Predictive Modeling of Brain Metastasis in Advanced Lung Adenocarcinoma: A Hybrid Approach Combining Traditional Radiomics and Deep Learning from Thoracic CT Images

doi:10.21203/rs.3.rs-4992307/v1

Purpose: Create a deep learning-based radiomics framework to anticipate prediction models for advanced lung adenocarcinoma with brain metastases. This aims to inform individualized treatment and prognosis, enhancing clinical decisions and patient outcomes.

Methods: Analyzed 404 patients' CT scans from two hospitals. Extracted handcrafted and deep learning features. Developed three models (Rad, DTL, Combined) to predict brain metastasis risk. The Combined model with clinical features formed the DLRN model. Evaluated using DCA and Calibration Curve.

Results: The Combined model outperformed others, with AUCs of 0.978 (training) and 0.833 (validation). When combined with clinical data, DLRN achieved AUCs of 0.979 (training) and 0.837 (validation), with high accuracy, sensitivity, and specificity. DCA showed DLRN's clinical benefit.

Conclusions: Developed and validated DLRN model for precise prediction of brain metastases.

Brain Metastasis Prediction

Advanced Lung Adenocarcinoma

Predictive Modeling

Radiomics

Deep Learning

Lung adenocarcinoma, a subtype of non-small cell lung cancer (NSCLC), is one of the leading causes of cancer-related deaths worldwide¹. Advanced stages of this malignancy, particularly those accompanied by brain metastases, significantly impact patient prognosis and treatment strategies². Early and accurate prediction of brain metastases in advanced lung adenocarcinoma patients is crucial for optimizing treatment plans, enhancing patient outcomes, and improving quality of life³.

In recent years, advances in medical imaging technologies, particularly CT, have enabled detailed visualization of anatomical and functional changes within the body⁴. Traditional radiomics, an emerging field at the intersection of medical imaging and data science, involves the extraction of high-dimensional quantitative features from medical images to decode tumor phenotypes and predict clinical outcomes⁵. However, the complexity and heterogeneity of tumors pose challenges for traditional radiomics approaches alone in achieving highly accurate predictions⁶.

Concurrently, deep learning, a subset of artificial intelligence, has revolutionized the field of medical image analysis by its ability to automatically learn hierarchical representations of data and extract abstract features from raw images⁷. By integrating deep learning with traditional radiomics, a hybrid approach may harness the power of both techniques to enhance the prediction accuracy of brain metastases in advanced lung adenocarcinoma.

To explore this potential, we designed a study that aims to develop and validate a predictive model for brain metastases in advanced lung adenocarcinoma patients, leveraging a hybrid framework that combines traditional radiomics and deep learning from thoracic CT images. This study utilizes a unique dataset comprising patients from two reputable medical institutions: Hebei University Affiliated Hospital serves as the training set, providing a rich source of annotated CT images and clinical information for model development, while Baoding No.1 Central Hospital functions as the validation set, enabling the assessment of the model's generalizability and robustness in an independent cohort. By leveraging this two-center approach, we hope to contribute to the field of precision oncology by developing a robust and clinically applicable predictive model for brain metastases in advanced lung adenocarcinoma.

Baseline characteristics

A total of 326 patients from the Affiliated Hospital of Hebei University were used as the training set, while 78 patients from the First Central Hospital of Baoding served as the validation set. In the validation set, the distributions of age, gender, smoking status, primary site, EGFR status, T stage, N stage, and bone metastasis were well-balanced, with P-values greater than 0.05, indicating no significant differences across these categories. However, within the train set, imbalances were observed in the distributions of age, EGFR status, and T stage, with P-values less than 0.05, suggesting statistical significance in the differences. For detailed insights, please refer to Table 1.

Univariate and Multivariate Analysis of Clinical Characteristics

Incorporating age, gender, smoking status, EGFR mutation status, T stage, N stage, and bone metastasis into a univariate analysis, factors with a p-value less than 0.05 were subsequently included in a multivariate analysis. Ultimately, age and EGFR mutation status emerged as significant statistical factors with substantial significance (Table 2).

Table 1 Stratification of Demographic and Clinicopathologic Variables Across the Entire Cohort Based on Brain Metastasis Status.

