Abstract PO-088: A deep-learning based approach to assess heterogeneity of histologies in non-small cell lung cancer

Abstract

Abstract Adenosquamous carcinomas of the lung carry a poor prognosis compared to other non-small cell lung cancers (NSCLC) (Tian. 2017, Baine et al. 2018, Wood et al. 2017, 2015 WHO Classification of Lung Tumors). Adenosquamous carcinoma (ASC) has features of both Adenocarcinoma (ADC) and Squamous cell carcinoma (SCC) in the same tumor. The incidence of ASC varies between studies but is estimated to account for 0.4 to 4% of all lung cancers (2015 WHO Classification of Lung Tumors). Diagnosis of these cancers depends on several factors including adequate sampling of the tumor, careful review and objective interpretation of histologic criteria. An automated method for light microscopy review could help standardize and speed up correct identification of this important subtype and lead to better targeted therapies. In addition, an automated technique to quantify the relative contributions of histologies in heterogeneous tumors would lead to a better understanding of tumor biology. 476 tissue samples from NSCLC patients were scanned at 0.5pixel/um resolution, and the tumor area was selected by a pathologist. Each slide tumor area was divided into smaller image ‘patches’ of 512 × 512 pixels. We trained a Convolution Neural Network based on the Inception V3 and Resnet18 architecture using transfer learning and weakly-supervised learning, to classify tissue patches using the whole-slide level labels. Once the network was trained, a patch-level prediction was performed on 20 unseen test slides. A whole-slide diagnosis was performed by choosing the most common histology predicted among all patches extracted for each slide. In addition, a patch-based heatmap was created to identify the heterogeneity of histologies within a single tissue sample. The model’s whole slide prediction of ADC or SCC histologies on the test set, matched the pathologist’s whole-slide diagnosis, with an F1-score of 0.91 at the individual tile level and 1.0 (perfect) in slide level. The trained model also identified a few likely adenosquamous carcinomas, where slides were found to have patches with features from both ADC and SCC at high proportions. We are currently examining whether samples identified as heterogeneous can be confirmed by pathologists’ visual inspection, and whether we can visualize the network features that contribute to the patch-level classification decision. The development of an algorithm to identify mixed histologic subtypes may enable better selection of patients responsive to drug treatment and elucidate the biological mechanism leading to tumor heterogeneity. Citation Format: Reheman Baikejiang, Jennifer Giltnane, Eloisa Fuentes, Cleopatra Kozlowski. A deep-learning based approach to assess heterogeneity of histologies in non-small cell lung cancer [abstract]. In: Proceedings of the AACR Virtual Special Conference on Tumor Heterogeneity: From Single Cells to Clinical Impact; 2020 Sep 17-18. Philadelphia (PA): AACR; Cancer Res 2020;80(21 Suppl):Abstract nr PO-088.