PET/CT Based EGFR Mutation Status Classification of NSCLC Using Deep Learning Features and Radiomics Features

Weicheng Huang, Jingyi Wang, Haolin Wang, Yuxiang Zhang, Fengjun Zhao, Kang Li, Linzhi Su, Fei Kang, Xin CaoFrontiers in Pharmacology 2022 · DOI

Abstract

Purpose: This study aimed to compare the performance of radiomics and deep learning in predicting EGFR mutation status in patients with lung cancer based on PET/CT images, and tried to explore a model with excellent prediction performance to accurately predict EGFR mutation status in patients with non-small cell lung cancer (NSCLC). Method: PET/CT images of 194 NSCLC patients from Xijing Hospital were collected and divided into a training set and a validation set according to the ratio of 7:3. Statistics were made on patients' clinical characteristics, and a large number of features were extracted based on their PET/CT images (4306 radiomics features and 2048 deep learning features per person) with the pyradiomics toolkit and 3D convolutional neural network. Then a radiomics model (RM), a deep learning model (DLM), and a hybrid model (HM) were established. The performance of the three models was compared by receiver operating characteristic (ROC) curves, sensitivity, specificity, accuracy, calibration curves, and decision curves. In addition, a nomogram based on a deep learning score (DS) and the most significant clinical characteristic was plotted. Result: In the training set composed of 138 patients (64 with EGFR mutation and 74 without EGFR mutation), the area under the ROC curve (AUC) of HM (0.91, 95% CI: 0.86-0.96) was higher than that of RM (0.82, 95% CI: 0.75-0.89) and DLM (0.90, 95% CI: 0.85-0.95). In the validation set composed of 57 patients (32 with EGFR mutation and 25 without EGFR mutation), the AUC of HM (0.85, 95% CI: 0.77-0.93) was also higher than that of RM (0.68, 95% CI: 0.52-0.84) and DLM (0.79, 95% CI: 0.67-0.91). In all, HM achieved better diagnostic performance in predicting EGFR mutation status in NSCLC patients than two other models. Conclusion: Our study showed that the deep learning model based on PET/CT images had better performance than radiomics model in diagnosing EGFR mutation status of NSCLC patients based on PET/CT images. Combined with the most statistically significant clinical characteristic (smoking) and deep learning features, our hybrid model had better performance in predicting EGFR mutation types of patients than two other models, which could enable NSCLC patients to choose more personalized treatment schemes.

Introduction

In non-small cell lung cancer (NSCLC), identifying epidermal growth factor receptor (EGFR) mutations is a critical step in clinical management. Patients harboring these mutations often show remarkable responses to targeted therapies, specifically tyrosine kinase inhibitors (TKIs). Currently, the clinical gold standard for identifying these mutations is tissue biopsy. However, biopsies are inherently invasive and carry risks such as pneumothorax. Furthermore, a single biopsy may fail to capture the full genetic profile of a heterogeneous tumor, potentially leading to incomplete diagnostic information.

Radiomics, the extraction of high-throughput quantitative data from medical images, has emerged as a promising non-invasive alternative. However, traditional handcrafted radiomics features often struggle to capture the complex, high-dimensional patterns inherent in tumor biology. This study investigates the potential of deep learning to overcome these limitations and explores whether a hybrid approach combining deep learning with clinical data can provide a superior diagnostic tool for predicting EGFR mutation status.

Methodology

The study utilized a retrospective dataset of 194 NSCLC patients from Xijing Hospital, collected between 2016 and 2021. The cohort was divided into a training set (n=138; 64 EGFR+, 74 EGFR-) and a validation set (n=57; 32 EGFR+, 25 EGFR-). All patients underwent standardized $^{18}$F-FDG PET/CT imaging.

Figure 2: The comprehensive workflow of the study, including ROI segmentation, feature extraction, selection, and model analysis.

The researchers developed and compared three distinct models:

  1. Radiomics Model (RM): Handcrafted features (shape, intensity, texture) were extracted using the pyradiomics toolkit, totaling 4,306 features from both PET and CT images.
  2. Deep Learning Model (DLM): A custom two-channel 3D Convolutional Neural Network (CNN) was employed. This network processed 64x64x64 voxel cuboids of the tumor from both PET and CT streams, extracting 2,048 deep learning features per patient.
  3. Hybrid Model (HM): This model integrated the deep learning score (DS) with statistically significant clinical characteristics, specifically smoking status.
Figure 1: The architecture of the 3D convolutional neural network used for feature extraction from PET/CT images.