Variable	Train set	no brain metastasis	brain metastasis	P-value	Validation set	no brain metastasis	brain metastasis	P-value
Age	62.65±9.37	63.61±9.15	60.31±9.53	0.003	62.64±9.80	63.04±9.23	61.80±11.08	0.606
Sex				0.722				0.438
Female	158(48.47%)	110(47.62%)	48(50.53%)		31(39.74%)	19(35.85%)	12(48.00%)
Male	168(51.53%)	121(52.38%)	47(49.47%)		47(60.26%)	34(64.15%)	13(52.00%)
Smoking				0.867				0.386
No	195(59.82%)	137(59.31%)	58(61.05%)		46(58.97%)	29(54.72%)	17(68.00%)
Yes	131(40.18%)	94(40.69%)	37(38.95%)		32(41.03%)	24(45.28%)	8(32.00%)
Primary_site				0.759				0.414
Right upper lobe	84(25.77%)	62(26.84%)	22(23.16%)		17(21.79%)	12(22.64%)	5(20.00%)
Right middle lobe	30(9.20%)	23(9.96%)	7(7.37%)		7(8.97%)	4(7.55%)	3(12.00%)
Right inferior lobe	60(18.40%)	39(16.88%)	21(22.11%)		18(23.08%)	15(28.30%)	3(12.00%)
Left upper lobe	90(27.61%)	63(27.27%)	27(28.42%)		26(33.33%)	17(32.08%)	9(36.00%)
Left inferior lobe	62(19.02%)	44(19.05%)	18(18.95%)		10(12.82%)	5(9.43%)	5(20.00%)
EGFR				0.003				0.73
Wild_type	148(45.40%)	119(51.52%)	29(30.53%)		40(51.28%)	26(49.06%)	14(56.00%)
19del	87(26.69%)	55(23.81%)	32(33.68%)		18(23.08%)	12(22.64%)	6(24.00%)
L858R	91(27.91%)	57(24.68%)	34(35.79%)		20(25.64%)	15(28.30%)	5(20.00%)
T_stage				0.008				0.331
T1	61(18.71%)	52(22.51%)	9(9.47%)		19(24.36%)	13(24.53%)	6(24.00%)
T2	115(35.28%)	70(30.30%)	45(47.37%)		33(42.31%)	23(43.40%)	10(40.00%)
T3	42(12.88%)	31(13.42%)	11(11.58%)		14(17.95%)	7(13.21%)	7(28.00%)
T4	108(33.13%)	78(33.77%)	30(31.58%)		12(15.38%)	10(18.87%)	2(8.00%)
N_stage				0.41				0.599
N0	64(19.63%)	47(20.35%)	17(17.89%)		25(32.05%)	17(32.08%)	8(32.00%)
N1	25(7.67%)	21(9.09%)	4(4.21%)		13(16.67%)	8(15.09%)	5(20.00%)
N2	128(39.26%)	87(37.66%)	41(43.16%)		16(20.51%)	13(24.53%)	3(12.00%)
N3	109(33.44%)	76(32.90%)	33(34.74%)		24(30.77%)	15(28.30%)	9(36.00%)
Bone metastasis				0.534				1
No	182(55.83%)	132(57.14%)	50(52.63%)		40(51.28%)	27(50.94%)	13(52.00%)
Yes	144(44.17%)	99(42.86%)	45(47.37%)		38(48.72%)	26(49.06%)	12(48.00%)

Table 2 Univariate and Multivariate Logistic Regression Analysis of Prognostic Factors for Brain Metastasis.

		Univariate analysis				Multivariate analysis
		OR	95%CI		p_value	OR	95%CI		p_value
Age		0.968	0.946	0.99	0.005	0.974	0.95	0.997	0.03
Sex	Female	Reference
	Male	0.873	0.569	1.338	0.532
Smoking_status	No	Reference
	Yes	0.829	0.532	1.293	0.409
Primary_site	Right upper lobe	Reference			0.891
	Right middle lobe	0.828	0.347	1.973	0.669
	Right inferior lobe	1.241	0.656	2.35	0.507
	Left upper lobe	0.987	0.547	1.781	0.966
	Left inferior lobe	1.142	0.593	2.2	0.691
EGFR_status	Wild_type	Reference			0.004				0.009
	19del	2.317	1.341	4.003	0.003	2.122	1.204	3.74	0.009
	L858R	2.036	1.187	3.492	0.01	2.078	1.197	3.609	0.009
T_stage	T1	Reference			0.083
	T2	2.347	1.219	4.519	0.011
	T3	2.053	0.928	4.539	0.076
	T4	1.784	0.899	3.542	0.098
N_stage	N0	Reference			0.298
	N1	0.612	0.238	1.572	0.307
	N2	1.192	0.662	2.144	0.558
	N3	1.385	0.768	2.499	0.279
Bone metastases	No	Reference
	Yes	1.046	0.681	1.606	0.837

Table 3 Assessment of the Efficacy of Radiomics Signatures and DLRN Prediction in Both the Train set and Validation set.

