Deep learning radiomics for the prediction of epidermal growth factor receptor mutation status based on MRI in brain metastasis from lung adenocarcinoma patients

Abstract

Background

Early and accurate identification of epidermal growth factor receptor (EGFR) mutation status in non-small cell lung cancer (NSCLC) patients with brain metastases is critical for guiding targeted therapy. This study aimed to develop a deep learning radiomics model utilizing multi-sequence magnetic resonance imaging (MRI) to differentiate between EGFR mutant type (MT) and wild type (WT).

Methods

In this retrospective study, 288 NSCLC patients with confirmed brain metastases were enrolled, including 106 with EGFR MT and 182 with EGFR WT. All patients were randomly divided into a training dataset (75%) and a validation dataset (25%). Radiomics and deep learning features were extracted from the brain metastatic lesions using contrast-enhanced T1-weighted (T1CE) and T2-weighted (T2W) MRI images. Features extraction and selection were performed using the least absolute shrinkage and selection operator (LASSO) and ResNet34. The predictive performance of the signatures for EGFR mutation status was assessed using receiver operating characteristic (ROC) curves and area under the curve (AUC) analyses.

Results

No significant differences were found between the training and validation datasets. A four-feature radiomics signature (RS) demonstrated excellent predictive accuracy for EGFR MT, with α-binormal-based and empirical AUCs of 0.931 (95% CI: 0.880–0.940) and 0.926 (95% CI: 0.877–0.933), respectively. Incorporating deep learning signature (DLS) further enhanced the model’s performance, achieving α-binormal-based and empirical AUCs of 0.943 (95% CI: 0.921–0.965) and 0.938 (95% CI: 0.914–0.962) in the training dataset. These findings were confirmed in the validation dataset, with AUCs of 0.936 (95% CI: 0.917–0.955) and 0.921 (95% CI: 0.901–0.941), demonstrating robust and consistent predictive performance.

Conclusions

The multi-sequence MRI-based deep learning radiomics model exhibited high efficacy in predicting EGFR mutation status in NSCLC patients with brain metastases. This approach, which integrates advanced radiological features with deep learning techniques, offers a non-invasive and accurate method for determining EGFR mutation status, potentially guiding personalized treatment decisions in clinical practice.

Background

Lung cancer remains one of the leading causes of cancer-related mortality worldwide, with lung adenocarcinoma (LADC) accounting for approximately 50% of all cases [1]. Brain metastasis (BM) occurs in about 30% of lung cancer patients during the course of the disease, often leading to poor prognosis. The median overall survival for these patients is limited to only 7 months [2, 3]. Currently, several therapeutic approaches are available for managing lung cancer with BM, including chemotherapy, whole-brain radiotherapy (WBRT), stereotactic radiosurgery (SRS), and surgical resection. However, each of these approaches has its own limitations [4,5,6]. Chemotherapy is often ineffective against BM, primarily because chemotherapeutic agents have difficulty crossing the blood-brain barrier (BBB) and reaching the metastatic lesions [7]. WBRT and SRS are associated with potential side effects such as cerebral edema and cognitive decline [8]. Additionally, surgical resection can be effective in certain cases, but it is invasive and often fails to completely remove the metastatic lesions, leading to high recurrence rates [9, 10]. Given the limitations of current treatments, there is an urgent need to explore novel therapeutic strategies to better address BM in lung cancer patients.

Cancer treatment has entered the era of precision medicine, where genetic mutations and phenotypic variations play a crucial role in predicting clinical outcomes. In patients with locally advanced or advanced non-small cell lung cancer (NSCLC), targeted therapy has emerged as a more effective treatment option compared to conventional chemotherapy [11]. Among the various genetic alterations, epidermal growth factor receptor (EGFR) mutations are particularly significant, serving as independent risk factors for the development of BM. Notably, patients with EGFR MT are more prone to BM than those with the EGFR WT [12]. EGFR tyrosine kinase inhibitors (EGFR TKIs), small-molecule targeted agents, can partially cross the BBB and have shown substantial efficacy in treating brain metastases in EGFR-mutant NSCLC patients [13]. Therefore, accurately determining the EGFR mutation status of brain metastases is critical for tailoring treatment strategies, optimizing therapeutic outcomes, and ultimately improving patient prognosis.