To mitigate overfitting, a rigorous feature selection process was implemented. This included Mann-Whitney U tests to identify features with significant differences between mutation groups, followed by Least Absolute Shrinkage and Selection Operator (LASSO) regression to select the most predictive features.

Results

The performance of the models was primarily evaluated using the Area Under the ROC Curve (AUC). The results demonstrated that deep learning features significantly outperformed handcrafted radiomics features.

  • Hybrid Model (HM): Achieved the highest performance with an AUC of 0.85 in the validation set and an overall accuracy of 82%.
  • Deep Learning Model (DLM): Followed with an AUC of 0.79 in the validation set.
  • Radiomics Model (RM): Showed the lowest predictive power with an AUC of 0.68.

Smoking status was identified as the most significant clinical predictor, with non-smokers being more likely to have EGFR mutations. The integration of this clinical data into the Hybrid Model provided the best overall diagnostic performance. Decision Curve Analysis (DCA) further confirmed that the Hybrid Model offered the highest net benefit for clinical decision-making.

Discussion and Conclusion

This study demonstrates that a deep learning model based on PET/CT images provides superior performance compared to traditional radiomics in diagnosing EGFR mutation status in NSCLC patients. By leveraging a two-channel CNN, the model successfully synthesized metabolic activity from PET and structural anatomy from CT, capturing a more comprehensive tumor phenotype.

Despite these promising results, several limitations remain. The study's single-center design and focus on a specific ethnic population necessitate further multi-center validation to ensure generalizability. Additionally, the "black box" nature of deep learning features presents a challenge for clinical interpretability. Future work should focus on incorporating additional clinical biomarkers and exploring explainable AI techniques to enhance transparency for clinicians.

In conclusion, the hybrid model combining deep learning features with clinical smoking history offers a powerful, non-invasive "virtual biopsy" tool. This approach could significantly aid in the selection of personalized treatment schemes for NSCLC patients, potentially improving outcomes through more accurate and less invasive diagnostic methods.

Summary updated

Abstract

<b>Purpose:</b> This study aimed to compare the performance of radiomics and deep learning in predicting EGFR mutation status in patients with lung cancer based on PET/CT images, and tried to explore a model with excellent prediction performance to accurately predict EGFR mutation status in patients with non-small cell lung cancer (NSCLC). <b>Method:</b> PET/CT images of 194 NSCLC patients from Xijing Hospital were collected and divided into a training set and a validation set according to the ratio of 7:3. Statistics were made on patients' clinical characteristics, and a large number of features were extracted based on their PET/CT images (4306 radiomics features and 2048 deep learning features per person) with the pyradiomics toolkit and 3D convolutional neural network. Then a radiomics model (RM), a deep learning model (DLM), and a hybrid model (HM) were established. The performance of the three models was compared by receiver operating characteristic (ROC) curves, sensitivity, specificity, accuracy, calibration curves, and decision curves. In addition, a nomogram based on a deep learning score (DS) and the most significant clinical characteristic was plotted. <b>Result:</b> In the training set composed of 138 patients (64 with EGFR mutation and 74 without EGFR mutation), the area under the ROC curve (AUC) of HM (0.91, 95% CI: 0.86-0.96) was higher than that of RM (0.82, 95% CI: 0.75-0.89) and DLM (0.90, 95% CI: 0.85-0.95). In the validation set composed of 57 patients (32 with EGFR mutation and 25 without EGFR mutation), the AUC of HM (0.85, 95% CI: 0.77-0.93) was also higher than that of RM (0.68, 95% CI: 0.52-0.84) and DLM (0.79, 95% CI: 0.67-0.91). In all, HM achieved better diagnostic performance in predicting EGFR mutation status in NSCLC patients than two other models. <b>Conclusion:</b> Our study showed that the deep learning model based on PET/CT images had better performance than radiomics model in diagnosing EGFR mutation status of NSCLC patients based on PET/CT images. Combined with the most statistically significant clinical characteristic (smoking) and deep learning features, our hybrid model had better performance in predicting EGFR mutation types of patients than two other models, which could enable NSCLC patients to choose more personalized treatment schemes.
BibTeX
@inproceedings{huang2022pet,
  author = {Weicheng Huang and Jingyi Wang and Haolin Wang and Yuxiang Zhang and Fengjun Zhao and Kang Li and Linzhi Su and Fei Kang and Xin Cao},
  title = {{PET/CT Based EGFR Mutation Status Classification of NSCLC Using Deep Learning Features and Radiomics Features}},
  booktitle = {Frontiers in Pharmacology},
  year = {2022},
  doi = {10.3389/fphar.2022.898529},
}
Figure