Abstract LB504: Automated grading of growth patterns in lung adenocarcinoma-from TRACERx to LATTICe-A

Abstract Lung adenocarcinoma exhibits distinct growth patterns (WHO, 2021) and the International Association for the Study of Lung Cancer (IASLC) grading system, based on the nature and proportion of histologic subtypes, is highly prognostic (Moreira et al. 2020). However, both recognition and quantification of growth patterns suffer from high interobserver variability among pathologists. Here, we present a large-scale application to automate the segmentation of histologic patterns and reproduce the IASLC grading system to stratify patients and predict prognosis. A deep learning model was trained to recognize and segment 6 histologic patterns (lepidic, acinar, papillary, micropapillary, solid, and cribriform) on 49 whole-slide images from TRACERx (AbdulJabbar et al. 2020). This model was directly applied to an independent cohort, consisting of 4324 hematoxylin and eosin-stained sections from 970 patients (the Leicester Archival Thoracic Tumor Investigatory Cohort, Moore et al. 2019). Growth pattern segmentation performance was first evaluated against 2433 hand annotations covering 6 patterns from 9 images at a pixel level using Dice coefficient. The average Dice was 0.502. Predicted predominant pattern per tumor was then compared to a subspecialty pathologist, achieving an overall agreement (51%), comparable to the interobserver rate among pathologists (52%, Boland et al. 2017). Discordant cases were more heterogeneous (p=4.5e-10, Shannon diversity based on pathological scores), underscoring the challenges posed by intratumor heterogeneity. Patients with a high proportion of micropapillary in the tumor, identified by deep learning, had significantly worse relapse-free survival (RFS) in multivariate analyses including clinical parameters (p=0.00127, Hazard Ratio (HR)=6.4, 95% confidence interval (CI)=2.07-19.8, n=827), consistent with previous publication (Cha et al. 2014). The Kaplan-Meier curve for RFS was significantly differentiated with both automated and pathological IASLC grading (p<0.0001). Moreover, patients with predominantly high-grade patterns (solid, micropapillary, cribriform) identified by deep learning had significantly reduced RFS (p=6.22e-4, HR=1.5, 95% CI=1.18-1.8, n=970). The prognostic effect was stronger using IASLC grading (cutoff: 20%, p=5.56e-6, HR=1.7, 95% CI=1.36-2.2), comparable to the pathological score at IASLC grade 3 (cutoff: 20%, p=2.41e-5, HR=1.8, 95% CI=1.35-2.3). A similar performance was observed for overall survival regarding above analyses. To the best of our knowledge, this study represents the largest application of deep learning to recognize tumor growth patterns in lung adenocarcinoma. Histologically heterogeneous growth patterns can be automatically identified using a method trained on an independent cohort. Automated tumor grading is significantly associated with patient outcomes, supporting its potential clinical utility. Citation Format: Xiaoxi Pan, Khalid AbdulJabbar, Jose Coelho-Lima, Anca-Ioana Grapa, Hanyun Zhang, Ho Kwan Alvin Cheung, Sarah J. Aitken, David A. Moore, Charles Swanton, John Le Quesne, Yinyin Yuan. Automated grading of growth patterns in lung adenocarcinoma-from TRACERx to LATTICe-A [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr LB504.