Classifying brain metastases originating from different pathological subtypes of lung cancer via a multimodal magnetic resonance imaging-based deep learning approach

Introduction

Brain metastases (BMs) represent one of the most common neurological complications associated with cancer and account for over half of all brain tumors (1). Lung cancer has the highest risk of BMs among all solid carcinomas (2). The management of lung cancer has transitioned into the era of personalized and precision medicine, with treatment strategies varying according to the specific pathological subtype of lung cancer. For instance, the first-line therapeutic approach for small-cell lung cancer (SCLC) is typically combination chemotherapy (3), while for advanced non-SCLC (NSCLC), the recommended first-line treatment is molecular targeted therapy as guided by driver oncogenes (4).

In clinical practice, for certain patients who undergo magnetic resonance imaging (MRI) examination due to brain symptoms, BMs may be discovered before the primary tumor or concurrently. Therefore, the prompt and efficient identification of different pathological subtypes of lung cancer through BMs is critical to initiating therapy as early as possible and improving prognosis. Although histopathological examination serves as the gold standard for diagnosing BMs, brain biopsy is limited due to its invasive nature, elevated risk, sampling bias, and patient reluctance, and thus obtaining histological confirmation of BMs remains challenging. Moreover, the positive rate of cerebrospinal fluid examination for lung cancer BMs is relatively low. Consequently, there is an urgent need to develop a method that can provide prompt, efficient, and noninvasive classification of BMs originating from different pathological subtypes of lung cancer poses in clinic.

MRI serves as the primary modality for diagnosing lung cancer BMs (5,6); however, it is difficult to distinguish BMs originating from different pathological subtypes of lung cancer with conventional MRI. Given recent advancements in artificial intelligence (AI), particularly the advance of deep learning (DL), the integration of DL with MRI presents a novel approach to identifying BMs originating from different pathological subtypes of lung cancer. DL is a subset of machine learning (ML) that utilizes computational units of multiple layers to facilitate automatic hierarchical feature learning (7,8). DL has been demonstrated to outperform the classical ML approach in various visual computer tasks and medical image analyses (9). DL has been applied in the field of neuro-oncology (8,10-12), while for classifying BMs originating from different pathological subtypes of lung cancer, several studies have applied classical ML (13,14) and a few have used DL (15). However, there is no consensus as to whether DL provides a superior classification performance, and confirming the pathological subtype of lung cancer through BMs remains challenging.

The aim of this study was thus to assess the performance of an MRI-based DL method in classifying BMs according to the pathological subtype of lung cancer. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1285/rc).

Methods

Participants

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Zibo Central Hospital (No. 2025214), and the requirement for informed consent was waived because of the retrospective nature of the study.

The image data were consecutively and retrospectively retrieved from our institutional electronic medical records. A total of 406 patients with confirmed lung cancer BMs who had undergone MRI examinations at Zibo Central Hospital were retrospectively recruited for this study between August 2019 and September 2021. A review panel consisting of two radiologists specialized in lung radiology (Z.H., with 15 years of experience, and H.Z., with 20 years of experience) and a lung oncologist (T.F., with 10 years of experience) was convened to confirm the diagnoses of BMs originating from lung cancer. The panel reviewed the clinical, pathology, and imaging data available immediately prior to and during follow-up to determine whether the brain lesions observed on MRI were indicative of BMs originating from lung cancer. After careful review, 406 patients were clinically diagnosed as having BMs. The inclusion criteria were as follows: (I) pathologically confirmed lung cancer; (II) a review panel-confirmed clinical diagnosis of BMs originating from lung cancer; and (III) MRI performed with T2 fluid attenuated inversion recovery (FLAIR), diffusion-weighted imaging (DWI), and T1-weighted contrast enhancement (T1CE) imaging. Meanwhile, the exclusion criteria were as follows: (I) patients with a brain biopsy or treatment history before the MRI examination (n=57); (II) tumors with a maximum axial diameter less than 1 cm (n=74); and (III) no availability of digital imaging data (n=13). Figure 1 shows the flowchart of patient selection. Ultimately, 262 patients comprising 357 BMs with a diameter greater than 1 cm were included in the study. Regarding the primary cancers, there were 154 cases of lung adenocarcinoma (LUAD), 48 cases of SCLC, and 60 cases of other pathological subtypes of lung cancer (OPTLC) (Figure 2). The detailed composition of the OPTLC category is provided in Table S1.

