Radiomics nomogram for prediction disease-free survival and adjuvant chemotherapy benefits in patients with resected stage I lung adenocarcinoma.

BACKGROUND: Robust imaging biomarkers are needed for risk stratification in stage I lung adenocarcinoma patients in order to select optimal treatment regimen. We aimed to construct and validate a radiomics nomogram for predicting the disease-free survival (DFS) of patients with resected stage I lung adenocarcinoma, and further identifying candidates benefit from adjuvant chemotherapy (ACT).
METHODS: Using radiomics approach, we analyzed 554 patients' computed tomography (CT) images from three multicenter cohorts. Prognostic radiomics features were extracted from computed tomography (CT) images and selected using least absolute shrinkage and selection operator (LASSO) Cox regression model to build a radiomics signature for DFS stratification. The biological basis of radiomics was explored in the Radiogenomics dataset (n=79) by gene set enrichment analysis (GSEA). Then a nomogram that integrated the signature with these significant clinicopathologic factors in the multivariate analysis were constructed in the training cohort (n=238), and its prognostic accuracy was evaluated in the validation cohort (n=237). Finally, the predictive value of nomogram for ACT benefits was assessed.
RESULTS: The radiomics signature with higher score was significantly associated with worse DFS in both the training and validation cohorts (P<0.001). The GSEA presented that the signature was highly correlated to characteristic metabolic process and immune system during cancer progression. Multivariable analysis revealed that age (P=0.031), pathologic TNM stage (P=0.043), histologic subtype (P=0.010) and the signature (P<0.001) were independently associated with patients' DFS. The integrated radiomics nomogram showed good discrimination performance, as well as good calibration and clinical utility, for DFS prediction in the validation cohort. We further found that the patients with high points (point ≥8.788) defined by the radiomics nomogram obtained a significant favorable response to ACT (P=0.04) while patients with low points (point <8.788) showed no survival difference (P=0.7).
CONCLUSIONS: The radiomics nomogram could be used for prognostic prediction and ACT benefits identification for patient with resected stage I lung adenocarcinoma. 2020 Translational Lung Cancer Research. All rights reserved.

Introduction

Lung adenocarcinoma remains a critical challenge for global public health as it gradually evolves into the largest subtype of non-small-cell lung cancer, with its proportion increasing to over 60% (1). Complete surgical excision is conducted as the primary intervention for patients with stage I lung adenocarcinoma, but the 5-year recurrence rate of these patients reaches 30% (2). Because postoperative recurrence results in an extremely poor survival outcome, adjuvant chemotherapy (ACT) has been used in an attempt to reduce its occurrence. The absolute survival improvement of ACT for patients with resected stage II-III lung cancer has been endorsed (3,4). But its effect in stage I disease remains undetermined, especially in the stage IB (5,6). Previous studies revealed that the adoption of ACT was beneficial for some patients with stage I disease but harmful for those who could be cured by surgery alone (7,8). Therefore, a more accurate stratification within the same stage may allow for identifying low and high-risk patients for recurrence and selecting those with more likelihoods to benefit from ACT.

Medical imaging is a vital technology in clinical management to aid decision making and direct personalized treatment (9). Radiomics, as an advanced imaging analytic process for extracting high dimensional features from medical images, has been demonstrated with promising performance in prognosis evaluation for various types of cancers (10-12). A nomogram integrating the radiomic biomarkers with clinicopathological features appears to improve the prognostic accuracy (13). To date, there are five radiomics nomograms have been developed for prognosis prediction in lung adenocarcinoma (). But a radiomics nomogram research which investigates its association with survival of stage I lung adenocarcinoma and further guides candidates’ selection for ACT, has not yet been fully reported.