Signature	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	Precision	Recall	F1	Threshold	Cohort
Rad	0.58	0.648	0.5815 - 0.7148	0.737	0.515	0.387	0.825	0.387	0.737	0.507	0.267	Train
DTL	0.759	0.786	0.7306 - 0.8421	0.611	0.821	0.586	0.836	0.586	0.611	0.598	0.331	Train
Combined	0.938	0.978	0.9652 - 0.9910	0.884	0.961	0.903	0.952	0.903	0.884	0.894	0.409	Train
Nomogram	0.91	0.979	0.9671 - 0.9913	0.958	0.891	0.784	0.981	0.784	0.958	0.863	0.15	Train
Rad	0.548	0.604	0.3982 - 0.8099	0.714	0.514	0.227	0.9	0.227	0.714	0.345	0.267	Validation
DTL	0.548	0.8	0.6275 - 0.9725	0.857	0.486	0.25	0.944	0.25	0.857	0.387	0.017	Validation
Combined	0.595	0.833	0.6637 - 1.0000	0.857	0.543	0.273	0.95	0.273	0.857	0.414	0.041	Validation
Nomogram	0.81	0.837	0.6737 - 0.9998	0.571	0.857	0.444	0.909	0.444	0.571	0.5	0.694	Validation

Features selection

The texture features, such as gray-level co-occurrence matrix (GLCM) and gray-level run-length matrix (GLRLM), characterize the patterns, or the second and high-order spatial distributions of the intensities. From each patient, a total of 107 handcrafted features were extracted, comprising of 16.8% first-order features, 13.1% shape features, and 1496 texture features (including 22.4% GLCM, 15% GLSZM, 15% GLRLM, 4.7%NGTDM, and 13.1% GLDM features). Details of the handcrafted features can be found in Fig. 2A-B. The integration of handcrafted features with deep learning features, followed by feature selection via Lasso regression (Fig. 3).

Predictive performance of three radiomics signatures and DLRN

Table 3 encompasses a comprehensive delineation of the scalar metrics pertaining to various predictive signatures and DLRN, as evaluated in both the train set and the validation set.

In this study, the deep learning model ResNet101 utilizes Grad-CAM (Gradient-weighted Class Activation Mapping) for visualization. Grad-CAM generates class activation maps by computing the gradients of the target class relative to the feature maps of convolutional layers, revealing the image regions that the network focuses on during decision-making (Fig. 4).

As depicted in Fig. 5, DTL achieved an AUC of 0.786 (95% CI: 0.731–0.842) within the train set, surpassing the traditional Rad, which relied on handcrafted features, with a validation set AUC of 0.604 (95% CI: 0.398–0.810). Subsequently, integrating handcrafted and deep learning features yielded Combined model, showcasing enhanced performance, exhibiting an AUC of 0.978 (95% CI: 0.965–0.991) for the train set and 0.833 (95% CI: 0.664-1.000) for the validation set. Furthermore, Fig. 6 illustrates the integration of the Combined model with demographic attributes (age and EGFR status) into DLRN, facilitating the visualization of brain metastasis in lung adenocarcinoma patients classification assessment. Among all radiomics models, DLRN demonstrated peak performance, recording an AUC of 0.979 (95% CI: 0.9671–0.9913), accuracy of 0.910, sensitivity of 0.958, and specificity of 0.894 in the train set, alongside an AUC of 0.837 (95% CI: 0.6737–0.9998), accuracy of 0.810, sensitivity of 0.571, and specificity of 0.857 in the validation set. The DCA curve, presented in Fig. 7, underscores DLRN's superiority over traditional and deep learning-based radiomics signatures, as well as the fusion model, in benefiting patients. Additionally, Fig. 8 showcases excellent concordance between DLRN's predicted outcomes and the actual risk stratification of brain metastasis in lung adenocarcinoma patients, as evidenced by the calibration curves.

The prognosis of brain metastasis in patients with advanced lung adenocarcinoma remains grim, significantly impacting survival rates and quality of life⁸. Early detection and prediction of this devastating complication are paramount, as they enable the implementation of timely and targeted therapeutic interventions⁹. Our study underscores the urgency and importance of developing accurate predictive models to identify patients at high risk of developing brain metastases, thereby facilitating personalized treatment strategies.

Radiomics, an evolving technique, transforms routine radiological images into quantifiable radiomic signatures, enabling the selection of pivotal features to formulate a distinctive pattern predictive of clinical outcomes or endpoints¹⁰. Within the context of brain metastasis in advanced lung adenocarcinoma, multiple investigations have delved into the utilization of CT-derived radiomics for forecasting this complication¹¹. These endeavors underscore the promising role of radiomics as a non-invasive, valuable asset in tailoring evaluations and guiding therapeutic strategies specific to patients with advanced lung adenocarcinoma and brain metastasis^11,12. Nevertheless, several notable shortcomings have emerged from these studies. Chiefly, despite promising outcomes achieved by various radiomic models, these endeavors often suffer from small sample sizes (typically below 200 patients), with a preponderance lacking external validation, thus compromising their generalizability¹³. Additionally, the radiomic features leveraged for model construction across previous studies were predominantly acquired through a preconceived, manual approach. While these handcrafted features embody domain expertise and can be abundant (potentially spanning tens of thousands), they are constrained by their intricate design process, often characterized by shallow and low-level imaging attributes that may inadequately encapsulate tumor heterogeneity^14,15. This limitation ultimately impedes the models' predictive capabilities and hinders their translation into clinical practice.

