https://orcid.org/0009-0001-3357-4495
Figures
Abstract
Objectives
The International Association for the Study of Lung Cancer (IASLC) grading system is key to the prognosis and treatment of Invasive Pulmonary Adenocarcinoma (IPA). However, current radiomics and other radiological approaches poorly capture tumor heterogeneity, limiting predictive power. This study aimed to develop an interpretable CT-based model that predicts the histological grading of IPA by decoding its intratumoral spatial heterogeneity.
Materials and methods
This multi‑center retrospective study enrolled 355 IPA patients, split into training/validation (7:3) and an independent test cohort. Tumors were graded as low‑grade (Ⅰ/Ⅱ) or high‑grade (Ⅲ) per IASLC criteria. Intratumoral subregions were generated via unsupervised clustering of CT images, and their spatial interaction heterogeneity was quantified using a Multi-regional Spatial Interaction (MSI) matrix. Five models (clinical‑radiological, radiomics, MSI, radiomics‑combined, MSI‑combined) were built using four preprocessors and five classifiers. The optimal model was selected based on the Receiver Operating Characteristic (ROC) curve in the validation cohort, with generalizability assessed in the test cohort. Performance was compared via the DeLong test, and SHapley Additive exPlanations (SHAP) analysis interpreted feature contributions.
Results
Three subregions were generated. The high-grade group exhibited a larger proportion of Subregion 1, while showing a smaller proportion of Subregion 2. The MSI model based on 10 MSI features achieved an AUC of 0.806 in the test cohort, outperforming clinical‑radiological, radiomics, and radiomics‑combined models (p = 0.002, 0.010, 0.022). Adding clinical‑radiological features did not improve the MSI model (p = 0.083). SHAP identified MSI_border_proportion_2_3 (relative border proportion between Subregions 2 and 3) as the most influential feature, with lower values indicating high‑grade IPA.
Citation: Qin S, Zhou S, Liu Y, Su D, Jia Q, Li Y, et al. (2026) Intratumoral spatial heterogeneity at non-contrast CT predicts histological grading of invasive pulmonary adenocarcinoma: a multicenter retrospective study. PLoS One 21(2): e0341163. https://doi.org/10.1371/journal.pone.0341163
Editor: Jincheng Wang, Hokkaido University: Hokkaido Daigaku, JAPAN
Received: November 2, 2025; Accepted: January 3, 2026; Published: February 2, 2026
Copyright: © 2026 Qin et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The anonymized dataset supporting the findings of this study is available in two forms within the submission: As a Supporting Information file (S8 Dataset, S9 Dataset, S10 Dataset, S11 Text). Accompanied by a detailed README file (README.txt) that describes all variables, the coding scheme for categorical values (e.g., 1=Male, 2=Female), and the anonymization procedure (e.g., age reported in aggregated groups). To protect participant privacy, all direct identifiers have been removed. Due to the detailed clinical nature of the data and the limited sample size in specific subgroups, the complete underlying patient-level dataset cannot be shared publicly to avoid potential re-identification. However, it is available to qualified researchers for validation purposes upon reasonable request. Requests for data access, subject to a formal data use agreement, should be directed to the Corresponding Author (Xiufu Zhang, xray1895@163.com) or the Ethics Committee of Jiangjin Central Hospital of Chongqing.
Funding: This study was funded by the Beijing Medical Award Foundation (Grant No. YXJL-2025-0483-0265).
Competing interests: Author Yang Li is employed by Shanghai United Imaging Intelligence Co., Ltd. This commercial affiliation does not alter our adherence to PLOS ONE policies on sharing data and materials. The other authors have declared that no competing interests exist.
