Artificial intelligence in digital pathology approach identifies the predictive impact of tertiary lymphoid structures with immune-checkpoints therapy in NSCLC.

Abstract

9065 Background: The presence of Tertiary Lymphoid Structures (TLS) in multiple cancer types has been recognized as a potential predictive biomarker for response to immune-checkpoint blockade. However, there is no standardized method to quantify their presence. In this context, Artificial Intelligence (AI)-based assessment of histology images may well contribute to improve reproducibility, accuracy and speed of TLS quantification. Methods: We developed an automated workflow for quantification of TLS on digitized H&E slides through A) pixel-level classification of tissue using supervised artificial neural networks model, B) object-level cell classification of candidate TLS regions, C) merging the two approaches for curation and validation of TLS versus non-TLS regions. 433 advanced stage non-small cell lung cancer (NSCLC) patients treated with first or subsequent line of anti-PD-(L)1 single agent at DFCI were included in this study. Results: TLS were detected in 37% (n = 161) of the patients H&E slides, with the highest score of 4.7 TLS per mm2 (interquartile range: Q1 = 0, Q2 = 0, Q3 = 0.03 TLS/mm2). TLS density (per mm2) was significantly higher in surgically resected (n = 246; TLSPOS= 49%) compared to bioptic samples (n = 187; TLSPOS= 21%). No association was observed between TLS and tumor mutational burden (TMB) or PD-L1 protein expression as continuous variables. Among clinically actionable mutations, EGFR (all subtypes) mutated patients (n = 38) had a significantly lower number of TLS compared to patients without EGFR mutations. Patients with ≥ 0.01 TLS/mm2 had a significantly higher objective response rate (32% vs 22%, p = 0.03), a significantly longer median progression-free survival (PFS, 4.8 vs 2.7 months, HR: 0.73, 95% CI: 0.59-0.90, p = 0.004), and a significantly improved median overall survival (OS, 16.5 vs 12.5 months, HR: 0.72, 95% CI: 0.57-0.92, p = 0.008). In multivariable analysis, after adjusting for PD-L1 (≥ vs < 50%), TMB (≥ vs < 10 mu/Mb), sex, age, ECOG score, smoking and line of treatment, TLS/mm2 (≥ vs < 0.01) levels were found to be an independent positive predictive factor for both PFS (HR:0.69, 95% CI: 0.54-0.88, p = 0.003) and OS (HR: 0.70, 95% CI: 0.52-0.93, p = 0.01). Conclusions: These findings suggest that TLS status is an independent predictor of immunotherapy effectiveness in NSCLC, with predictive value similar to that of PD-L1 expression and TMB. This novel AI system has potential for automated identification and quantification of the TLS on digital histopathological slides, and could be utilized in a standard pathology workflow with relative ease. These findings are currently being validated in other solid tumors and cohorts.

The Emerging Role of Tertiary Lymphoid Structures in Breast Cancer: A Narrative Review

“…Lastly, digital and computational pathology, integrating deep learning and artificial intelligence, can prove instrumental for both identifying and quantifying TLSs [59]. Multiplex fluorescent immunohistochemistry/immunofluorescence (mIHC/IF) has demonstrated superior performance compared with immunohistochemistry.…”

B Cells in Breast Cancer Pathology

“…Tools to automatically annotate TLS in breast carcinomas have not yet been developed. For lung cancer patients, deep learning-based approaches to capture the presence and frequency of TLS have shown their potential as a predictor of tumour recurrence [87][88][89].…”

The Emerging Role of Tertiary Lymphoid Structures in Breast Cancer: A Narrative Review

“…Lastly, digital and computational pathology, integrating deep learning and artificial intelligence, can be instrumental in both identifying and quantifying TLSs [48]. Multiplex fluorescent immunohistochemistry/immunofluorescence (mIHC/IF) has demonstrated superior performance compared to immunohistochemistry.…”