Figure 1 The flowchart of patient selection. ADC, apparent diffusion coefficient; DWI, diffusion-weighted imaging; FLAIR, fluid-attenuated inversion recovery; MRI, magnetic resonance imaging; T1CE, T1-weighted contrast enhancement.

Figure 2 Pie charts for the distribution of patients with BMs originating from different pathological subtypes of lung cancer. BM, brain metastasis; LUAD, lung adenocarcinoma; OPTLC, other pathological subtypes of lung cancer; SCLC, small cell lung cancer.

MRI acquisition

MRI examinations were performed on one of three MRI scanners: a 3.0-T Signa HDXT device (GE HealthCare, Chicago, IL, USA), a 1.5-T Excelart Vantage MRT-1503 device (Canon Medical Systems Corporation, Otawara, Japan), and a 3.0-T MAGNETOM Skyra device (Siemens Healthineers, Erlangen, Germany) with an eight-channel sensitivity-encoding head coil. The Signa, Vantage, and Skyra devices were used in 116, 124, and 22 patients with BMs originating from lung cancer, respectively. The imaging protocols included T2 FLAIR, DWI, and T1CE sequences. Apparent diffusion coefficient (ADC) maps were automatically generated through postprocessing software. The detailed imaging parameters are summarized in Table S2.

Image segmentation

MRI data were transferred to a personal computer, where ITK-SNAP (http://www.itksnap.org/) was used to perform image segmentation in a semiautomatic manner. Before image segmentation, unified training for segmentation was provided for the review radiologists, and a stable segmentation standard was established. Image segmentation was processed by two radiologists, a segment radiologist (Y.L., with 10 years of experience) and a review radiologist (B.A., with 15 years of experience). Disagreements were resolved by discussion until a consensus was reached. Radiologists were blinded to the patients’ diagnosis. The largest slice of the tumor was selected on the T2 FLAIR, DWI, ADC, and T1CE sequences, respectively, and the region of interest was drawn along the tumor edge. Ultimately, 356 lesions were segmented on the T2 FLAIR sequence, 357 lesions on the DWI sequence, 282 lesions on the ADC maps, and 349 lesions on the T1CE sequence. In total, 1,344 samples were segmented.

Dataset partitioning

The data from the four sequences were randomly divided into training, validation and testing sets in a ratio of 7:1:2. The training set is used for model parameter fitting, the validation set for hyperparameter tuning and early stopping, and the testing set for final evaluation. The distribution of the dataset was shown in Table 1.

DL model architecture

We employed ResNet-18 (Table S3) as the basic classification framework and performed classification detection on the T2 FLAIR, DWI, ADC, and T1CE sequences. ResNet-18 is a convolutional neural network (CNN) belonging to the ResNet-xx family of models (16), known for its efficiency and effectiveness in image recognition tasks. Compared to traditional CNN, ResNet-18 reduces training error while adding more layers (17). In all BM classification experiments, the initial learning rate was set to 1e−3. We adopted an adaptive learning rate adjustment strategy, in which the learning rate for the next cycle is adjusted to be 0.5 times the current learning rate based on the validation set detection accuracy.

Statistical analysis

Demographic data

The normality of the distribution was tested via the Shapiro-Wilk test. The Mann-Whitney test and the Chi-squared test were used to determine the statistical significance of differences between age groups and gender groups, respectively. P<0.05 was considered statistically significant.

Performance evaluation

The discrimination performances of T2 FLAIR, DWI, ADC, and T1CE sequences for classification of BMs originating from different pathological subtypes of lung cancer were assessed by using receiver operating characteristic (ROC) curve analysis. Additionally, the precision, recall, F1-score, and accuracy of the T2 FLAIR, DWI, ADC, and T1CE sequences were calculated. All final performance metrics reported were based on the testing set. P<0.05 was considered statistically significant.

The workflow of this study was shown in Figure 3.

Figure 3 The workflow of the image segmentation, DL, and model employed in the study. Red regions indicate the tumor manually delineated for image segmentation. ADC, apparent diffusion coefficient; DL, deep learning; DWI, diffusion-weighted imaging; FLAIR, fluid-attenuated inversion recovery; ROC, receiver operating characteristic; T1CE, T1-weighted contrast enhancement.

Results

Patient demographics

A total of 1,344 BM samples from 262 patients were included in the analysis. There was no statistically significant difference in the age distribution among patients with LUAD, SCLC, or OPTLC BMs. However, there was a higher proportion of female patients with LUAD BMs than with SCLC or OPTLC BMs. The demographic characteristics are summarized in Table 2.