Table S1

Radiomic nomograms for prognosis prediction of lung adenocarcinoma in the literatures

Literature	Year	Patient	Sample size	Variates included in nomogram
Huang et al. (13)	2016	Early stage (I and II) NSCLC	282	Radiomic signature, age, clinical stage, histologic grade, gender
Desseroit et al. (44)	2016	Stage I–IIIb NSCLC	116	Overall stage, selected radiomics features (metabolic volume, PET entropy, CT zone percentage)
Wang et al. (45)	2019	Advanced NSCLC	118	Radiomics signature, age, lymph node, lymphocyte2, NLR1
Yang et al. (46)	2019	Stage I–IIIb NSCLC	371	Radiomics signature, age, sex, T stage, N stage
Akinci D’Antonoli et al. (47)	2020	stage Ia–IIb NSCLC	124	Radiomics signature, overall stage

Therefore, the aim of present study was to construct and validate a nomogram based on radiomics and clinical features to estimate the DFS in resected stage I lung adenocarcinoma patients and to further explore its potential value for predicting survival benefit from ACT. Additionally, we intended to explore the potential biological basis of radiomics with imaging and gene expression data.

We present the following article in accordance with the STROBE Reporting Checklist (available at http://dx.doi.org/10.21037/tlcr-19-577).

Methods

Data collection

The study was approved by The Institutional Review Boards of Shanghai Pulmonary Hospital and the informed consent was waived for this retrospective study (No. k19-134Y). Consecutive patients who received complete resection for lung cancer in Shanghai Pulmonary Hospital from January 2011 to December 2012 were retrieved. The patients diagnosed with stage I adenocarcinoma according to the eighth edition TNM staging system were identified (14). Patients were excluded based on the following criteria: (I) patients without thin-slice CT images (1 mm) within one month prior to surgery; (II) patients with incomplete clinicopathologic data and follow-up records; (III) patients whose lesions cannot be accurately segmented. The patients undergone surgical resection before December 2011 were set as the training cohort (n=238), and after December 2011 as an independent validation cohort (n=237). In addition, a dataset comprised CT imaging data and matched RNA sequencing data of 79 resected stage I adenocarcinoma patients was obtained from the NSCLC Radiogenomics Dataset (15) of the Cancer Imaging Archive (TCIA) to evaluate the biological process of radiomics signature.

The follow-up data was obtained through electronic medical records and telephone interviews as a complete. We set the endpoint of this study as disease-free survival (DFS), which was defined as the duration from the date of surgery to that of recurrence, death, or the last follow-up. The histologic subtype was re-evaluated according to the current lung adenocarcinoma classification system (16).

Imaging data acquisition and radiomics processing

All details of CT image acquisition protocols in our study were described in Supplementary material I. After downloading CT images from the institutional Picture Archiving and Communication Systems with DICOM format, tumor segmentation was performed in an open-source software (3D-slicer, v4.10.1, www.slicer.org) (17). The regions of interest (ROI) on CT images were manually delineated layer-by-layer by two junior radiologists (T.W. and Y.Y., with 3 and 5 years of experience, respectively) and verified by a senior radiologist (X.S., with 30 years of experience in chest imaging diagnosis). An open-source platform, PyRadiomics in Python, was utilized to extract 107 radiomics features from the non-filtered segmented ROI (18). The details of radiomics features were descripted in Supplementary material II and .

