Artificial intelligence for lung adenocarcinoma growth-pattern classification on H&E whole-slide images: a diagnostic test accuracy meta-analysis of lepidic, acinar, papillary, micropapillary, and solid patterns

Ayan Myssayev, Amin Tamadon, Nadiar Maratovich Mussin, Reza Shirazi, Afshin Zare

Diagnostic Pathology, 2026

Полный текст Открытый доступ: Gold Метаданные (CSL-JSON)
DOI 10.1186/s13000-026-01810-x
Достоверность данных 70 средняя
Как считается эта оценка

Складывается из проверяемых признаков самой записи: есть ли у статьи DOI, сколько независимых источников подтвердили её данные, заполнены ли обязательные поля и насколько свежо последнее подтверждение. Снижают оценку расхождения между источниками и отсутствие идентификатора. Значение от 80 означает, что запись опубликована без дополнительной проверки; от 60 до 79 — что она опубликована с пониженной достоверностью и ручная проверка желательна. Запись, выверенную редактором вручную, пересчёт не опускает ниже 90. Это оценка наших данных о статье, а не оценка самой статьи.

Аннотация

Artificial intelligence (AI) may improve reproducible assessment of lung adenocarcinoma (LUAD) growth patterns on H&E whole-slide images, but pattern-specific diagnostic test accuracy remains uncertain. We systematically searched Embase, Europe PMC, PubMed, Scopus, Semantic Scholar, and Web of Science on May 1, 2026, for English-language studies evaluating AI-based classification of lepidic, acinar, papillary, micropapillary, and solid LUAD patterns. The search identified 844 records, and 20 studies were included in the quantitative synthesis. Only count-complete diagnostic estimates were entered into the inferential bivariate random-effects meta-analysis of logit sensitivity and specificity; SROC performance, likelihood ratios, diagnostic odds ratios, heterogeneity, QUADAS-2 quality, and exploratory Deeks asymmetry were assessed. AI classifiers showed high specificity across all patterns. The inferential synthesis included seven count-complete estimates for lepidic, acinar, and papillary patterns and six for micropapillary and solid patterns. Pooled sensitivity/specificity were 0.929/0.988 for lepidic, 0.899/0.971 for acinar, 0.793/0.984 for papillary, 0.759/0.991 for micropapillary, and 0.977/0.996 for solid patterns. SROC AUCs were approximately 0.994, 0.984, 0.981, 0.986, and 0.999, respectively. Solid classification showed the strongest diagnostic performance, with an LR + of 230.66, an LR- of 0.023, and a DOR of 9919.55. Papillary and micropapillary patterns showed excellent specificity but lower sensitivity. Heterogeneity was high, while exploratory Deeks testing showed no significant evidence of funnel asymmetry or small-study effects in the combined count-complete dataset. AI-based LUAD growth-pattern classification on H&E whole-slide images showed high diagnostic accuracy in published research settings, especially for solid, lepidic, and acinar patterns. Broader external validation, standardized reporting, prospective pathologist-in-the-loop studies, and careful evaluation in mixed-pattern tumors are needed before routine clinical implementation.