In various clinical applications, multiple radiomics studies have investigated and incorporated the unique aspects of deep learning. The CT-based deep learning radiomics signature demonstrates clinically viable predictive capabilities for PD-L1 expression, presenting itself as a promising surrogate imaging biomarker and a complementary tool to immunohistochemistry assessment¹⁶. Furthermore, the fusion of deep learning and habitat radiomics enhances prediction of immunotherapy response in NSCLC patients, offering a potential avenue for personalized immunotherapy strategies¹⁷. In a study, the PET/CT-informed deep learning radiomics model accurately forecasts PD-L1 expression in NSCLC patients, presenting clinicians with a non-invasive method to identify PD-L1-positive candidates for treatment¹⁸. Through amalgamating TNM stage, CT radiomic signature, and deep learning markers, the devised nomograms enable prediction of individual prognosis for NSCLC patients undergoing chemotherapy, potentially enhancing personalized treatment and precise patient care ¹⁹. A study revealed that a PET/CT-based deep learning model surpassed a radiomics model in diagnosing EGFR mutation status in NSCLC patients. By integrating the most statistically relevant clinical factor (smoking history) with deep learning attributes, our hybrid model demonstrated superior accuracy in predicting EGFR mutation types, empowering NSCLC patients with more tailored treatment options²⁰.

To the best of our understanding, this research endeavor marks a pioneering step in developing and rigorously validating sophisticated deep learning approaches for predicting advanced lung adenocarcinoma patients exhibiting brain metastases. By harnessing high-resolution CT scans sourced from an extensive, multi-institutional dataset, we aim to refine the understanding of disease progression and facilitate the development of tailored treatment strategies for this challenging condition. Our research demonstrates that the DTL surpasses conventional Rad, attaining an AUC of 0.935 in the internal cohort and 0.710 in the independent validation cohort. By integrating deep learning signatures with traditional radiomic attributes, we constructed combined model, which exhibited even more refined predictive prowess, with an external AUC of 0.757. These discoveries underscore the primacy of deep learning-derived features over manually crafted ones in predicting the risk stratification of lung adenocarcinoma with brain metastases, and hint at their potential to augment existing radiomic models. Additionally, our study pinpoints the crucial radiomic characteristics originating from the deep learning framework, emphasizing the pivotal role these features play in model formulation. This is not unexpected, given the distinctive capacity of deep learning to discern intricate imaging patterns and encapsulate a broader spectrum of imaging variability, compared to traditional radiomic features. The intricate architecture of neural networks facilitates non-linear transformations of imaging data, bridging the gap between input and output spaces, thereby enabling deep learning features to capture higher-order imaging nuances and greater heterogeneity.

Incorporating deep learning, specifically CNNs, and advanced CNN architectures like ResNet-101, into the predictive modeling of brain metastasis in advanced lung adenocarcinoma from thoracic CT images, presents a compelling hybrid approach that marries the strengths of traditional radiomics with the power of deep learning. CNNs, as a subclass of neural networks, have proven to be highly effective in extracting discriminative features from grid-structured data, such as medical images²¹. Their hierarchical layers, consisting of convolutional, pooling, and fully connected operations, enable them to learn increasingly complex representations of the input data, which is crucial for accurate diagnosis and prognosis in medical imaging tasks. ResNet-101, a pioneering CNN architecture, further enhances the capabilities of deep learning for medical image analysis by introducing residual learning²². This mechanism, embodied in residual blocks with skip connections, alleviates the vanishing gradient problem that hinders the training of very deep networks. By enabling the network to learn the residual between layers, ResNet-101 is able to achieve remarkable performance gains, even at depths exceeding 100 layers, making it an ideal candidate for handling complex medical imaging tasks like predicting brain metastasis in lung cancer patients²³.

In the context of our hybrid approach, combining traditional radiomics, which relies on handcrafted features extracted from images, with ResNet-101's deep learning capabilities, offers a comprehensive solution. Radiomics provides a wealth of quantitative information that captures the shape, texture, and intensity patterns within medical images, while ResNet-101 automatically discovers even more subtle and discriminative features that may not be readily apparent to human observers. By fusing these two modalities, our approach aims to enhance the predictive accuracy and robustness of brain metastasis modeling in advanced lung adenocarcinoma, ultimately facilitating earlier detection and more targeted treatment strategies.