Table S2

Extracted radiomics features

Intensity	Shape	Texture
		GLCM	GLRLM	GLSZM	NGTDM	GLDM
10 Percentile	Elongation	Autocorrelation	Gray level nonuniformity	Gray level nonuniformity	Busyness	Busyness
90Percentile	Flatness	Cluster Prominence	Gray level nonuniformity normalized	Gray level nonuniformity normalized	Coarseness	Coarseness
Energy	Least axis length	Cluster Shade	Gray level variance	Gray level variance	Complexity	Complexity
Entropy	Major axis length	Cluster Tendency	High gray level run emphasis	High gray level zone emphasis	Contrast	Contrast
Interquartile range	Maximum 2D diameter column	Contrast	Long run emphasis	Large area emphasis	Strength	Strength
Kurtosis	Maximum 2D diameter row	Correlation	Long run high gray level emphasis	Large area high gray level emphasis		Dependence entropy
Maximum	Maximum 2D diameter slice	Difference Average	Long run low gray level emphasis	Large area low gray level emphasis		Dependence nonuniformity
Mean absolute deviation	Maximum 3D diameter	Difference Entropy	Low gray level run emphasis	Low gray level zone emphasis		Dependence nonuniformity normalized
Mean	Mesh volume	Difference Variance	Run entropy	Size zone nonuniformity		Dependence variance
Median	Minor axis length	Id	Run length nonuniformity	Size zone nonuniformity normalized		Gray level nonuniformity
Minimum	Sphericity	Idm	Run length nonuniformity normalized	Small area emphasis		Gray level variance
Range	Surface area	Idmn	Run percentage	Small area high gray level emphasis		High gray level emphasis
Robust mean absolute deviation	Surface volume ratio	Idn	Run variance	Small area low gray level emphasis		Large dependence emphasis
Root mean squared	Voxel volume	Imc1	Short run emphasis	Zone entropy		Large dependence high gray level emphasis
Skewness		Imc2	Short run high gray level emphasis	Zone percentage		Large dependence low gray level emphasis
Total energy		Inverse variance	Short run low gray level emphasis	Zone variance		Low gray level emphasis
Uniformity		Joint average				Small dependence emphasis
Variance		Joint energy				Small dependence high gray level emphasis
		Joint entropy				Small dependence low gray level emphasis
		MCC
		Maximum probability
		Sum average
		Sum entropy
		Sum squares

Radiomics feature selection and signature construction

We determined the optimum cutoff value for every radiomics feature based on its association with the patients’ DFS using X-tile program (version 3.6.1, Yale University School of Medicine, New Haven, CT, USA) (19). Accordingly, the significant features (P<0.05) which could divide patients into different risk groups were selected. To reduce the redundancy, the least absolute shrinkage and selection operator (LASSO) Cox regression model was adopted to select the optimal radiomics feature subset. Then a radiomics signature was obtained through the selected features and their respective weighted coefficients.

Survival analysis based on the radiomics signature

The radiomics signature stratified the patients of training cohort into low- and high-risk groups based on its optimal cutoff value which was identified by X-tile program. The Kaplan-Meier survival curves and log-rank test were delineated to seek the survival difference between the risk groups. These calculation and cutoff value were applied to the validation cohort. Meanwhile, subgroups analyses based on the clinicopathologic risk factors were performed for its robustness assessment.

Gene set enrichment analysis (GSEA)

We performed a GSEA to explore the biological basis of the developed radiomics signature for prognosis prediction. The RNA-seq expression data in Fragments Per Kilobase of transcript per Million mapped reads (FPKM) values for all samples were obtained from GSE103584 and transformed to log2 (FPKM + 0.1). Genes with missing expression values in more than 50% of all samples were removed, resulting in 12,901 genes. Then we created a rank of these genes using a fold change which was defined as the difference value the mean gene expression between the high- and low-risk group. The pre-ranked GSEA was performed using the PIANO R package (20) with 10,000 permutations. As gene sets, we tested expert-curated pathways from the C2 Reactome collection version 7.0 available at MSigDB (21). Gene sets were restricted to sizes between 15 and 500, resulting in 805 tested gene sets. The Enrichment Score (ES) was utilized to quantify the association of the rank of genes with pathways and validated with false discovery rate (FDR) to corrected for multiple comparison.

Development and validation of the radiomics nomogram

The univariate and multivariate Cox regression analyses with stepwise selection were carried out in the training cohort to identify the independent prognostic factors. Then a nomogram was applied for DFS prediction visualization. Corresponding calibration curves was developed in the validation cohort; the Harrell concordance index (C-index) quantified its discrimination ability; a decision curve analysis exhibited its clinical utility by measuring the net benefits at different threshold probabilities.