Relying solely on the EGFR mutation status of the primary lung cancer to determine the mutation status of BM is often unreliable. Studies have shown that up to 30–53% of gene expression between primary tumors and brain metastases are inconsistent [14, 15]. Although biopsy and pathological examination remain the gold standard for confirming genotype, these methods have several limitations. Traditional biopsy techniques are invasive and may not be feasible for tumors located in certain anatomical regions. Furthermore, biopsies typically sample only a small portion of the tumor, which may not fully capture the heterogeneity of the entire lesion [16]. While circulating tumor DNA (ctDNA) analysis offers a non-invasive alternative, it is limited by the risks of false negatives and high costs [17, 18]. Therefore, to accurately assess the EGFR mutation status of brain metastases, biopsy or surgical resection followed by pathological examination is often necessary. However, these procedures carry inherent risks, including postoperative bleeding, pulmonary embolism, and cerebrospinal fluid leakage [19]. There is an urgent need to explore non-invasive methods for predicting the EGFR mutation status of brain metastases.

Magnetic resonance imaging (MRI) is an effective neuroimaging technique widely used for the detection, diagnosis, and monitoring of brain metastases. In the field of lung cancer, artificial intelligence (AI) has made significant strides, particularly in image analysis. Machine learning (ML) and deep learning (DL) techniques are capable of rapidly capturing tumor heterogeneity and extracting biological information that traditional visual inspection may not reveal [20]. These computational methods can uncover molecular and genetic insights that often exceed the capabilities of conventional imaging. Previous studies have primarily focused on CT imaging, using radiomics and deep learning to predict genetic mutations in lung cancer [21, 22]. Although studies by Mutic and Li et al. have shown that dual-modality MRI provides superior information and enhances predictive accuracy for genetic mutations compared to single-modality CT [23, 24], research on deep learning-based MRI analysis for predicting EGFR mutation status in brain metastases remains limited. In this retrospective study, we extracted conventional radiomics and deep learning features from T1 contrast-enhanced (T1-CE) and T2-weighted (T2W) MRI sequences. We then developed machine learning and deep learning models that integrate these features to differentiate between EGFR mutant (MT) and wild-type (WT) brain metastases.

Methods

Patient selection

Between January 2019 and December 2023, a total of 341 NSCLC patients with brain metastases were enrolled in this study. The inclusion criteria were as follows: (1) pathological diagnosis of NSCLC confirmed by biopsy or bronchoscopy; (2) baseline T2-weighted and T1-weighted contrast-enhanced (T1-CE) MRI sequences obtained prior to anti-tumor treatment; (3) information on the specific EGFR mutation sites and mutation status; (4) tumors having a diameter greater than 5 mm. The exclusion criteria included: (1) incomplete clinical data, including but not limited to the lack of patient clinical factors and complete electronic medical records; (2) a history of malignancies other than NSCLC; (3) presence of central nervous system (CNS) disorders unrelated to cancer, including but not limited to cerebral infarction, traumatic brain injury, and neurodegenerative diseases such as Alzheimer’s; (4) poor-quality or incomplete MRI scans due to artifacts or technical issues. After applying these inclusion and exclusion criteria, 288 patients were included in the final analysis. Clinical and genetic mutation information were obtained from the patients’ electronic medical records. This study received approval from the Research Ethics Committee of Shandong Provincial Hospital affiliated with Shandong First Medical University. All procedures adhered to the guidelines and ethical principles outlined in the 1964 Helsinki Declaration. Informed consent was obtained from all participants.

Detection of EGFR mutation

Histopathological diagnoses of lung cancer were made for all patients based on samples collected via bronchoscopic or percutaneous needle-guided. Genomic DNA was extracted from formalin-fixed, paraffin-embedded (FFPE) tissue sections using the DNeasy Isolation Kit (Qiagen, Valencia, CA, USA). EGFR mutations were identified using the amplification refractory mutation system real-time PCR method with the Human EGFR Gene Mutation Detection Kit (Beijing ACCB Biotech Ltd).

