This study was approved by the ethics committee of our hospital. A total of 230 patients were enrolled from center 1 between January 2017 and December 2021 and served as the primary cohort. 80 patients were enrolled from center 2 (between Jul. 2014 and Feb. 2022), and 78 patients were enrolled from center 3 (between Jan. 2020 and Dec. 2022), and served as the external validation cohort 1 and 2, respectively. The Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 (Eisenhauer et al., 2009) was used to determine treatment response to EGFR-TKI therapy. The inclusion criteria include (Sculier, 2013): underwent complete T1CE and T2W brain MRI scans before treatment, and (Boire et al., 2020) had complete gene test results. The exclusion criteria include (Sculier, 2013): with poor MRI image quality (Boire et al., 2020); age less than 18, and (Arbour and Riely, 2019) carrying a primary brain tumor or other tumor diseases. Patients from center 1 were divided into training and internal validation cohorts by random stratified sampling in a ratio of 8:2. Patients from centers 2 and 3 were used as independent sets to validate our DL methods. Figure 1 shows the screening process for patients from all three centers.

Patients from center 1 were scanned by a 3.0-T MRI scanner (Siemens Verio, Erlangen, Germany), patients from center 2 were scanned by a 3.0-T MRI scanner (Siemens Magnetom Skyra, Erlangen, Germany), and patients from center 3 were scanned by a 3.0-T MRI scanner (Philips, Ingenia). In center 1, the T1CE MRI scanning parameters were as follows: Repeat time (TR) = 270 ms; Echo time (TE) = 2.48 ms; slice thickness = 5 mm, FOV = 194 × 230 mm, and matrix size = 320 × 216. The T2W MRI scanning parameters were as follows: TR = 3630 ms, TE = 87 ms; slice thickness = 5 mm; FOV 194 × 230 mm, and matrix size = 384 × 227 mm. In center 2, the T1CE MRI scanning parameters were as follows: TR = 1400 ms; TE = 9 ms; slice thickness = 6 mm, FOV = 179 × 230 mm and matrix size = 320 × 187. The T2W MRI scanning parameters were as follows: TR = 3500 ms, TE = 99 ms; slice thickness = 6 mm; FOV = 194 × 230 mm and matrix size = 320 × 270 mm. T1CE MRI images were taken 5 min after Gd-DTPA injection. In center 3, the parameters of T1CE and T2W MRI were as follows: T1CE: TR = 180 ms; TE = 2.3 ms; slice thickness = 6 mm, and matrix size = 256 × 256. T2W: TR = 2000 ms; TE = 80 ms; slice thickness = 6 mm, and matrix size = 256 × 256. The dose was 0.2 mL/kg, and the injection speed was 3 mL/s. The segmentation of regions of interest (ROIs) of the brain metastasis (BM) was performed using the ITK-SNAP (version 3.6.1). A radiologist with 5 years’ experience was invited to manually segment the ROI of BM, who was blinded to the clinicopathological information of the patients, except for the tumor location. And a senior radiologist with 15 years’ experience was invited to validate all manual delineations. Volume of peritumoral edema (VPE) was calculated using ITK-SNAP.

The proposed automated artificial intelligence EGFR-TKI system (ETS) consists of two main components: (i) automatic tumor region segmentation and (ii) EGFR genotype prediction. The EGFR-Model of ETS can automatically recognize the region of interest (ROI), and directly predict the EGFR mutation status. For patients with EGFR mutation, the TKI-Model of ETS predicts response to EGFR-TKI therapy. The architecture of the ETS is shown in Figure 2.

The proposed automated artificial intelligence EGFR-TKI system (ETS) consists of two main components: (i) automatic tumor region segmentation and (ii) EGFR genotype prediction. Specifically, ETS first segments the brain metastasis region using a modified FC-DenseNet with LeakyReLU and external attention (EA), and then predicts EGFR mutation status using a DenseNet-121–based classifier. For patients with EGFR mutation, the system further predicts the response to EGFR-TKI therapy. The architecture of ETS is shown in Figure 2.

The segmentation subnetwork for the ETS is based on the FC-Densenet (Jégou et al., 2017) backbone and uses the LeakyRelu nonlinear activation function to replace the ReLU nonlinear activation function. In addition, an EA is added to the network’s downsampling and upsampling process (Guo et al., 2023). To train the segmentation network, we first performed data augmentation to increase the diversity of training samples and improve the robust performance of the training model. Each MRI image is randomly rotated by 90 degrees, and in addition, each image is randomly selected for data enhancement by one of three non-rigid body transformations: Elastic transform, Grid distortion, and Optical distortion. In the training process, the model is optimally trained by adaptive moment estimation (Adam) (Kinga and Adam, 2015) with a learning rate of 0.0001, the total number of iterations of the training model is 100, and the input size of the model is 128 × 128 × 3.