Classification performance of the DL approach

The comprehensive results for the classifications of LUAD, SCLC, and OPTLC BMs are shown in Table 3. Respectively, the precision, recall, F1-score, and accuracy were 0.618, 0.567, 0.581, and 0.653 with the T2 FLAIR sequence; 0.617, 0.575, 0.582, and 0.634 with the DWI sequence; 0.649, 0.719, 0.667, and 0.684 with ADC mapping; and 0.689, 0.735, 0.708, and 0.757 with the T1CE sequence.

The single category results of the DL approach based on the T2 FLAIR, DWI, ADC, and T1CE sequences in classifying LUAD, SCLC, and OPTLC BMs in terms of precision, recall, and F1-score are shown in Table 4.

In the ROC curve analysis, the area under the curve (AUC) for the DL approach in classifying BMs originating from different pathological subtypes of lung cancer as well as microaverage and macroaverage were, respectively, 0.71, 0.66, 0.66, 0.74, and 0.69 with the T2 FLAIR sequence; 0.67, 0.65, 0.65, 0.71, and 0.67 with the DWI sequence; 0.75, 0.92, 0.88, 0.83, and 0.86 with ADC mapping; and 0.74, 0.88, 0.82, 0.83, and 0.83 with the T1CE sequence (Figures 4,5).

Figure 4 The ROC curves for classifying different pathological subtypes of lung cancer BMs. ADC, apparent diffusion coefficient; BM, brain metastasis; DWI, diffusion-weighted imaging; FLAIR, fluid-attenuated inversion recovery; ROC, receiver operating characteristic; T1CE, T1-weighted contrast enhancement.

Figure 5 Heat maps for the AUC values of the ROC curves of the DL approach based on multimodal MRI. ADC, apparent diffusion coefficient; AUC, area under the curve; BM, brain metastasis; DL, deep learning; DWI, diffusion-weighted imaging; FLAIR, fluid-attenuated inversion recovery; LUAD, lung adenocarcinoma; MRI, magnetic resonance imaging; OPTLC, other pathological subtypes of lung cancer; ROC, receiver operating characteristic; SCLC, small-cell lung cancer; T1CE, T1-weighted contrast enhancement.

Discussion

In this study, we used a DL approach to classify BMs originating from different pathological subtypes of lung cancer and assessed it classification performance. Accurate classification of BMs originating from different pathological subtypes of lung cancer is clinically significant for treating and managing lung cancer BMs in patients, as it can inform the selection of treatment strategy.

Our study demonstrated that a DL approach based on multimodal MRI has the potential serve as a noninvasive tool for effectively identifying BMs originating from different pathological subtypes of lung cancer. Another key finding from our study was that among the sequences examined, the ADC map and T1CE sequences exhibited the best classification performance. The ADC value can quantify the extent of restricted diffusion of water molecules in tumors (18), while the T1CE sequence can reflect the blood supply of tumors, with enhancement indicating the destruction of the blood-brain barrier (19). Researches (20,21) indicate that ADC mapping is able to discriminate well between BMs originating from different pathological subtypes of lung cancer. A previous study that used a T1CE single sequence based on DL to differentiate the pathological subtypes of SCLC and NSCLC causing BMs reported a mean overall accuracy of 0.7515 (22). In our study, the DL approach based on ADC maps and TICE sequences had the highest AUC and ROC curve for classifying SCLC, with values of 0.92 and 0.88, respectively. This suggests that our DL approach would be particularly valuable in recognizing BMs originating from SCLC, especially as SCLC is more aggressive and tends to disseminate earlier than does NSCLC. Moreover, SCLC is typically managed with combined chemotherapy rather than surgical treatment. The development of a noninvasive approach for accurately diagnosing BMs preoperatively could greatly facilitate the implementation of individualized treatment strategies.