Image acquisition and BM segmentation

Patients who were enrolled in this study were scanned at 3.0-T MRI scanner (Philips Medical Systems, Netherlands). Based on conventional sequences for diagnosing BM, we selected the following sequences and corresponding imaging acquisition parameters for analysis. The acquisition parameters of contrast-enhanced T1-weighted (T1CE) fast spin-echo sequence were as follows [25]: Repeat time = 250ms; Echo time = 2.46ms; matrix = 320 × 320; slice thickness = 5 mm; FOV = 240 × 240 mm. A T2-weighted sequence was obtained using a 3.0T MRI with TR = 4000 ms, TE = 113ms; slice thickness = 5 mm; FOV = 240 × 240 mm and matrix size = 352 × 352 mm. Additionally, the contrast agent utilized was gadolinium diethylenetriaminepentaacetic acid (Gd-DTPA, Bayer, Berlin, Germany). T1CE MR images were acquired 5 min post-injection of the contrast agent. The administered dosage was 0.2 mL/kg with an injection rate of 3 mL/s.

The two imaging sequences - T2-weighted and T1-CE were initially transformed into a uniform geometric space using the Elastix toolbox [26]. This registration step facilitated subsequent analyses and was executed via the open-source 3D Slicer software platform (Version 4.10, http://www.slicer.org). Regions of interest (ROIs) were automatically contoured using AccuContour software (version 3.0, Manteia Medical Technologies Co. Ltd., Xiamen, China) based on deep learning algorithms, then manually modified by the first radiologist with 7 years’ experience and the second radiologist with 10 years of experience. Lesions with a diameter less than 5 mm were not included. Two independent radiologists were blinded to the clinicopathological information. Any disagreement of 5% or more was resolved through consensus [27, 28].

Radiomics features and deep learning features extraction

Quantitative radiomics feature extraction was automatically performed with PyRadiomics packages, which enable feature calculation in 3D Slicer software (Version 4.10, http://www.slicer.org). 93 features were extracted from each ROI and included first-order intensity histogram (IH) and statistical matrix (SM) features, grey-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighboring gray-tone difference matrix (NGTDM), and gray-level dependence matrix (GLDM) features. Furthermore, 744 wavelet features were extracted for IH and SM from 8 wavelet decompositions. All radiomic features were z-score normalized to a mean of 0 and a standard deviation of 1. To evaluate intra-observer reproducibility, two radiologists independently segmented 25 randomly selected lesions from the dataset. The radiomic features extracted from the two independently drawn ROIs were compared using intra-class correlation coefficients (ICCs). Features exhibiting an ICC ≥ 0.8 were retained, and features with an ICC < 0.8 were initially excluded from further analyses due to unsatisfactory intra-observer reproducibility.

The processed tumor images served as input for convolutional neural network (CNN) models. For this investigation, we employed ResNet34 as our pre-trained CNN architecture. This model is a variant of the Residual Network (ResNet) family, designed to mitigate the challenges of gradient vanishing and exploding during deep network training. ResNet34 achieves this through the incorporation of skip connections, allowing the network to learn residual functions with reference to the layer inputs, thereby facilitating the training of deeper neural networks [29]. The network takes a 224 × 224 × 3-pixel natural image as input and, after multiple consecutive convolutional and pooling layers, outputs a 1,000-dimensional vector, which we considered the deep feature extracted from the image. The weights were pretrained using the open-source dataset ImageNet-1k (https://www.image-net.org/download.php). These slices, including the central tumor slice and two adjacent slices on either side, and the ROI regions were extracted. To ensure data quality, all samples underwent manual inspection to verify image integrity and exclude partial volume artifacts. These images were resized to 224 × 224 using bilinear interpolation and replicated across 3 channels for input into ResNet34. After inputting the ROI images into the model, a 1,000-dimensional deep feature was extracted from each ROI region of each sequence.

Features selection

Feature selection technique was implemented to pinpoint and exclude irrelevant radiomic features that could potentially impair the predictive performance for differentiating EGFR MT and WT in the following three steps. First, the Mann–Whitney U test was performed to each feature, and features with p < 0.05 were retained. Then, to mitigate the risks of overfitting and selection bias, the least absolute shrinkage and selection operator (LASSO) [30], a machine learning technique, was employed for optimal feature selection. The objective was to identify the most discriminative radiomic features capable of distinguishing EGFR MT and WT. For the binary logistic regression model, the tuning parameter λ in LASSO was determined through 5-fold cross-validation, minimizing the criteria using the “glmnet” package in R software (Version 3.4, http://www.r-project.org/) [31]. All wavelet filtering, image preprocessing, and subsequent radiomics analysis were conducted using the Multimodal Radiomics Platform, as described in prior publications (accessible online [32]: http://cfu.lab.nycu.edu.tw/MRP_MLinglioma.html). The Akaike Information Criterion (AIC) was used as the evaluation index, and the optimal feature subset was obtained using the backward stepwise search algorithm.

