← Back to Home

Lung Cancer PD-L1 Scoring Shows Only 70% Agreement: AI and Human Assessments Diverge Near the Low-Expression Threshold

A study of 993 lung cancer patients highlights differences near pathology scoring thresholds. These conference data have yet to establish whether AI or pathologists better predict immunotherapy benefit.

By SURL BioNews

A scoring threshold in cancer tissue may influence how physicians consider immunotherapy, but AI and pathologists may not assign the same classification to the same specimen. A study involving 993 Japanese patients with non-small cell lung cancer found 73% agreement between their PD-L1 classifications, with differences concentrated particularly between negative and low expression.

The LC-SCRUM-Advantage/MRD study used surgically resected specimens, independently scored by pathologists and Lunit SCOPE PD-L1. It is scheduled for presentation as poster 313P at the European Society for Medical Oncology (ESMO) 2026 Congress on October 25. Japan’s National Cancer Center Hospital East also listed the comparative study in its official program and confirmed Yoshitaka Zenke of the Department of Respiratory Medicine as the presenter.

PD-L1 is one of the biomarkers used when evaluating cancer immunotherapy. The tumor proportion score (TPS) used in the study describes the proportion of tumor cells showing PD-L1 staining. According to News1, agreement between AI and human assessments was 80%, 95%, and 51% in the ≥50%, 1–49%, and <1% expression ranges, respectively. In other words, disagreement was most likely in the group assessed as negative by human readers.

Details released by Lunit further delineate this boundary: among 211 tumors assigned a TPS below 1% by human readers but at least 1% by AI, 204 had AI scores of 1–10%. This indicates that the vast majority of cases crossing the negative threshold in this dataset still fell within the low-expression range. However, a difference in classification alone cannot establish which assessment is more accurate, much less directly imply that AI identified more patients who would benefit from immunotherapy.

OncoDaily republished commentary by Aakash Desai, an oncologist at the University of Alabama at Birmingham in the United States, cautioning that these data came from surgical specimens, which differ from the small biopsy specimens commonly used in advanced lung cancer, and that the findings cannot be directly applied to that setting. The study abstract also did not establish which scoring method better predicts immunotherapy benefit. Determining whether differences that cross the 1% threshold have clinical significance will still require linking scores to actual treatment outcomes.

This lung cancer analysis is one of five SCOPE platform studies announced by Lunit on this occasion. Another, the AIforHER project led by France’s Institut de Cancérologie de l’Ouest (ICO), compares AI tools’ ability to identify low and ultralow HER2 expression in breast cancer. According to the company’s announcement, Lunit has advanced to subsequent evaluation in clinical settings. Both studies address a challenge in pathology measurement: as treatment research increasingly focuses on weak biomarker expression, differences in assessments near classification boundaries also require more careful validation.

These findings currently remain conference research. Lunit explicitly states that SCOPE products are for research use only and must not be used in diagnostic procedures or decision-making. The practical value of these data lies in identifying scoring ranges that need further examination. Moving from identifying differences to a reliable clinical tool will require supporting evidence from different specimen settings and treatment outcomes.

References

  1. Lunit
  2. National Cancer Center Hospital East, Japan
  3. News1
  4. OncoDaily