Our work was novel in that we identified BMs originating from different pathological subtypes of lung cancer with a DL based on multimodal MRI, whereas prior studies have generally used a classical ML approach. For instance, one study used support vector machine (SVM) and binary logistic regression models to differentiate BMs originating from primary adenocarcinoma and squamous cell carcinoma, with the AUCs ranging from 0.765 to 0.828 (23). Furthermore, a classical ML approach based on T1CE and the extreme gradient boosting (XGBoost) algorithm was developed to distinguish BMs originating from different pathological subtypes in patients with NSCLC, yielding an AUC of 0.8 (14). In our study, the DL approach based on the ADC map and T1CE sequences achieved AUCs of 0.75 to 0.92 and 0.74 to 0.88 for the ROC curves in the test set, respectively. The microaverage and macroaverage of ROC curves were used to evaluate the classification value of the DL approach in the test set, and the microaverage AUCs based on the ADC map and T1CE sequences were 0.83 and 0.83, respectively, while the macroaverage values were 0.83 and 0.86 in set, respectively. The semetrics serve as the comprehensive evaluation indicators for assessing the multicategory classification structures. In this study, we opted to report the microaverage instead of the macroaverage, due to its suitability in the presence of class imbalance within multicategory classification structures (24). The microaverage we obtained indicated favorable performance as compared to that reported in previous classical ML studies (14,23). We are currently undergoing a transition period from classical ML to DL (25). Our study provides robust evidence supporting the feasibility of using DL to classify BMs originating from different pathological subtypes of lung cancer based on multimodal MRI. The results reported in this paper may be used as preliminary findings to inform future prospective studies on the use of DL approaches in classifying BMs originating from different pathological subtypes of lung cancer.

There are several limitations to our study which should be addressed. First, despite being one of the larger retrospective studies on classifying BMs originating from different pathological subtypes of lung cancer using DL approach, the sample size was somewhat small. Secondly, the current study was conducted solely within a single institution that may impact the generalizability of our findings, and external validation in a larger multicenter study across institutions is required in the future. Thirdly, while DL enables exploring abstract imaging-clinical relationships, its black-box nature (26) and limited explanatory capabilities due to feature self-learning (25) constrain applicability. Future work should develop advanced interpretable frameworks to improve predictive performance. Lastly, our study utilized 2D magnetic resonance (MR) images and involved semi-automatic segmentation of BMs, which takes much time and manpower, future work will explore 3D approaches. DL-based automated tumor segmentation algorithms with robust performance have been proposed, offering reduced subjectivity and significant gains in efficiency, facilitating larger-scale studies in neuro-oncology.

Conclusions

In conclusion, our DL approach based on multimodal MRI was able to classify BMs originating from different pathological subtypes of lung cancer, particularly utilizing ADC and T1CE sequences. These findings provide a basis for further development of non-invasive diagnostic tools.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1285/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1285/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1285/prf