Introduction
Adenocarcinoma is the most common lung cancer subtype [1]. Invasive Pulmonary Adenocarcinoma (IPA) is more malignant and has a poorer prognosis than either adenocarcinoma in situ or minimally invasive adenocarcinoma [2,3]. Establishing a histological grading system is critical for evaluating the malignancy grade of IPA. Therefore, in 2020, the International Association for the Study of Lung Cancer (IASLC) proposed a grading system for IPA, which is based on the most predominant histological subtype and the proportion of high-grade components [4]. The IASLC grading system not only serves as an independent predictor of Overall Survival (OS) and Recurrence-Free Survival (RFS) in IPA but also informs clinical decision-making. It aids in identifying candidates for preoperative neoadjuvant therapy [5–7]. Furthermore, for patients who undergo surgery, the grading system provides a critical rationale for postoperative adjuvant chemotherapy, as evidenced by studies demonstrating superior outcomes in high-grade (Grade 3) stage Ib-III patients receiving such treatment [8,9]. Therefore, the accurate preoperative assessment of IASLC grade is crucial for predicting prognosis and formulating optimal treatment strategies in patients, specifically determining the extent of resection (lobectomy or segmentectomy) and deciding whether to use neoadjuvant/adjuvant therapy.
The minimally invasive biopsies with high diagnostic yield are generally considered the preferred option for the preoperative diagnosis of lung adenocarcinoma. However, preoperative biopsy is limited by tumor heterogeneity and insufficient sample size, making a comprehensive assessment of histological grades challenging [10]. CT imaging holds significant value in diagnosing the malignancy grade and pathological subtypes of lung cancer [11,12]. A higher IASLC grade of IPA was associated with more invasive features on CT, suggesting that CT imaging can help characterize tumor grade [13,14]. However, due to the diversity and overlap of CT features in IPA—coupled with the fact that assessment results are prone to being influenced by physicians’ subjective experience—these factors can all lead to suboptimal predictive performance.
Radiomics enables the extraction of a series of high-dimensional quantitative features from medical images, thereby helping reveal underlying imaging patterns imperceptible to the naked eye [15]. In recent years, it has been applied to histological grading, invasiveness prediction, etc., of IPA [16,17]. However, conventional radiomics analysis methods often treat tumors as homogeneous wholes and overlook intratumoral subregional variations; this practice ultimately leads to the masking of local heterogeneity information. To address this limitation, a new analytical method is required: it employs algorithms (e.g., unsupervised clustering) to partition voxels with similar imaging features into distinct subregions, which may exhibit unique growth and invasion patterns. This approach not only overcomes the drawbacks of conventional radiomics but also better captures intratumoral heterogeneity [18–20]. For instance, an intratumoral heterogeneity score derived from subregion analysis has proven valuable for distinguishing histological subtypes in lung adenocarcinoma presenting as pure ground-glass nodules [21] and for predicting high-grade patterns in solid lung adenocarcinoma [22]. Furthermore, extending this concept to a three-dimensional space incorporating multiple perspective heterogeneity metrics has demonstrated superior performance in preoperative IASLC grading [23]. The aforementioned studies have effectively characterized tumor heterogeneity by quantifying subregional attributes and their spatial distribution. To further capture the potential spatial interactions among subregions, the Multi-regional Spatial Interaction (MSI) matrix is introduced to quantify the adjacency and spatial co-occurrence relationships of subregions in three-dimensional space, thereby providing a complementary characterization of tumor spatial heterogeneity [24]. Previous studies have established that MSI matrix-derived spatial heterogeneity features are independent prognostic factors for OS and RFS in Non-Small Cell Lung Cancer (NSCLC) [25]. However, the relationship between this specific form of spatial heterogeneity, quantified by the MSI matrix, and the histological grading of IPA remains to be fully investigated.
Therefore, this study aims to investigate the predictive value of spatial heterogeneity quantified by the combined approach of density-based subregion analysis and MSI matrix for the histological grades of IPA. Our goal is to provide valuable supplementary information to pathological diagnosis from an imaging perspective, ultimately to assist in improving preoperative grading accuracy.
