Güncel Patoloji Dergisi 2026, Cilt 10, Sayı 2
Türkçe Özet Özet PDF Benzer Makaleler Yazara Mail Gönder
Pathologist–Artificial Intelligence Concordance in PD-L1 Scoring in Non-Small Cell Lung Carcinoma: A Single-Center Validation Study
Sevilay ÖZMEN 1, Demir Kürşat YILDIZ 2, Rabia DEMİRTAŞ 1, Berk MADEN 1, Burçin ERGÜL 3, Numan BULUT 4, Özge BOZKURT 5, Serenay DEMİRCİ 6, Arzu AKÇAY 7, Ömer ARAZ 8, Yakup HİLAL 9, Salim Başol TEKİN 10
1 Atatürk Üniversitesi Tıp Fakültesi, Tıbbi Patoloji Anabilim Dalı, Erzurum, Türkiye
2 Van Yüzüncü Yıl Üniversitesi Tıp Fakültesi, Tıbbi Patoloji Anabilim Dalı, Van, Türkiye
3 Bursa Şehir Hastanesi Tıbbi Patoloji Laboratuvarı, Bursa, Türkiye
4 Erzurum Şehir Hastanesi Tıbbi Patoloji Laboratuvarı, Erzurum, Türkiye
5 Ankara Atatürk Sanatoryumu Eğitim ve Araştırma Hastanesi Patoloji Laboratuvarı, Ankara, Türkiye
6 Iğdır Dr. Nevruz Erez Devlet Hastanesi Tıbbi Patoloji Laboratuvarı, Iğdır, Türkiye
7 Yeni Yüzyıl Üniversitesi Tıp Fakültesi, Tıbbi Patoloji Anabilim Dalı, İstanbul, Türkiye
8 Atatürk Üniversitesi Tıp Fakültesi, Göğüs Hastalıkları Anabilim Dalı, Erzurum, Türkiye
9 Kars Harakani Devlet Hastanesi Patoloji Laboratuvarı, Kars, Türkiye
10 Acıbadem Bursa Hastanesi Onkoloji Kliniği, Bursa, Türkiye
DOI : 10.5146/jcpath.2026.728 Objective: Classification of programmed death ligand 1 (PD-L1) tumor proportion scores at clinical thresholds may vary by evaluator. This study assessed concordance between the primary pathologist assessment (KY-1) and the Gemini 3.1 Pro artificial intelligence model (YZ), and explored a blinded second pathologist assessment (KY-2) in discordant cases. Material and Methods: This single-center retrospective validation study included 25 archival cases. Four-micrometer formalin-fixed paraffin-embedded sections were stained with CAL10 and ultraView Universal DAB on a BenchMark ULTRA platform. At least 10 Full HD images covering all tumor-bearing areas and collectively containing at least 100 viable tumor cells were captured per case using a 20× objective. Each Turkish prompt was run once, and a single YZ output was recorded. Scores were categorized as <1%, 1–49%, or ≥50%. KY-1–YZ agreement was assessed with Cohen’s kappa and exact McNemar tests; the KY-2 subset was analyzed exploratorily. Results: KY-1 classified 15 cases as negative, 6 as low-positive, and 4 as high-positive; YZ classified 14, 5, and 6 cases, respectively. Exact agreement was 16/25 (64.0%), Cohen’s kappa was 0.377, and quadratically weighted kappa was 0.531. YZ assigned a higher category in five cases and a lower category in four. Marginal differences were absent at the 1% and 50% thresholds (p=1.000; p=0.625). In eight discordant cases with KY-2 data, KY-2 matched KY-1 in seven cases (kappa = 0.784); the KY-2–YZ kappa was -0.366. Conclusion: KY-1 and YZ showed limited-to-moderate agreement, without a systematic directional difference at clinical thresholds. KY-2 reproduced KY-1 in most selected discordant cases. Because no independent multi-reader consensus standard was available, YZ accuracy could not be established. Local validation, standardized imaging and prompting, and blinded adjudication are required before clinical implementation. Anahtar Kelimeler : Carcinoma, Non-small-cell lung, Programmed cell death 1 ligand 1 protein, Immunohistochemistry, Observer variation, Artificial intelligence