We delve deeper into the implications and potential of our proposed nomogram model. This model, crafted as an intuitive scoring system, seamlessly integrates diverse variables, thereby enhancing individualized prediction accuracy and streamlining clinical application. Our initial design philosophy emphasized the incorporation of objective, quantitative attributes that adeptly capture imaging heterogeneity, a pivotal step towards bolstering both model efficacy and its practical applicability in clinical realms. It is noteworthy that we consciously excluded radiological characteristics, such as tumor shape, heterogeneity, contour, calcification presence, enhancement degree, and mediastinal fat infiltration, from our model's construction. This decision stemmed from the fact that these imaging features, primarily derived through visual inspection, inherently introduce subjectivity and potential discrepancies among various radiologists. By eschewing such subjectivity, we aimed to foster a more consistent and reliable predictive tool. Drawing upon the optimal radiomics signature (DLR), coupled with demographic factors like age and gender, we devised the DLRN, a visual aid that proficiently predicts brain metastasis in advanced lung adenocarcinoma. This model, having achieved optimal performance, underscores the efficacy of our hybrid approach that harmoniously blends traditional radiomics with deep learning techniques. Furthermore, the Decision Curve Analysis (DCA) underscores the profound clinical significance of our preoperative DLRN. Its implementation promises substantial benefits for patients, highlighting its potential as a non-invasive yet invaluable clinical instrument. This approach can guide surgical decision-making, adjuvant radiotherapy planning, and even anticipate tumor recurrence and survival, thereby personalizing treatment strategies for advanced lung adenocarcinoma patients and ultimately enhancing their outcomes. In conclusion, our study underscores the transformative potential of integrating traditional radiomics with deep learning methodologies. This innovative hybrid approach not only enhances prediction accuracy but also fosters a more streamlined and patient-centric clinical workflow. As we continue to refine and validate our model, we envision it evolving into a cornerstone in the management of advanced lung adenocarcinoma, particularly in the context of brain metastasis.

While the results are promising, the present research endeavor carries several caveats. Firstly, the retrospective nature of the investigation, despite leveraging a vast multicenter dataset for model development, might have inadvertently introduced a bias in participant selection. For a rigorous validation of the predictive nomogram, a forward-looking study encompassing an even broader patient population is indispensable. Furthermore, the integration of multifaceted data sources, for instance, CT imaging, holds promise to enrich the informational landscape and refine the accuracy of the predictive framework in subsequent investigations. Additionally, the third limitation of this research lies in its exclusive focus on advanced lung adenocarcinoma. This narrow focus restricts the applicability and generalizability of the findings to other types of lung cancer or even beyond lung malignancies. To truly advance the field and provide a comprehensive predictive framework, future endeavors must endeavor to broaden the scope to encompass a wider range of pathological entities, thereby fostering a more holistic understanding of lung cancer and its prognosis.

Study population

The current investigation retrospectively analyzed CT images of patients with advanced lung adenocarcinoma to predict the occurrence of brain metastases. The study population comprised 326 patients from Hebei University Affiliated Hospital serving as the training set and 78 patients from Baoding First Central Hospital constituting the validation set. The inclusion criteria for this study were: 1) histologically confirmed advanced lung adenocarcinoma with the potential for brain metastasis, 2) availability of standard contrast-enhanced CT scans performed within 30 days prior to any surgical or non-surgical intervention, and 3) absence of any antitumor treatments prior to CT examination. Patients were excluded if: 1) the CT image quality was compromised due to artifacts or other factors, or 2) contrast-enhanced CT scans were unavailable. Data was extracted and collated from the electronic medical record systems of the participating hospitals. Ethical approval was granted by the Institutional Review Boards of all contributing institutions, with a waiver of informed consent granted due to the retrospective nature of the study. In total, 404 patients with advanced lung adenocarcinoma were enrolled across the two hospitals and stratified into two cohorts. The train set, comprising 326 patients from Hebei University Affiliated Hospital, was utilized for model development and internal validation. The validation set, consisting of 78 patients from Baoding First Central Hospital, served as an external test set to evaluate the model's performance.

This study conforms to the following declarations: (i) The experiments were approved by the Ethics Committee of the Affiliated Hospital of Hebei University, with all necessary permissions and licenses obtained prior to the commencement of the study. Any additional relevant details, such as approval numbers or specific conditions, have been duly noted and adhered to by the committee. (ii) All experiments were performed in strict accordance with the relevant guidelines and regulations set forth by the Ethics Committee of the Affiliated Hospital of Hebei University, ensuring the highest standards of ethical conduct and scientific integrity were maintained throughout the study.