ACT benefit analysis based on the nomogram

A polynomial equation extracted from the prognostic nomogram were used for calculating the total risk point, which might reflect the weight of each predictor and correspond to the survival probability. The optimal cutoff risk point of stage IB patients was identified to divide them into low- and high-risk groups and the survival analyses were performed to evaluate the benefits of ACT.

Statistical analysis

Patients’ baseline characteristics in two cohorts were compared using analyses of independent t test for continuous variates and a chi-squared test for categorical variates. Statistical analyses were accomplished via R software, version 3.5.3 (http://www.R-project.org) and SPSS for Windows, version 20.0 (IBM, Armonk, NY, USA). The packages used in R programming were showed in Supplementary material III. A two-sided P value below 0.05 was considered statistically significant.

Results

Clinicopathologic characteristics

The baseline characteristics of all included patients were listed in . The median follow-up time were 66.3 and 64.7 months, the 5-year DFS with 80.3% and 79.5% in the training and validation cohorts, respectively. There was no significant difference between two cohorts in terms of clinicopathologic factors or follow-up time.

Table 1

The baseline characteristics of 475 patients with resected stage I lung adenocarcinoma in the training and validation cohorts

Characteristic	Training cohort, N=238, n (%)	Validation cohort, N=237, n (%)	P value
Age, years			0.436
<65	165 (69.3)	172 (72.6)
≥65	73 (30.7)	65 (27.4)
Gender			0.966
Female	129 (54.2)	128 (54.0)
Male	109 (45.8)	109 (46.0)
Smoking			0.107
No	164 (68.9)	179 (75.5)
Yes	74 (31.1)	58 (24.5)
p-TNM stage			0.383
IA	121 (50.8)	111 (46.8)
IB	117 (49.2)	126 (53.2)
Surgery types			0.396
Lobectomy	221 (92.9)	215 (90.7)
Sub-lobar resection	17 (7.1)	22 (9.3)
Histologic subtype			0.589
LPA	83 (34.9)	83 (35.0)
APA	91 (38.2)	88 (37.1)
PPA	37 (15.5)	46 (19.4)
SPA	17 (7.1)	15 (6.3)
MPA	10 (4.2)	5 (2.1)
Pathologic tumor size, mm, mean ± SD	20.2±8.2	20.1±8.4	0.948
Follow-up time, median	66.3	64.7	0.054

SD, standard deviation; LPA, lepidic predominant adenocarcinoma; APA, acinar predominant adenocarcinoma; PPA, papillary predominant adenocarcinoma; MPA, micropapillary pattern-predominant adenocarcinoma; SPA, solid predominant adenocarcinoma.

Radiomics feature-based signature construction

A total of 58 radiomics features were significant (P<0.05) for risk stratification in training cohort. Then an optimal feature subset consisting of eight radiomics features, including maximum, minimum, IDN, joint energy, long run low gray level emphasis, gray level variance, coarseness, and large dependence emphasis, were identified mostly related to the DFS by the LASSO Cox regression analysis (). The radiomics score (rad-score), was constructed with the calculation formula generating in the LASSO model (Supplementary material IV).

Figure S1

Radiomics feature selection using the least absolute shrinkage and selection operator (LASSO) regression model. (A) LASSO coefficient profiles of the radiomics features. As the tuning parameter (λ) increased using 5-fold cross-validation, more coefficients tended to approach 0 and the optimal non-zero coefficients generated, which yielded a set of the optimal radiomics features; (B) the partial likelihood deviance from the LASSO regression cross-validation procedure was plotted against log(λ).

The distribution of rad-score and survival status indicated that patients with higher score generally had worse survival than did those with lower score (, left panel). The optimal cutoff value of rad-score was 1.125, which divided all patients into low- (rad-score <1.125) and high-risk group (rad-score ≥1.125). The survival analyses showed that a significant difference existed between these groups (P<0.001) both in the training and validation cohorts (, right panel). Furthermore, the survival analyses applied in the clinicopathologic subgroups of all patients, consisting of age, sex, pathological TNM stage and histologic subtype, further proved that this signature remained significant for survival stratification (P<0.01, ).