The classification subnetwork uses the Densenet-121 (Huang G. et al., 2017) as the backbone network. The fully connected layer of the Densenet-121 was replaced with the global average pooling (GAP) (Lin et al., 2025) for discriminating the EGFR mutation status. We applied the ideology of transfer learning, where the classification network was pre-trained on the ImageNet-1k dataset to increase the learning efficiency of the network. We evaluated four model variants, No Seg–VPE, No VPE, No Seg, and Seg–VPE—to isolate the contributions of the segmentation network and the volumetric peritumoral edema (VPE) feature.

To predict EGFR-TKI therapy response, we extracted DL features and handcrafted features from patients with EGFR mutation. The analysis of variance (ANOVA) and principal component analysis (PCA) (Witten et al., 2013) were applied to dimensionality reduction and screen features. Finally, we used a decision tree model to predict the response to EGFR-tyrosine TKI therapy. To enhance interpretability and reveal spatial correlations between image regions and prediction results, we applied Grad-CAM (Selvaraju et al., 2017) to the final convolutional layer of the DenseNet-121 classifier. This allowed us to visualize the discriminative regions that most influenced the EGFR mutation prediction. Since the classifier receives input features extracted from the segmented tumor region, the resulting attention maps reflect localized regions within the BM that are most relevant to the model’s decision-making process. In the training process, the model is optimally trained by adaptive moment estimation (Adam) (Kinga and Adam, 2015) with a learning rate of 0.0001; the epoch of the training model was set to 100. All DL experiments were performed in Python (v.3.6) using Keras (version 2.3) on a single GPU (Nvidia GeForce 3090) workstation.

To validate the predictive performance of the ETS for both EGFR‐mutation status and EGFR‐TKI response, we conducted independent evaluations on three datasets: an internal hold-out set (20% of the development data) and two external validation cohorts. The fully trained ETS was applied to each dataset. For each task and each cohort (Internal Validation, External Validation 1, External Validation 2), we generated receiver operating characteristic (ROC) curves and calculated the area under the curve (AUC), accuracy, F1 score, precision, and recall. Optimal decision thresholds were selected by maximizing Youden’s index.

All statistical analysis was performed in R software (version 3.6.0). ANOVA was performed for continuous variables, and the chi-square test was used for discrete (categorical) variables. Factors with a p-value less than 0.05 were considered statistically significant. The performance of the ETS was evaluated using area under the curve (AUC), accuracy, F1 score, precision, and recall. All evaluation metrics were implemented in Python (v.3.6) using the scikit-learn library. The Gradient Weighted Class Activation Map (Grad-CAM) was implemented on PyTorch (Version 1.12.0). Figure 3 depicts the workflow of our study.

Table 1 listed demographic and clinical characteristics of the patients with BM originated from primary NSCLC. From Table 1, there was no statistical significance in terms of age, gender, and smoking history.

Table 2 compared the performance of the proposed EGFR-Model^{No Seg−VPE}, EGFR-Model^{No VPE}, EGFR-Model^{No Seg} and EGFR-Model^Seg-VPE for predicting the EGFR mutation status. Without the subnetwork for segmentating the BM, the EGFR-Model^{No Seg−VPE} yielded lower AUCs, accuracy, F1-score, precision, and recall compared with EGFR-Model^{No VPE} in primary and external cohorts. The decreased predictive performance in EGFR-Model^{No Seg−VPE} suggested the necessity of the segmentation subnetwork. By integrating VPE, the EGFR-Model^{No Seg} showed better performance than EGFR-Model^{No Seg−VPE} in terms of AUC, accuracy, F1-score, precision, and recall. This indicated that the VPE can provide additional information to improve the capability of predicting the EGFR mutation status. The EGFR-Model^Seg-VPE, integrating both VPE and segmentation subnetworks, performed the best among all models for predicting the EGFR mutation status. ROC curves of all models on primary and external sets were shown in Figure 4. As shown in Figure 5, the Grad-CAM heatmaps highlight high-response areas within the segmented tumor region, indicating that the prediction of EGFR mutation status is driven by biologically relevant features. These results illustrate a link between the model architecture, particularly the segmentation-guided feature extraction, and the spatial mapping of predictive regions.

Table 3 compared the performance of the proposed TKI-Model^{No Seg−VPE}, TKI-Model^{No VPE}, TKI-Model^{No Seg} and TKI-Model^Seg-VPE for predicting response to EGFR-TKI. The TKI-Model^{No Seg−VPE} without the subnetwork for segmenting the BM genarated lower AUC and ACC compared with TKI-Model^{No VPE} that has the segmentation subnetwork. The result indicates the necessity of the segmentation subnetwork. The TKI-Model^{No Seg} integrating VPE outperformed the TKI-Model^{No Seg−VPE} that is without VPE in terms of AUC and ACC in primary and external cohorts. This suggested that the VPE holds additional information correlated to the efficacy of EGFR-TKI. The TKI-Model^Seg-VPE integrating both VPE and the segmentation subnetwork achieved the best predictive performance with AUCs of 0.747, 0.726, and 0.728 in the internal validation, external validation 1 and external validation 2 cohort, respectively. Figure 6 depicted the ROC curves of the TKI-Model for predicting response to EGFR-TKI.