Image acquisition and tumor regions of interest segmentation

Patients from two esteemed medical centers, Hebei University Affiliated Hospital and Baoding No.1 Central Hospital, underwent customized CT scanning procedures utilizing distinct systems and parameters. Prior to and subsequent to intravenous injection of iodine-based contrast agents, chest CT scans were administered. Patients were positioned supine, ensuring that the scanning encompassed all affected regions. A volume ranging from 80 to 100 mL of contrast material was administered intravenously via the antecubital vein at a steady rate of 2.5 ml/s. The acquired transverse CT images were retrieved from the respective picture archiving and communication systems (PACS) for this study. To identify tumor regions of interest (ROIs), two seasoned radiologists, each with over ten years of experience and blinded to the pathological details, manually outlined each axial, contrast-enhanced CT slice using ITK-SNAP (version 3.8).

Image preprocessing

Recognizing that CT images from different institutions may exhibit variations in pixel values due to differences in imaging techniques, we implemented a preprocessing step to normalize these values. Specifically, we sorted all pixel values and adjusted intensities to fall within the 0.5th to 99.5th percentile range, thereby mitigating outliers. Furthermore, to address potential variations in voxel spacing, we applied spatial normalization. This was complemented by a fixed resolution resampling method (1 × 1 × 1 mm³), ensuring consistency across images and facilitating accurate analysis in our experiments.

Radiomics features extraction

In this study, radiomics feature extraction encompassed a blend of traditional, manually crafted features (geometry, intensity, and texture) and deep learning features autonomously discerned by a Convolutional Neural Network (CNN) trained on the dataset. The manually crafted features, extracted utilizing the Pyradiomics tool (https://github.com/Radiomics/pyradiomics), were organized into three categories: geometry, intensity, and texture. Geometry features encapsulate the tumor's 3D shape attributes, intensity features portray the primary statistical distribution of voxel intensities within the tumor, offering a comprehensive yet non-redundant assessment.

In this study, ResNet101 served as the CNN framework and was employed for extracting deep learning features. ResNet101 is a deep CNN architecture that utilizes residual learning to facilitate training of very deep networks. It consists of 101 layers, making it a powerful model for tasks such as image recognition and classification. By introducing residual blocks, ResNet101 enables the network to learn residuals, simplifying the learning task and mitigating issues like vanishing gradients, thereby enhancing performance and generalization ability.

Feature selection and signatures construction

In the derivation cohort, a rigorous feature selection process was undertaken. For manually crafted features, we secured the reliability of ROI-derived attributes by double-checking segmentation in a random subset of 100 patients, independently performed by two radiologists. Subsequently, inter-rater agreement was assessed using ICC, retaining only those features exceeding an ICC threshold of 0.85 for robustness. To pinpoint predictors highly discriminative of risk categories, statistical significance (p < 0.05) via t-tests was imposed. Additionally, we evaluated feature redundancy via Pearson correlation, eliminating pairs with coefficients > 0.9 to avoid duplication. A greedy approach iteratively pruned the feature set, targeting the most redundant, ensuring comprehensive yet concise representation. Lastly, the LASSO logistic regression model was leveraged to further condense the feature pool, identifying the most pertinent subset for crafting an informative signature.

Given the substantial dimensionality of 2048 in deep learning features, we applied Principal Component Analysis (PCA) as a strategy to harmonize these features and reduce their dimensionality to 32. This reduction enhanced the model's generalization capabilities and helped mitigate the potential for overfitting.

Furthermore, we delved into a hybrid strategy by merging meticulously crafted radiomics features with deep learning features, leveraging the fusion of these two realms to bolster our model. Specifically, we employed an early fusion approach, incorporating features sifted through both traditional handcrafted methods and our deep learning framework, to form a comprehensive and complementary feature ensemble. The subsequent feature selection process mirrored that employed for the handcrafted features, ensuring consistency in our methodology.

Following this feature selection, we crafted two distinct signatures: a conventional radiomics signature (Rad) utilizing handcrafted features, and a deep transfer learning signature (DTL) leveraging deep learning features. Additionally, we ventured into constructing a combined model, a deep transfer learning-based radiomics signature, which integrated both types of features.

To unleash the full potential of our predictive endeavors, we harnessed XGBoost, a cutting-edge gradient boosting framework renowned for its adeptness in navigating intricate and nonlinearly structured datasets, surpassing a myriad of other algorithms. To refine our model to its optimal configuration, we implemented a rigorous grid search strategy, intricately intertwined with a stringent 5-fold cross-validation approach, within our derivation cohort. This comprehensive approach ensured that our model was not only accurate but also robust and reliable.

Construction of DLRN

To gain deeper insights into the classification performance, we crafted a unique logistic regression model, incorporating a meticulously crafted deep learning-rooted radiomics network (DLRN) enriched with a comprehensive set of clinical markers. This ensemble surpassed mere age and gender, embracing vital clinical features like tumor localization, EGFR mutation status, the presence or absence of bone metastasis, thereby offering a multidimensional lens for assessing patient stratification.