Figure 1

The distribution of radiomics score (left panel) and KM survival (right panel) in the training (A) and validation cohorts (B). The P values of survival curves were calculated using the log-rank test.

Figure S2

Kaplan-Meier survival analysis for all 475 patients with stage I lung adenocarcinoma according to the eight-feature-based radiomics signature stratified by clinicopathological risk factors.

Biological basis of the radiomics signature

The biological basis of the radiomic signature was evaluated in the independent Radiogenomics dataset with CT images and RNA-sequence-based gene expression data. The pre-ranked GSEA showed that the significant enriched pathways (FDR <0.1) among the top associations with the radiomics signature were mostly correlated to various metabolic processes and immune system (). This revealed that the developed imaging biomarker might reflect the different characteristic metabolic changes and immune system during the cancer progression, which could provide more potential for early prognosis prediction of lung adenocarcinoma.

Figure 2

Gene set expression patterns—Radiogenomics dataset. The radiomics signature was linked to gene expression patterns using a pre-ranked gene set enrichment analysis (GSEA). Positive and negative enrichments are shown in red and blue, respectively. The top 10 enrichments in each category are highlighted.

Performance analysis of the radiomics nomogram

In the multivariate Cox regression analysis (), age (HR: 1.893; 95% CI: 1.061–3.379; P=0.031), pathologic TNM stage (HR: 2.109; 95% CI: 1.039–4.338; P=0.043), histologic subtype (LPA: reference; APA/PPA: HR: 2.703; 95% CI: 1.039–7.035, P=0.042; MPA/SPA: HR: 5.300; 95% CI: 1.782–15.767, P=0.003) and the radiomics signature (HR: 7.794; 95% CI: 3.185–19.078; P<0.001) were identified as independent predictors for DFS of patients with stage I adenocarcinoma.

Table 2

Univariate and multivariate analysis of disease-free survival for patients in training cohort

Variables	Univariate			Multivariate
	HR (95% CI)	P		HR (95% CI)	P
Age (≥65 vs. <65)	2.506 (1.414–4.442)	0.002		1.893 (1.061–3.379)	0.031
Gender (Male vs. Female)	2.120 (1.176–3.819)	0.012		1.568 (0.850–2.890)	0.150
Smoking status (Yes vs. No)	1.602 (0.894–2.871)	0.113		NA	NA
p-TNM stage (IB vs. IA)	4.243 (2.109–8.535)	<0.001		2.109 (1.039–4.338)	0.043
Operation type (Lobe vs. Sub-lobe)	0.966 (0.300–3.115)	0.954		NA	NA
Histologic subtype		0.001			0.010
LPA	1.000 (Reference)			1.000 (Reference)
APA/PPA	4.720 (1.839–12.118)	0.001		2.703 (1.039–7.035)	0.042
MPA/SPA	7.854 (2.682–22.998)	0.002		5.300 (1.782–15.767)	0.003
Pathologic tumor size	1.763 (1.266–2.455)	<0.001		1.097 (0.779–1.545)	0.596
Radiomics signature	11.528 (4.890–27.178)	<0.001		7.794 (3.185–19.078)	<0.001

HR, hazard ratio; CI, confidence interval; LPA, lepidic predominant adenocarcinoma; APA, acinar predominant adenocarcinoma; PPA, papillary predominant adenocarcinoma; MPA, micropapillary pattern-predominant adenocarcinoma; SPA, solid predominant adenocarcinoma.

Then a nomogram that incorporated these independent clinical factors and radiomics signature was generated (). We found that the radiomics nomogram showed good discrimination performance (C-index: 0.713; 95% CI: 0.646–0.780) in the validation cohort, as well as satisfactory agreements among prediction outcomes and actual observations (). The decision curves analysis exhibited that the integrating radiomics nomogram achieved higher clinical usefulness relative to the clinicopathologic factors and radiomics signature alone ().

