EfficientNet-Based System for Detecting EGFR-Mutant Status and Predicting Prognosis of Tyrosine Kinase Inhibitors in Patients with NSCLC

Nan Xu, Jiajun Wang, Gang Dai, Tao Lu, Shu Li, Kexue Deng, Jiangdian SongJ. Imaging Inform. Medicine 2024 · DOI

TL;DR

This paper introduces an EfficientNet-based deep learning system (EME) that can non-invasively detect EGFR mutations in non-small cell lung cancer (NSCLC) patients from CT scans and predict their prognosis when treated with EGFR tyrosine kinase inhibitors (TKIs). The EME system outperforms traditional radiomics approaches and offers biologically interpretable insights into treatment response.

What Problem Does This Paper Address?

Lung cancer, specifically non-small cell lung cancer (NSCLC), is a major global health concern. Identifying epidermal growth factor receptor (EGFR) mutations is crucial for guiding treatment with EGFR tyrosine kinase inhibitors (TKIs), which have significantly improved patient outcomes. However, current methods for detecting EGFR mutations rely on invasive tumor tissue biopsies. These biopsies can be impractical for advanced-stage patients, and the genetic heterogeneity of NSCLC can lead to unreliable results. Furthermore, not all patients with EGFR mutations respond to TKIs, and predicting who will benefit remains a challenge. Existing deep learning models often focus on a single task (mutation detection or prognosis) and lack biological interpretability. This paper aims to address these limitations by developing a single, non-invasive deep learning system that can both detect EGFR mutations and predict TKI treatment prognosis, while also providing insights into the underlying biological mechanisms.

Methods

The researchers developed an EfficientNetV2-L-based deep learning model, termed EME (EfficientNet-based Model for EGFR), for predicting EGFR-mutant status. The model was trained on pre-therapy computed tomography (CT) images. To capture relevant features, patches of the tumor region and its surrounding 10mm periphery were extracted and resized. For prognostic stratification of EGFR-mutant patients receiving TKIs, a separate EME-prognostic model was constructed using deep learning features extracted from the EME model, followed by LASSO Cox regression and stepwise regression for feature selection.

To compare performance, traditional radiomics features were extracted using the PyRadiomics library. Fusion models combining EME and radiomics features were also created. To enhance interpretability, the researchers employed Gradient-weighted Class Activation Maps (Grad-CAM) to visualize the regions of the CT images that the EME model focused on. Additionally, t-distributed Stochastic Neighbor Embedding (t-SNE) was used to reduce feature dimensions and visualize the relationships between EME features and EGFR mutation status. Finally, Gene Set Variation Analysis (GSVA) was performed on data from The Cancer Genome Atlas (TCGA) lung adenocarcinoma cohort to explore the biological associations between the EME prognostic score and EGFR-related pathways.

Datasets

The study utilized a retrospective, multicenter dataset of 485 patients with NSCLC from four hospitals in China. This dataset was split into a training set (339 patients) and an independent test set (146 patients). For biological correlation analysis, the TCGA lung adenocarcinoma (TCGA-LUAD) dataset was used, which includes CT images and RNA sequencing data.

Key Insights and Ideas

  • Non-invasive EGFR Mutation Detection: The EME system demonstrates that CT images, analyzed by deep learning, can accurately detect EGFR mutation status without the need for invasive biopsies. This is a significant step towards more accessible and less burdensome diagnostic procedures.
  • Predicting TKI Treatment Response: Beyond mutation detection, the EME-prognostic model effectively stratifies patients into high- and low-risk groups for TKI therapy, suggesting it can help identify patients who are more likely to benefit from these treatments.
  • Biologically Interpretable AI: The use of Grad-CAM and t-SNE visualizations, along with the correlation analysis with TCGA data, provides crucial insights into why the EME model makes its predictions. This moves beyond the "black box" nature of many deep learning models.
  • Richer Heterogeneity in Mutant Tumors: Heat maps of EGFR-mutant patients showed richer heterogeneity in intratumoral and peritumoral areas, suggesting this heterogeneity is a key imaging marker associated with EGFR mutations.
  • EME Outperforms Radiomics: The EME model consistently outperformed traditional radiomics approaches in both EGFR mutation detection and prognostic prediction, highlighting the power of deep learning for complex image analysis.

Key Results

  • EGFR Mutation Detection: The EME model achieved an Area Under the Curve (AUC) of 0.907 (95% CI 0.840–0.926) on the test dataset for predicting EGFR-mutant status, which was significantly superior to the radiomics model (AUC 0.825, P=0.007). A fusion model achieved a slightly higher AUC (0.941) but was not statistically significantly better than EME alone.
  • Prognostic Stratification: The EME-prognostic model demonstrated strong performance in stratifying PFS, achieving a C-index of 0.711 in the test dataset. Patients with an EME-prognostic score below a determined threshold showed significantly better PFS.
  • Biological Correlations: The EME-prognostic score showed significant negative correlations with key biological pathways related to EGFR signaling and EGF stimulus, suggesting that higher EME scores (indicating poorer prognosis) are associated with increased activity in these pathways.

Limitations and Future Work

  • Limited Ethnicity: The study primarily included Asian patients, limiting the generalizability of the EME model to other ethnicities. Future work should focus on validating the model across diverse populations.
  • EGFR Subtypes: The study did not deeply investigate the prognostic implications of specific EGFR mutation subtypes (e.g., exon 19 deletions vs. L858R). Further research into these subtypes is recommended.
  • Histopathological Integration: The authors suggest that incorporating histopathological images could provide more comprehensive analysis of EGFR-mutant NSCLC patients.
  • Clinical Feature Fusion: Future studies could benefit from fusing clinical features with the EME model to further improve prognostic prediction.
  • In Vitro/In Vivo Validation: While biological correlations were found, experimental validation (in vitro or in vivo) is needed to substantiate these associations.

Summary updated

BibTeX
@article{xu2024efficientnet,
  author = {Nan Xu and Jiajun Wang and Gang Dai and Tao Lu and Shu Li and Kexue Deng and Jiangdian Song},
  title = {{EfficientNet-Based System for Detecting EGFR-Mutant Status and Predicting Prognosis of Tyrosine Kinase Inhibitors in Patients with NSCLC}},
  journal = {J. Imaging Inform. Medicine},
  year = {2024},
  doi = {10.1007/s10278-024-01022-z},
}
Figure