Model construction and validation for EGFR mutation status

After feature selection, we developed a logistic regression model to validate the effectiveness of the selected radiomics features and deep learning features. The signature was computed as a linear combination of the selected features, weighted by their respective LASSO coefficients in the training group. The stacking fusion method was used to construct the DLR signature in a logistic regression model to predict the mutation status of EGFR. Sample size limitations and uneven group distributions can introduce bias, potentially distorting performance estimates. In cases of imbalanced datasets, traditional ROC curve analysis may yield misleadingly high AUC values, obscuring true model efficacy. To address this issue, we employed precision-recall curves (PRC), which plot positive predictive value against true positive rate across all thresholds. The area under the PRC, termed average precision (AP), provides a more robust evaluation metric for imbalanced classification tasks, offering complementary insights to conventional receiver operating characteristics (ROC) analysis. We employed an α-binormal model to generate smooth ROC and precision-recall curves, addressing previous limitations. The model’s discriminative ability was evaluated using these curves. Developed on the training dataset and validated independently, the model yielded total points for each patient, facilitating comprehensive assessment across cohorts. Generally, to evaluate the predictive performance of the signatures, ROC analysis, decision curve analysis and calibration curve analysis were conducted to evaluate the model’s accuracy and clinical applicability.

Statistics analysis

Statistical analysis was performed using R software version 3.4.2 (Auckland, New Zealand). Statistical analyses were conducted using appropriate tests for categorical and continuous variables. Chi-squared or Fisher’s exact tests were applied for clinical categorical factors, such as gender, ECOG PS status and smoking history etc. While Mann-Whitney U tests were used for continuous variables, involving ages and preliminary selecting for distinguishing features of EGFR mutation statuses. The “rms” and “pROC” R packages were utilized to generate comparison, calibration, and precision-recall curves, providing comprehensive visual representations of model performance. The “glmnet” package was used to execute the LASSO algorithm, and the “pROC” package was used to assess predictive values. All statistical analyses were two-sided, and a p-value < 0.05 was considered statistically significant.

Results

Patient characteristics

A total of 288 EGFR positive NSCLC patients with brain metastasis were included in this retrospective study. The training set consisted of 83 patients with EGFR MT and 133 patients with EGFR WT. The validation set included 23 patients with EGFR MT and 49 patients with EGFR WT mutations. The median age at baseline was 61.4 ± 12.3 and 62.8 ± 11.7 in training and validation datasets, respectively (p = 0.761). Demographic and clinical characteristics, including smoking history (p = 0.415), ECOG PS (p = 0.674), gender (p = 0.463), and genetic mutation profile (p = 0.323), were comparable across both cohorts, with no statistically significant differences observed. Table 1 summarizes the clinical characteristics of the patients, while Table 2 presents the clinicopathological factors of patients in the training set, specifically with regard to EGFR mutation status. No significant predictive value for EGFR mutation status was found for gender (p = 0.364) or ECOG PS (p = 0.513). Additionally, smoking history did not reveal any statistically significant differences when identifying EGFR mutant status (p = 0.470).

Table 1 Clinical factors of patients in the training and validation datasets

Table 2 Clinical factors of patients with EGFR MT and EGFR WT status in the training dataset

Radiomics signatures construction and validation

The radiomics analysis included 837 features extracted from segmented pretreatment MRI images of patients. After applying an interclass correlation coefficient cut-off value of 0.8 for the reproducibility test, 541 features were retained for further analysis. LASSO regression identified four radiomics features with non-zero coefficients, which were used to construct the signature in the training set. As illustrated in Fig. 1, these features were weighted according to their respective coefficients. The radiomics signature (RS) demonstrated high discriminative power between EGFR MT and EGFR WT. In the training set, the signature achieved α-binormal and empirical AUCs of 0.931 (95% CI: 0.880–0.940) and 0.926 (95% CI: 0.877–0.933), respectively. This robust performance was further validated in the independent cohort, yielding even higher AUCs (α-binormal: 0.892 (95% CI: 0.870–0.913), empirical: 0.887 (95% CI: 0.865–0.908)). Table 3 summarizes these results, including the average precision (AP) values. Overall, the model developed based on brain metastasis MR images demonstrated robust capability in discriminating EGFR mutation phenotypes in lung cancer (Fig. 2).