Performance assessment and model comparison

To comprehensively evaluate the predictive prowess of our model for brain metastasis in advanced lung adenocarcinoma patients, we harnessed the power of receiver operating characteristic (ROC) analysis alongside a suite of quantitative metrics (AUC, sensitivity, specificity, accuracy, and F1 score) tailored to the optimal ROC threshold. We conducted a rigorous comparison of these metrics across three distinct radiomic signatures (Rad, DTL, and Combined model) as well as our DLRN, ensuring a holistic assessment across all cohorts. To ensure fairness and robustness, we employed a five-fold cross-validation strategy within the derivation cohort, with an additional fixed external test cohort for independent validation. Furthermore, we leveraged decision curve analysis (DCA) to gauge the clinical relevance and applicability of our predictive model. The calibration curve served as a vital tool to assess the congruence between observed outcomes and model predictions. Figure 1 elegantly outlines the comprehensive study design and workflow.

Statistical analysis

In terms of clinical data analysis, we applied either independent t-tests or Mann-Whitney U tests, depending on the data distribution, to scrutinize continuous variables. For categorical variables, we opted for Fisher's exact test or chi-square tests, ensuring a comprehensive assessment. Statistical significance was established at a two-sided p-value threshold of < 0.05, thereby identifying genuine differences with confidence.

In conclusion, extending our previous research paradigm, we have ventured into the realm of brain metastases arising from advanced lung adenocarcinoma. Our study harnessed the power of enhanced CT imaging to devise and validate a deep learning-infused radiomics signature, specifically tailored for predicting the risk stratification of these intricate brain metastases within a multi-institutional patient cohort. This methodology surpassed the predictive prowess of traditional handcrafted feature-based radiomics, underscoring the potential of deep learning in deciphering complex tumor-brain interactions.

Moreover, we introduced a groundbreaking fused DLRN model, a harmonious blend of deep learning and meticulously crafted features. This innovative union not only elevated our model's performance to unparalleled heights but also showcased its capability to offer supplementary diagnostic insights. Ultimately, this approach holds immense promise in enhancing personalized treatment strategies for patients battling advanced lung adenocarcinoma with brain metastases, paving the way for more precise and effective therapeutic interventions.

Acknowledgements

I am deeply grateful to the Onekey platform for its invaluable assistance in facilitating my research. The user-friendly interface and comprehensive data support have been instrumental in the success of this study.

Author contributions

As a key contributor, Qie Shuai played a pivotal role in developing the hybrid approach that combines traditional radiomics and deep learning techniques. He likely led the efforts in designing and implementing the radiomics features extraction pipeline from the thoracic CT images, ensuring their accuracy and relevance for predicting brain metastasis. Liu Sukun's contributions likely focused on the deep learning aspect of the study. She may have been responsible for selecting, training, and optimizing the deep learning architectures used in the analysis. Shi Hongyun's contributions likely revolved around the clinical and medical aspects of the study. She may have provided valuable insights into the clinical significance of brain metastasis in advanced lung adenocarcinoma, guiding the research questions and objectives. As a member of the research team, Liu Ming's contributions were likely multifaceted. He may have assisted in data collection, preprocessing, and analysis, ensuring that the study's foundation was solid and reliable.

Data availability statement

The data used in this study, including thoracic CT images and associated clinical information, were sourced from the Affiliated Hospital of Hebei University and the First Central Hospital of Baoding. Due to privacy and ethical considerations, the raw data are not publicly available. However, the processed data and features used in the development of the predictive model, as well as the code for implementing the hybrid approach combining traditional radiomics and deep learning, will be provided upon reasonable request to qualified researchers, subject to appropriate data sharing agreements and institutional approvals. Please note that the corresponding author Professor Liu Ming should be contacted if someone wants to request the data from this study.

Competing interests

The authors declare no competing interests.