Funding: This study was supported by the Zibo Key Research and Development Plan (Nos. 2020ZC010169 and 2020ZC010263) and the Shandong Medical and Health Technology Project (No. 202309010704).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1285/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Zibo Central Hospital (No. 2025214), and the requirement for informed consent was waived because of the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Nelson TA, Wang N. Targeting lung cancer brain metastases: a narrative review of emerging insights for anaplastic lymphoma kinase (ALK)-positive disease. Transl Lung Cancer Res 2023;12:379-92. [Crossref] [PubMed]
2. Wei Y, Xu Y, Wang M. Immune checkpoint inhibitors for the treatment of non-small cell lung cancer brain metastases. Chin Med J (Engl) 2023;136:1523-31. [Crossref] [PubMed]
3. Chung HC, Piha-Paul SA, Lopez-Martin J, et al. Pembrolizumab After Two or More Lines of Previous Therapy in Patients With Recurrent or Metastatic SCLC: Results From the KEYNOTE-028 and KEYNOTE-158 Studies. J Thorac Oncol 2020;15:618-27.
4. Qin K, Hou H, Liang Y, et al. Prognostic value of TP53 concurrent mutations for EGFR- TKIs and ALK-TKIs based targeted therapy in advanced non-small cell lung cancer: a meta-analysis. BMC Cancer 2020;20:328. [Crossref] [PubMed]
5. Zhang M, Young GS, Chen H, et al. Deep-Learning Detection of Cancer Metastases to the Brain on MRI. J Magn Reson Imaging 2020;52:1227-36. [Crossref] [PubMed]
6. Kaufmann TJ, Smits M, Boxerman J, et al. Consensus recommendations for a standardized brain tumor imaging protocol for clinical trials in brain metastases. Neuro Oncol 2020;22:757-72. [Crossref] [PubMed]
7. Mathema VB, Sen P, Lamichhane S, et al. Deep learning facilitates multi-data type analysis and predictive biomarker discovery in cancer precision medicine. Comput Struct Biotechnol J 2023;21:1372-82. [Crossref] [PubMed]
8. Yoo Y, Ceccaldi P, Liu S, et al. Evaluating deep learning methods in detecting and segmenting different sizes of brain metastases on 3D post-contrast T1-weighted images. J Med Imaging (Bellingham) 2021;8:037001. [Crossref] [PubMed]
9. Wang L, Zhou X, Nie X, et al. A Multi-Scale Densely Connected Convolutional Neural Network for Automated Thyroid Nodule Classification. Front Neurosci 2022;16:878718. [Crossref] [PubMed]
10. Shin I, Kim H, Ahn SS, et al. Development and Validation of a Deep Learning-Based Model to Distinguish Glioblastoma from Solitary Brain Metastasis Using Conventional MR Images. AJNR Am J Neuroradiol 2021;42:838-44. [Crossref] [PubMed]
11. Jayachandran Preetha C, Meredig H, Brugnara G, et al. Deep-learning-based synthesis of post-contrast T1-weighted MRI for tumour response assessment in neuro-oncology: a multicentre, retrospective cohort study. Lancet Digit Health 2021;3:e784-94. [Crossref] [PubMed]
12. Bae S, An C, Ahn SS, et al. Robust performance of deep learning for distinguishing glioblastoma from single brain metastasis using radiomic features: model development and validation. Sci Rep 2020;10:12110. [Crossref] [PubMed]
13. Sun L, Luan S, Yu L, et al. Radiomic signatures of brain metastases on MRI: utility in predicting pathological subtypes of lung cancer. Transl Cancer Res 2024;13:6825-36. [Crossref] [PubMed]
14. Deng F, Liu Z, Fang W, et al. MRI radiomics for brain metastasis sub-pathology classification from non-small cell lung cancer: a machine learning, multicenter study. Phys Eng Sci Med 2023;46:1309-20. [Crossref] [PubMed]
15. Grossman R, Haim O, Abramov S, et al. Differentiating Small-Cell Lung Cancer From Non-Small-Cell Lung Cancer Brain Metastases Based on MRI Using Efficientnet and Transfer Learning Approach. Technol Cancer Res Treat 2021;20:15330338211004919. [Crossref] [PubMed]
16. Abunadi I, Senan EM. Multi-Method Diagnosis of Blood Microscopic Sample for Early Detection of Acute Lymphoblastic Leukemia Based on Deep Learning and Hybrid Techniques. Sensors (Basel) 2022;22:1629. [Crossref] [PubMed]
17. Zhang Y, Peng L, Ma G, et al. Dynamic Gesture Recognition Model Based on Millimeter-Wave Radar With ResNet-18 and LSTM. Front Neurorobot 2022;16:903197. [Crossref] [PubMed]
18. Du N, Shu W, Li K, et al. An initial study on the predictive value using multiple MRI characteristics for Ki-67 labeling index in glioma. J Transl Med 2023;21:119. [Crossref] [PubMed]
19. Zhu D, Shao Y, Yang Z, et al. Magnetic resonance imaging characteristics of brain metastases in small cell lung cancer. Cancer Med 2023;12:15199-206. [Crossref] [PubMed]
20. Müller SJ, Khadhraoui E, Neef NE, et al. Differentiation of brain metastases from small and non-small lung cancers using apparent diffusion coefficient (ADC) maps. BMC Med Imaging 2021;21:70. [Crossref] [PubMed]
21. Bozdağ M, Er A, Çinkooğlu A. Histogram Analysis of ADC Maps for Differentiating Brain Metastases From Different Histological Types of Lung Cancers. Can Assoc Radiol J 2021;72:271-8. [Crossref] [PubMed]
22. Sui L, Chang S, Xue L, et al. Deep Learning Based on Enhanced MRI T1 Imaging to Differentiate Small-cell and Non-small-cell Primary Lung Cancers in Patients with Brain Metastases. Curr Med Imaging 2023;19:1541-8. [Crossref] [PubMed]
23. Zhang J, Jin J, Ai Y, et al. Differentiating the pathological subtypes of primary lung cancer for patients with brain metastases based on radiomics features from brain CT images. Eur Radiol 2021;31:1022-8. [Crossref] [PubMed]
24. Solano-Rojas B, Villalón-Fonseca R. A Low-Cost Three-Dimensional DenseNet Neural Network for Alzheimer's Disease Early Discovery. Sensors (Basel) 2021;21:1302. [Crossref] [PubMed]
25. Cho SJ, Sunwoo L, Baik SH, et al. Brain metastasis detection using machine learning: a systematic review and meta-analysis. Neuro Oncol 2021;23:214-25. [Crossref] [PubMed]
26. Fufurin I, Berezhanskiy P, Golyak I, et al. Deep Learning for Type 1 Diabetes Mellitus Diagnosis Using Infrared Quantum Cascade Laser Spectroscopy. Materials (Basel) 2022;15:2984. [Crossref] [PubMed]

(English Language Editor: J. Gray)