Fig. 1

Radiomics features selection through LASSO with a binary regression model. (A) The LASSO coefficient profile plot was produced against the log lambda sequence. (B) Tuning parameter (log lambda) selection in the LASSO via minimum criteria. AUC: Area under the curve

Fig. 2

The performances of the developed radiomics signature. (A) Receiver operating characteristics (ROC) curves. (B) Precision-recall curve (PRC)

Table 3 The predictive performances of RS for EGFR mutant status

Deep learning signature construction and validation

LASSO regression analysis reduced the deep learning-derived features from 1520 to 16 potential predictors, retaining only those with non-zero coefficients. We developed a model by deep learning features obtained from the optimized imaging model. The deep learning signature (DLS) model demonstrated superior predictive efficacy for differentiating EGFR mutation phenotypes compared to the RS alone. In the training cohort, it achieved the highest AUC with α-binormal and empirical AUCs of 0.943 (95% CI: 0.921–0.965) and 0.938 (95% CI: 0.914–0.962), respectively, highlighting its robust discriminative capability. External validation of the model in an independent cohort confirmed its exceptional discriminative performance, yielding an α-binormal AUC of 0.936 and an empirical AUC of 0.921 (95% CI: 0.901–0.941) (Fig. 3) (Table 4).

Fig. 3

The performance of the developed deep learning signatures. (A) Receiver operating characteristics (ROC) curves. (B) Precision-recall curve (PRC).The performance of the developed deep learning signatures. (A) Receiver operating characteristics (ROC) curves. (B) Precision-recall curve (PRC)

Fig. 4

Calibration curve of the deep learning signatures shows as a red line

Table 4 The predictive performances of DLS for EGFR mutant status

The calibration curve indicated a strong concordance between the predicted EGFR mutation status probability and the actual observed outcomes using restricted cubic splines, with an AUC of 0.870 (95% CI: 0.781–0.959) (Fig. 4). The degree of alignment of the calibration curve with the diagonal line was directly related to the DLS’s predictive accuracy. This close alignment signifies that the DLS’s predictions were highly reliable. To further evaluate clinical utility, decision curves for both the RS and the DLS were plotted. These decision curves illustrated the net benefit of each signature across a range of threshold probabilities. Figure 5 confirmed that DLS exhibited superior performance in distinguishing between EGFR WT and MT cases, providing a greater net benefit compared to the RS.

Fig. 5

The decision of the radiomics signature, deep learning signatures and two extreme curves were plotted based on the validation dataset. The figure illustrated that the utilize of deep learning signatures to predict EGFR MT has a greater benefit that the radiomics signature alone

Discussion

In the era of precision medicine, the rapid and accurate prediction of EGFR mutations is crucial for implementing targeted therapies, as it enables clinicians to tailor treatment plans based on the patient’s specific mutation profile, thereby significantly improving therapeutic outcomes. In this study, we used LASSO and ResNet34 to extract and select features, demonstrating the potential of MRI images in predicting the EGFR mutation status in lung adenocarcinoma patients with brain metastases. More importantly, we developed a deep learning-based predictive model designed to assist clinicians in forecasting EGFR mutation status, providing an effective tool to enhance treatment precision in this complex clinical context.

Brain metastasis is a common and serious complication in lung cancer, accounting for 40–50% of all brain metastasis cases [33]. EGFR mutations are associated with a higher risk of brain metastasis in NSCLC patients [34, 35], and EGFR TKIs have shown significant efficacy in treating brain metastases, making them a viable first-line treatment option [36]. EGFR-mutated patients should be prioritized for targeted therapy, while wild-type patients need to assess PD-L1 expression and select appropriate local treatments [37]. The effectiveness of treating central nervous system metastases largely depends on the ability of drugs to penetrate BM. Drug distribution is influenced by factors such as BBB disruption, tumor growth characteristics, and the activity of efflux transporters [38, 39]. EGFR TKIs have demonstrated significant efficacy in treating brain metastases in EGFR-mutated patients, with first- and second-generation EGFR inhibitors changing the treatment landscape for NSCLC [40]. Third-generation TKIs, such as osimertinib, effectively overcome resistance caused by the EGFR T790M mutation and can penetrate both the blood-brain barrier and blood-tumor barrier [41, 42]. These findings underscore the importance of studying EGFR mutation subtypes to guide personalized treatment strategies.