Figure 3

A prediction performance analysis of patients with stage I lung adenocarcinoma. (A) The nomogram for predicting 3- and 5-year DFS after surgery; (B) plots depict the calibration of the nomogram in terms of agreement between predicted and observed 3- and 5-year DFS. Performance of the validation cohort was shown on the plot relative to the 45-degree line, which represented perfect prediction; (C) decision curve analysis for the comparison of prognostic factors. The y-axis measures the net benefit.

ACT benefit analysis based on the radiomics nomogram

The adoption of ACT (n=129/243) did not show survival benefit in all patients with stage IB lung adenocarcinoma (P=0.11, ). The polynomial equation (Supplementary material V) extracted from the radiomics nomogram was used to calculate the total risk point for all stage IB patients and divided them into low- and high-point groups with its optimal cutoff value. Compared with the patients in the low-group, the high-point group accounted for more patients with older age, male, micropapillary and solid predominant adenocarcinoma, larger tumor size and higher radiomics signature (P<0.05, ). Interestingly, while patients with a low risk point (risk point <8.788) defined by the radiomics nomogram showed no survival difference with or without ACT (P=0.7; ), patients with high point (total point ≥8.788) obtained a favorable response to ACT (P=0.04; ). These results indicated that the integrated radiomics nomogram had great potential to successfully select these high-risk patients with stage IB lung adenocarcinoma who were suitable candidates for ACT.

Figure S3

The Benefit Analysis of adjuvant chemotherapy (ACT) in All Patients with Stage IB Adenocarcinoma. These patients showed no survival difference between with and without ACT.

Table 3

The Clinicopathological features of 243 patients with resected stage IB lung adenocarcinoma in the low- and high-risk groups defined by the radiomics nomogram

Characteristic	Low-group, N=129, n (%)	High-group, N=114, n (%)	P value
Age, years			<0.001
<65	102 (79.1)	62 (54.4)
≥65	27 (20.9)	52 (45.6)
Gender			0.004
Female	86 (66.7)	55 (48.2)
Male	43 (33.3)	59 (51.8)
Smoking			0.467
No	97 (75.2)	81 (71.1)
Yes	32 (24.8)	33 (28.9)
Surgery types			0.225
Lobectomy	125 (96.9)	107 (93.9)
Sub-lobar resection	4 (3.1)	7 (6.1)
Histologic subtype			<0.001
LPA	55 (42.6)	17 (14.9)
APA	48 (37.2)	51 (44.7)
PPA	25 (19.4)	20 (17.5)
SPA	0 (0.0)	18 (15.8)
MPA	1 (0.8)	8 (7.0)
Pathologic tumor size, mm, mean ± SD	21.9±7.8	24.8±8.9	0.007
Radiomics signature, mean± SD	0.789±0.336	1.373±0.297	<0.001

SD, standard deviation; LPA, lepidic predominant adenocarcinoma; APA, acinar predominant adenocarcinoma; PPA, papillary predominant adenocarcinoma; MPA, micropapillary pattern-predominant adenocarcinoma; SPA, solid predominant adenocarcinoma.

Figure 4

The Benefit analysis of adjuvant chemotherapy (ACT) in different subgroups. (A) Patients in the low-risk groups defined by the nomogram showed no survival difference between with and without ACT; (B) patients in the high-risk groups defined by the nomogram could obtain survival benefits from ACT.

Discussion

Patients with stage I lung adenocarcinoma have substantial risks for recurrence even after complete resection, and whether ACT could provide survival benefit remains controversial. Postoperative DFS evaluation is vital for guiding a patient’s individualized follow-up strategies and subsequent treatment option. In this research, the constructed radiomics signature could successfully categorize stage I adenocarcinoma patients into high- and low-risk groups with significant different in DFS and the integrated radiomics nomogram could identify the suitable patients benefit from ACT. Furthermore, the genomic studies revealed the correlations between radiomics signature and tumor metabolic changes and immune system.

