Deep learning signature from chest CT and association with immunotherapy outcomes in EGFR/ALK-negative NSCLC.

Abstract

9061 Background: Many clinicopathological and molecular features are associate with clinical benefit from immune checkpoint inhibitors (ICIs) for patients with non-small-cell lung cancer (NSCLC), yet none was exclusive underscoring the heterogeneity of lung cancers. As images may provide a holistic view of cancer, we attempted deep learning to chest CT scans to derive a predictor of response to ICIs and test its benefit relative to known clinicopathological factors. Methods: 928 stage IV, EGFR/ALK-negative NSCLC patients treated with ICIs alone or in combination (MD Anderson GEMINI Database) were divided into training (CTtr = 572), validation (CTva = 78), and testing (CTte = 278) cohorts, balancing the distribution of clinicopathological and radiological factors. Progression-free (PFS) and overall survival (OS) were defined as outcomes. We analyzed whole lung, including tumor and normal parenchyma of chest CT images ≤ 3 months prior to ICI treatment. An ensemble learning model (CT-deep-learning) to clustering patients into high vs low risk groups of PFS or OS was developed by fusing risk scores from four independent deep learning networks (supervised, unsupervised, and hybrid). This CT-deep-learning model was further evaluated in different clinicopathological subgroups. Finally, a composite model (CT-Clinic-path) was built by combining image model with clinicopathological factors. Antolini's concordance index (C-index) was used to assess model performance. Results: Median PFS and OS were shorter in the high-risk vs low-risk group as defined by CT-deep-learning: PFS (CTtr: 4.2 vs 9.6 mons; HR 1.96; 95% CI 1.62-2.38; P < 0.0001; CTva: 3.7 vs 10.2 mons; HR 2.32; 95% CI 1.32-4.07; P = 0.0025; CTte: 3.6 vs 9.1 mons; HR 1.89; 95% CI 1.39-2.56; P < 0.0001) and OS (CTtr: 16.0 vs 31.4 mons; HR 2.19; 95% CI 1.72-2.79; P < 0.0001; CTva: 12.7 vs 28.6 mons; HR 2.01; 95% CI 1.04-3.88; P = 0.035; CTte: 14.8 vs 32.0 mons; HR 1.84; 95% CI 1.31-2.60; P = 0.0004). CT-deep-learning outperformed clinicopathologic features known to associate with ICI benefit, such as histology, smoking status, PD-L1 expression, and remained to be an independent (P < 0.001) prognostic factor on multivariate analysis. Furthermore, integrating CT-deep-learning to clinicopathological variables improved prediction performance with a net reclassification up to 7% (Clinic-path model, C-indices 0.60 – 0.62 vs CT-clinic-path model, 0.64 - 0.65 for PFS; Clinic-path model 0.64 – 0.67 vs CT-clinic-path model 0.69 – 0.71 for OS). Conclusions: We have developed and validated a deep learning signature associated with PFS and OS in ICI-treated NSCLC patients, which appears to be independent of and superior to known clinicopathological biomarkers. If validated, this signature may strengthen the predictive value of clinicopathological factors and facilitate selecting appropriate patients for ICI-based therapies.