In clinical practice, the evaluation of EGFR mutation status typically relies on two methods: plasma ctDNA detection and tissue biopsy, each of which has its inherent limitations. Plasma ctDNA detection, while non-invasive and convenient, is often limited by the low concentration of ctDNA in plasma, and the high costs associated with the procedure are significant barriers to its widespread use [18, 43, 44]. In contrast, tissue biopsy provides direct genomic information but is an invasive procedure that carries risks, such as vascular or neural damage, and the potential for tumor cell dissemination. Furthermore, biopsy samples are typically obtained from a specific region of the lesion, which may not fully capture the heterogeneity of the entire tumor [45]. In situations where pathological sampling is not feasible, MRI analysis of brain metastases offers an effective alternative for predicting the type of EGFR mutation, thereby improving diagnostic accuracy and providing valuable insights for the development of personalized treatment strategies [46].

Radiomics-based binary classification models have shown significant potential in predicting EGFR mutation status by automating the extraction of features from medical imaging data. Liu et al. demonstrated that integrating CT radiomics features with clinical data enhanced the predictive power for EGFR mutation in a cohort of 298 lung adenocarcinoma patients, improving the AUC from 0.690 to 0.778 [47]. Similarly, Zhao et al. achieved an increase in AUC from 0.645 to 0.758 by combining CT radiomics features with deep learning techniques [48]. However, CT-based deep learning models face several inherent limitations. A major obstacle is the accurate delineation and segmentation of lung lesions in CT images, particularly when complex anatomical structures such as major blood vessels, lymph nodes, and adjacent organs are involved. These complexities can impair the accuracy of segmentation and adversely affect model performance [49,50,51]. In contrast, MRI offers distinct advantages over CT in this application, including superior soft tissue contrast and the ability to capture multiple sequence parameters [23, 52]. These advantages of MRI may enhance the accuracy and robustness of radiomics-based prediction models for EGFR mutation status, providing a more reliable alternative for clinical use.

Our research focuses on the radiomic analysis of MRI images of brain metastases, differentiating from previous studies that primarily concentrated on CT radiomics of primary lung tumors. Previous studies have explored the application of deep learning models based on MRI for predicting EGFR mutation status. Wang et al. evaluated a radiomic signature for predicting two EGFR subtypes but reported unsatisfactory results, with low AUCs [46]. Li et al. demonstrated that a deep learning approach using multi-sequence MRI could predict EGFR mutation status in NSCLC patients with brain metastases [53]. However, their study lacked external validation and only provided a basic distinction between EGFR mutations and wild-type status. Our encouraging results can be attributed to the following two factors. First, we selected lesions with a diameter greater than 5 mm to avoid the potential influence of noise, blurring, or low contrast, which could hinder the accurate identification of boundaries by segmentation algorithms. This is particularly important in low-resolution or unclear medical images, where small tumors may be misclassified as background or other structures. Furthermore, the chosen segmentation approach combines automatic segmentation with manual adjustments by experienced physicians. Traditional threshold-based segmentation algorithms may perform well for large tumors but often fail to detect smaller ones. In contrast, deep learning-based algorithms, by learning more complex features, can adapt to segmentation tasks involving tumors of varying sizes. This approach not only reduces errors inherent in purely manual segmentation but also enhances labelling consistency and repeatability. Third, the application of advanced ROIs segmentation algorithms. Enhancing the accuracy of these segmentation algorithms is currently crucial for improving the performance of medical image analysis and predictive models. Research on tumor segmentation has been conducted across various tumor types, with a particular emphasis on methods designed to enhance segmentation accuracy in complex anatomical areas and small regions. Mudassar et al. combined an attention-enhanced U-Net model with cGAN, achieving two key breakthroughs: the block discriminator in cGAN can generate highly realistic synthetic images for data augmentation, addressing the limitation of annotated data availability; and the attention mechanism embedded in U-Net can accurately capture complex structural features, such as tumor boundaries [54]. This approach performed excellently on the brain MRI dataset, achieving a Dice coefficient of 98.61%. Nomura et al. developed an adaptive input strategy based on fixed-size VOI (Volume of Interest), enabling the model to handle lesions of varying sizes [55]. When applied to brain metastasis tumor segmentation, the Dice coefficient was 0.727 ± 0.115. Therefore, future research may focus on developing adaptive algorithms that automatically adjust to tumor size variations, combined with multi-scale and multi-task learning approaches, to further improve segmentation accuracy and prediction performance.