Siegel, R. L., Giaquinto, A. N. & Jemal, A. Cancer statistics, 2024. CA Cancer J Clin 74, 12-49, doi:10.3322/caac.21820 (2024).
Xu, J. et al. Distinguishing EGFR mutation molecular subtypes based on MRI radiomics features of lung adenocarcinoma brain metastases. Clin Neurol Neurosurg 240, 108258, doi:10.1016/j.clineuro.2024.108258 (2024).
Zhang, X. et al. Deep learning PET/CT-based radiomics integrates clinical data: A feasibility study to distinguish between tuberculosis nodules and lung cancer. Thorac Cancer 14, 1802-1811, doi:10.1111/1759-7714.14924 (2023).
Wang, C. et al. Deep Learning to Predict EGFR Mutation and PD-L1 Expression Status in Non-Small-Cell Lung Cancer on Computed Tomography Images. J Oncol 2021, 5499385, doi:10.1155/2021/5499385 (2021).
Ye, G. et al. CT-based quantification of intratumoral heterogeneity for predicting pathologic complete response to neoadjuvant immunochemotherapy in non-small cell lung cancer. Front Immunol 15, 1414954, doi:10.3389/fimmu.2024.1414954 (2024).
Laqua, F. C. et al. Transfer-Learning Deep Radiomics and Hand-Crafted Radiomics for Classifying Lymph Nodes from Contrast-Enhanced Computed Tomography in Lung Cancer. Cancers (Basel) 15, doi:10.3390/cancers15102850 (2023).
Guo, J. et al. Predicting bone metastasis-free survival in non-small cell lung cancer from preoperative CT via deep learning. NPJ Precis Oncol 8, 161, doi:10.1038/s41698-024-00649-z (2024).
Ohno, M. et al. Development of a scoring system to predict local recurrence in brain metastases following complete resection and observation. J Neurooncol, doi:10.1007/s11060-024-04790-4 (2024).
Wang, Z. et al. Single-cell transcriptomic analyses provide insights into the cellular origins and drivers of brain metastasis from lung adenocarcinoma. Neuro Oncol 25, 1262-1274, doi:10.1093/neuonc/noad017 (2023).
Kong, C., Yin, X., Zou, J., Ma, C. & Liu, K. The application of different machine learning models based on PET/CT images and EGFR in predicting brain metastasis of adenocarcinoma of the lung. BMC Cancer 24, 454, doi:10.1186/s12885-024-12158-0 (2024).
Shi, J. et al. Using Radiomics to Differentiate Brain Metastases From Lung Cancer Versus Breast Cancer, Including Predicting Epidermal Growth Factor Receptor and human Epidermal Growth Factor Receptor 2 Status. J Comput Assist Tomogr 47, 924-933, doi:10.1097/rct.0000000000001499 (2023).
Deng, F. 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 46, 1309-1320, doi:10.1007/s13246-023-01300-0 (2023).
Cao, R. et al. Radiomics evaluates the EGFR mutation status from the brain metastasis: a multi-center study. Phys Med Biol 67, doi:10.1088/1361-6560/ac7192 (2022).
Cong, P. et al. Development and validation a radiomics nomogram for diagnosing occult brain metastases in patients with stage IV lung adenocarcinoma. Transl Cancer Res 10, 4375-4386, doi:10.21037/tcr-21-702 (2021).
Wang, G. et al. Radiomics signature of brain metastasis: prediction of EGFR mutation status. Eur Radiol 31, 4538-4547, doi:10.1007/s00330-020-07614-x (2021).
Xu, T. et al. CT-based deep learning radiomics biomarker for programmed cell death ligand 1 expression in non-small cell lung cancer. BMC Med Imaging 24, 196, doi:10.1186/s12880-024-01380-8 (2024).
Caii, W. et al. Integration of deep learning and habitat radiomics for predicting the response to immunotherapy in NSCLC patients. Cancer Immunol Immunother 73, 153, doi:10.1007/s00262-024-03724-3 (2024).
Li, B., Su, J., Liu, K. & Hu, C. Deep learning radiomics model based on PET/CT predicts PD-L1 expression in non-small cell lung cancer. Eur J Radiol Open 12, 100549, doi:10.1016/j.ejro.2024.100549 (2024).
Chang, R. et al. Nomograms integrating CT radiomic and deep learning signatures to predict overall survival and progression-free survival in NSCLC patients treated with chemotherapy. Cancer Imaging 23, 101, doi:10.1186/s40644-023-00620-4 (2023).
Huang, W. et al. PET/CT Based EGFR Mutation Status Classification of NSCLC Using Deep Learning Features and Radiomics Features. Front Pharmacol 13, 898529, doi:10.3389/fphar.2022.898529 (2022).
Manini, C. et al. Impact of training data composition on the generalizability of CNN aortic cross section segmentation in 4D Flow MRI. J Cardiovasc Magn Reson, 101081, doi:10.1016/j.jocmr.2024.101081 (2024).
Hlavata, R., Kamencay, P., Radilova, M., Sykora, P. & Hudec, R. Automated Method for Intracranial Aneurysm Classification Using Deep Learning. Sensors (Basel) 24, doi:10.3390/s24144556 (2024).
Arian, R. et al. SLO-Net: Enhancing Multiple Sclerosis Diagnosis Beyond Optical Coherence Tomography Using Infrared Reflectance Scanning Laser Ophthalmoscopy Images. Transl Vis Sci Technol 13, 13, doi:10.1167/tvst.13.7.13 (2024).

No competing interests reported.

Predictive Modeling of Brain Metastasis in Advanced Lung Adenocarcinoma: A Hybrid Approach Combining Traditional Radiomics and Deep Learning from Thoracic CT Images

Status:

Version 1

Abstract

Figures

Introduction

Result