AI analysis of elastin staining sections reveals the biological implications of cancer-related fibrosis in small-sized lung adenocarcinomas.

OBJECTIVE: Pathological "invasion" influences patient prognosis in lung adenocarcinoma (LUAD); however, diagnosis is often associated with high interobserver variability because non-lepidic adenocarcinoma (NLA) and cancer-related fibrosis (CRF) are intricately mixed, and CRF contains both invasive and non-invasive components. In this study, to enable quantitative and reproducible diagnosis of areas showing high intratumoral heterogeneity, we proposed an artificial intelligence (AI)-based analysis distinguishing between NLA and CRF. This approach could reduce interobserver variability in prognostic prediction and help examine the different biological meanings of NLA and CRF. METHODS: To clarify the physiological structure of lung parenchyma, we used elastin staining specimens. CRF and NLA were separately annotated in the first cohort (n = 35) and used for supervised learning. The AI then analyzed whole non-lepidic areas in the second cohort (n = 188); we then examined the relationship between clinicopathological features and the AI analysis. RESULTS: For the first cohort, the accuracy was 89.2% on average. For the second cohort, groups with high CRF ratios (>50%) in the AI analysis showed a statistical correlation with types B-C in Noguchi's classification (p < 0.001), grades 1-2 in the WHO grading system (p = 0.018), and a higher 5-year disease-free survival (DFS) rate (85.7% vs. 69.4%, p = 0.006). In groups with a low CRF ratio (n = 111), a central distribution of CRF was associated with a lower DFS rate (41.2% vs. 74.5%, p = 0.001). CONCLUSION: The AI engine for distinguishing between NLA and CRF enabled prognostic prediction of LUAD with reduced interobserver variability. It may also be useful for examining the biological significance of NLA and CRF.