Predicting EGFR mutational status from pathology images using a real-world dataset

Abstract

Treatment of non-small cell lung cancer is increasingly biomarker driven with multiple genomic alterations, including those in the epidermal growth factor receptor (EGFR) gene, that benefit from targeted therapies. We developed a set of algorithms to assess EGFR status and morphology using a real-world advanced lung adenocarcinoma cohort of 2099 patients with hematoxylin and eosin (H&E) images exhibiting high morphological diversity and low tumor content relative to public datasets. The best performing EGFR algorithm was attention-based and achieved an area under the curve (AUC) of 0.870, a negative predictive value (NPV) of 0.954 and a positive predictive value (PPV) of 0.410 in a validation cohort reflecting the 15% prevalence of EGFR mutations in lung adenocarcinoma. The attention model outperformed a heuristic-based model focused exclusively on tumor regions, and we show that although the attention model also extracts signal primarily from tumor morphology, it extracts additional signal from non-tumor tissue regions. Further analysis of high-attention regions by pathologists showed associations of predicted EGFR negativity with solid growth patterns and higher peritumoral immune presence. This algorithm highlights the potential of deep learning tools to provide instantaneous rule-out screening for biomarker alterations and may help prioritize the use of scarce tissue for biomarker testing.

Deep Learning for Lung Cancer Diagnosis, Prognosis and Prediction Using Histological and Cytological Images: A Systematic Review

“…The highest AUC (0.917) for ALK rearrangements prediction was in the study by Chen et al [31], while EGFR mutations were predicted with an AUC of 0.824, 0.84, and 0.83 in the studies by Wang, Yang, and Coudray et al, respectively [31,50,56,87]. In the most recent study by Pao et al, the prediction of EGFR mutational status in 2099 lung ADC tissue specimens reached an AUC of 0.87 [95]. As far as PD-L1 quantification is concerned, the majority of studies included datasets consisting of WSIs stained with the 22C3 antibody.…”

“…For example, clustering-constrained-attention MIL (CLAM), developed by Lu et al [47], is a weakly supervised method that uses attention-based learning to automatically identify subregions of high diagnostic value and, thus, accurately classify the whole slide. Other works combine the MIL approaches with well-known architectures of CNNs, such as ResNet [48,65,95], EfficientNetB1 [22], and SimCLR [59]. Moreover, Teramoto et al [101] compared several CNNs as backbones (LeNet, AlexNet, ResNet, Inception, DenseNet) using MIL and an attention mechanism, while Hou et al [57] presented 14 different combinations of expectation maximization (EM)-based MIL approach with Logistic Regression and Support Vector Machine (SVM).…”

The future of artificial intelligence in thoracic surgery for non-small cell lung cancer treatment a narrative review

“…In parallel, Pao et al.’s ( 17 ) study illuminates the significant role of AI in enhancing intraoperative decision-making. DL algorithms are showcased for their ability to discern subtle pathological features intraoperatively, providing instantaneous insights into tissue histology.…”

“…( 16 ) and Pao et al. ( 17 ) demonstrated the proficiency of DL in predicting ALK rearrangements and EGFR mutations, respectively, achieving commendable Area Under the Curve (AUC) values. These findings lay a foundation for personalized treatment strategies based on the molecular characteristics of the tumor.…”

“…The integration of digital pathology, as advocated by Pao et al. ( 17 ), necessitates a paradigm shift in infrastructure, storage, and data-sharing practices. The ethical dimension becomes pronounced, with concerns regarding patient data privacy and the development of regulatory frameworks surfacing prominently.…”