With the widespread adoption of technologies such as CNNs and Deep Neural Networks (DNNs), numerous studies have shown that deep learning surpasses traditional radiomics signatures in distinguishing EGFR mutations. Song et al. reported that deep learning-based methods outperformed radiomics in classifying EGFR mutation subtypes in lung adenocarcinoma patients [56]. Similarly, research by Nguyen et al. confirmed these findings, demonstrating that deep learning algorithms performed better than conventional machine learning (cML), achieving higher AUC (0.822 vs. 0.775) and sensitivity (80.1% vs. 71.1%) [57]. Our findings are consistent with these studies. By integrating deep neural networks, MRI quantitative features, and EGFR genetic data, our DeepSurv model accurately predicts the EGFR mutation status of patients based on imaging features, demonstrating its potential in survival analysis. Additionally, we have addressed the challenge of small labeled medical image datasets through multi-sequence image fusion. Moreover, deep learning models that incorporate genomic mutation data show greater prognostic accuracy compared to traditional clinical staging, highlighting the crucial role of sequence mutation analysis in advancing molecular typing for lung cancer.

Several factors can influence the accuracy of predictive models. Ahn et al.‘s study suggests that diagnosing EGFR mutation status in large brain metastases (diameter > 10 mm) may be less effective than in smaller lesions, possibly due to the presence of necrotic centers, which can impair the accuracy of machine learning classifications [58]. Additionally, large brain metastases often correspond to smaller datasets, which may lead to overfitting issues. In contrast, Zhou et al. pointed out that accurately genotyping EGFR mutations in patients with small brain metastases is challenging, primarily due to the difficulty in extracting imaging features from these very small lesions [59]. Previous studies have also emphasized the importance of the metastasis/brain parenchyma (M/BP) interface in predicting EGFR mutations. Fan et al. demonstrated that the M/BP interface and the tumor’s active regions could provide complementary information regarding EGFR mutation status and response to EGFR-TKI therapy [60]. The complexity of peritumoral edema, especially in irregularly shaped lesions, may also affect image segmentation accuracy, making precise delineation more difficult. Moreover, brain metastases exhibit diverse imaging characteristics across different brain regions (e.g., cerebrum, cerebellum, meninges), which complicates the development of unified predictive models and highlights the need for region-specific analytical approaches. Future research should focus on addressing these challenges to enhance the accuracy of prediction models and their applicability in clinical decision-making.

In this study, the model demonstrated exceptional predictive accuracy for EGFR mutations, with α-binormal-based and empirical AUCs of 0.931 and 0.926, respectively. These results were also confirmed in the validation dataset. Notably, our approach utilized multi-center data while maintaining high predictive accuracy, which underscores the model’s strong generalization capabilities. Both our study results and the prediction models constructed using deeper features show that their predictive performance is influenced not only by the segmentation algorithm but also by other factors. The imaging protocol during MRI acquisition can affect the extraction and values of radiomics or deep learning features, such as voltage, current, and image slice thickness. We applied the LASSO algorithm to eliminate features with high correlation, redundancy, or instability, thereby selecting the optimal features to enhance the predictive performance of deep learning and machine learning models. The value of the feature parameters identified in this study in other research remains to be validated, but these factors have the potential to improve predictive performance. Conversely, the conclusions and parameters from previous studies have only been validated within specific datasets, and their applicability in our study requires further validation.

This study has several limitations that warrant further consideration. Firstly, the retrospective design of the study inherently introduces the potential for selection bias. Secondly, tumor segmentation on the MRI images was performed using a semi-automatic approach by a multidisciplinary team of experienced radiologists and oncologists. Future development of fully automated segmentation methods could not only reduce the time and cost associated with treatment planning but also improve the reproducibility of radiomic features. Thirdly, due to the relatively small sample size, this study did not predict different EGFR mutation subtypes. Future research could focus on predicting the prognosis of different EGFR mutation sites, such as 19 del, L858R, T790M, and related TKI treatments.

Conclusions

In conclusion, our study highlights the effectiveness of a multisequence MRI-based model in predicting EGFR mutation status in metastatic NSCLC patients. By integrating complex radiological features with advanced machine learning techniques, this novel approach provides a non-invasive and accurate method for determining EGFR mutation status. The model shows significant potential as a valuable clinical decision support tool, especially in cases where biopsy is difficult or contraindicated.