Dedicated software then extracts hundreds of descriptors from that mask: first-order variance, texture metrics from the gray-level co-occurrence matrix (GLCM) and the gray-level run-length matrix (GLRLM), and zone-based descriptors (GLSZM). Together they form a quantitative fingerprint of tumor heterogeneity — something the human eye can’t measure reliably.
Machine-learning algorithms (random forest, SVM, decision trees, lightweight neural nets) then pick the subset of descriptors that best separates PD-L1-positive from PD-L1-negative patients. Internal validation typically uses five-fold cross-validation, and the more mature models now ship with external cohort testing.
What the recent literature shows
The ISMRM result echoes independent findings. A Memorial Sloan Kettering Cancer Center study published in Cancers (2021) evaluated 62 women with TNBC and built a three-feature model (first-order variance, run-length variance and large-zone low-gray-level emphasis). Performance beat the qualitative radiologist read: 90.7% sensitivity, 85.1% specificity and 88.2% diagnostic accuracy. Crucially, classical BI-RADS features did not associate with PD-L1 status, reinforcing that the useful signal lives in measurements that escape visual inspection.
In 2024, a South China group took the approach multicenter. Five cohorts were combined, with TCGA RNA-seq from 1,089 patients integrated alongside 94 TCIA MRIs to construct a radiomics signature for tumor microenvironment phenotypes — including immune-inflamed versus immune-desert subtypes. The signature held up in an independent cohort, suggesting the radiomic signal is not a single-center artifact.
Implications for daily practice
The appeal is clear: estimate PD-L1 across the entire primary tumor with no fresh biopsy and no sampling bias. That helps in three scenarios — small tumors where biopsy removes most of the lesion, borderline PD-L1 cases (CPS 1-10) where lab-to-lab discordance is highest, and neoadjuvant monitoring, where repeated biopsy to track immune response is impractical. Imaging algorithms have already surpassed radiologists at other oncology tasks, and the trend is to integrate these readouts into the structured report.
In emerging markets, where access to PD-L1 immunohistochemistry is uneven, an MRI-derived biomarker could screen patients ahead of the more expensive pathology workup — provided it is validated on local cohorts. Commercial AI platforms for mammography have shown measurable specificity gains in screening; replicating that level of validation rigor for radiomics pipelines is now the challenge.
Limitations and next steps
Enthusiasm needs to be tempered by methodology. Cohorts remain small, magnetic field strength and acquisition protocols vary, and models trained on one scanner may not work on another without harmonization (ComBat, signature normalization). Reproducibility across vendors — Siemens, GE, Philips — is an open challenge. Initiatives combining large-scale MRI with AI have pushed the community toward reproducible pipelines and open-source code.
The regulatory bar is the other concern. To become standard of care, an MRI-based PD-L1 classifier would need prospective multicenter validation, head-to-head comparison with immunohistochemistry, and eventually FDA, EMA or Anvisa authorization. Explainability is part of that package: regulators expect clear rationales for which descriptors drive the prediction, and patient advocacy groups have asked for transparency on which subgroups the model was trained on. Sex, age, ancestry and prior treatment all interact with both tumor biology and MRI signal, and a model trained on a North American cohort can drift quietly when deployed in Brazil or Mexico.
What the ISMRM 2026 talk does change is the conversation. It moves PD-L1 prediction from “interesting research” to “actively in the pipeline,” which is the trigger for vendors and academic centers to invest in standardized acquisition, dataset sharing and reference benchmarks. The next two years will likely see the first prospective trials reading PD-L1 from MRI alongside biopsy, with response to pembrolizumab as the outcome — the moment when an imaging biomarker actually steers therapy rather than merely correlating with it. The ISMRM 2026 presentation marks an important — but not final — step toward an imaging biomarker that could reshape decision flow in breast cancer.
Source: AuntMinnie — ISMRM: Predictive model evaluates PD-L1 status in breast cancer (May 10, 2026)