The analysis of medical images has gradually evolved from subjective operator-dependent criteria to a more objective and quantitative evaluation (22,23). Radiomics could provide additional information in oncologic practice related to benign and malignant nodules differentiation (24), mutation types identification (10), subtype classification (25) and treatment response assessment (26). As for prognostic prediction based on lung CT, Lee et al. (27) found that two radiomics features of 339 patients were independent prognostic factor of lung adenocarcinoma survival. However, this method of describing gross lesions with few single features required further improvement as it might cause an underestimation of the application of radiomics (28). Hence, in radiomics analysis, a multi-feature-based radiomics signature has significantly greater prognostic accuracy than that of each selected feature alone, considering the interactions between different features (12,13,29). In our study, eight-feature-consisted radiomics signature based on multicenter cohorts of 475 patients reached a favorable performance for survival assessment of stage I lung adenocarcinoma.

As for the administration of ACT in lung adenocarcinoma, current guidelines do not recommend its adoption in stage IA disease but remain more ambiguous in stage IB disease (30,31). Our data supported that the adoption of ACT in all stage IB patients without selection could not bring survival benefit (P=0.11). Numerous studies have demonstrated that stage IB patients with older age are associated with greater recurrence risk (8,32,33). The histologic features of lung adenocarcinoma, especially the micropapillary and solid predominant patterns, remained independent predictor for evaluating tumor recurrence and ACT survival benefits (3,34-36). Our study is consistent with these studies. Moreover, we add to these findings by demonstrating that the radiomics nomogram integrating the proposed imaging signature with these clinicopathologic predictors is more clinically relevant because accurate survival prediction allows for good identification of patients who derive therapeutic benefit from ACT. The radiomics signature provides the incremental value for guiding the adoption of ACT in patients with stage IB disease that complemented clinicopathologic factors (37).

With respect to elaborating the underlying mechanism that drives biological process ultimately reflected in imaging phenotypes, series studies published by Aert et al. (38-40) have found that some biological pathways related to lung cancer prognosis, including mitosis, cell cycle and transcription, can be captured by imaging biomarkers. Our study used a GSEA in an external cohort to reveal how imaging biomarkers can be connected to postoperative prognosis, suggesting that different tumor metabolism and immune system change may result in different survival outcomes. Cancers have much more complex metabolic pathways to support tumor cell growth and proliferation (41). The radiomics features for constituting the signature can correlate with these metabolic pathways and immune system editing, which allows for tumor development and progression, and thereby worsening the prognosis (40,42).

Some limitations existed in our study besides for those commonly associated with retrospective studies. First, despite the constructed signature has been verified a favorable performance in identifying high-risk patients for recurrence and suitable candidates for ACT benefit on an independent validation cohort, multicenter datasets were still necessary for validating its robustness and generalization. Second, the task of ROIs delineation was completed in a manual method which was commonly considered as a gold criterion for images processing at present (43). However, the exploitation of a more efficient and perfectly accurate method for lesion segmentation will be an important direction in the field of future radiomics research. Third, only adenocarcinoma patients were enrolled in this study. Due to its enormous proportion in lung cancer and the heterogeneity in tumor behavior, a homogeneous study cohort only including lung adenocarcinoma patients may be better for the model analysis of radiomics.

In conclusion, our results suggest that the developed radiomics signature may successfully categorize stage I adenocarcinoma patients into high- and low-risk groups of disease recurrence, and the integrated radiomics nomogram may further identify the suitable patients benefit from ACT. Moreover, we evaluated the biological basis of the proposed signature and found to be correlated with different characteristic metabolic changes and immune system. These findings may potentially facilitate the clinical impact in guiding pertinent therapeutic administration and improved survival outcomes of patients with stage I lung adenocarcinoma.

The article’s supplementary files as