Materials and methods
Patients
This retrospective study was conducted in accordance with the Declaration of Helsinki and received approval from the Institutional Review Boards of all participating hospitals: the Ethics Committee of Jiangjin Central Hospital of Chongqing (Center 1, KY20250619−002), the Ethics Committee of Chongqing General Hospital (Center 2, IIT-S2025-041–01), and the Ethics Committee of the People’s Hospital of Chongqing Hechuan (Center 3, HX-2025-020-001). We retrospectively collected data from 599 patients with pathologically confirmed lung adenocarcinoma who underwent surgical resection between June 2022 and February 2025. The data were accessed for research purposes on August 1, 2025. During the initial data collection, the authors had access to identifying information (including patient names and medical record numbers); however, all such identifiers were permanently removed and replaced with anonymous study codes before statistical analysis to ensure confidentiality. The requirement for informed consent was waived. To avoid clustering effects, only one largest lesion was selected for subsequent analysis in patients with multiple lesions.
A sample size estimation was performed to ensure adequate statistical power for model development. The calculation was based on the standard formula for estimating a diagnostic proportion with a specified precision [26]:
where =1.96 corresponds to a two-sided significance level of
=0.05,
is the expected proportion (sensitivity or specificity), and
is the allowable error (set at 0.08). Using expected sensitivity (80%) and specificity (70%) as reference, the minimum required sample size was calculated to be 224.
Inclusion Criteria: (1) Chest CT scan performed within one month prior to surgery; (2) Availability of detailed postoperative pathological reports. Exclusion Criteria: (1) Poor-quality CT images; (2) History of preoperative treatment; (3) Pathological subtypes including minimally invasive adenocarcinoma, invasive mucinous adenocarcinoma, and other variants of adenocarcinoma (e.g., colloid adenocarcinoma, fetal adenocarcinoma, or enteric adenocarcinoma).
Ultimately, 355 patients with IPA were enrolled in this study. From Center 1 and Center 2, a total of 227 patients were randomly assigned to a 161 (n = 161) or a validation cohort (n = 66) in a 7:3 ratio. An additional 128 patients from Center 3 formed an independent test cohort.
According to the IASLC grading system, grade 3 is defined as invasive adenocarcinoma with a total proportion of high-grade patterns (solid, micropapillary, or complex glandular patterns) of 20% or more, irrespective of the predominant subtype. Grade 3 IPA is associated with a poorer prognosis and benefits from adjuvant chemotherapy, demonstrating significant differences compared to grade 1 and 2 tumors [4,5,7]. Based on this, we stratified the enrolled patients into a low-grade group (grades 1 and 2) and a high-grade group (grade 3). The flowchart of patient enrollment is shown in Fig 1.
CT imaging protocol
Non-enhanced chest CT scans were performed at each center using the following scanners: Center 1 utilized the United Imaging UCT530 40-slice CT scanner (United Imaging Healthcare, Shanghai, China), Center 2 employed the Philips IQon Spectral 64-slice CT scanner (Philips Medical Systems, Best, The Netherlands), and Center 3 used the Philips Brilliance iCT 128-slice scanner. Detailed scanning parameters for all three centers are provided in S1 Table.
Clinical and radiological characteristics
Patient baseline characteristics, including gender, age, smoking history, alcohol consumption, and clinical stages, were recorded using the electronic medical record system. Two radiologists with 10 and 7 years of experience in thoracic imaging, respectively, independently and blindly evaluated the following radiological characteristics: involved lobe, diameter, shape, attenuation, lobulation, air bronchogram, spiculation, pleural tag, and cavity. Any discrepancies between the two readers were resolved by consensus.
Tumor segmentation
Tumor regions were automatically segmented following image preprocessing using the deep learning model VB-Net within the uAI Research Portal (version: 20240730). The model achieved an average Dice Similarity Coefficient (DSC) of 91.5% in delineating pulmonary nodules [27]. The two-dimensional slices were reconstructed into three-dimensional volumes of interest (VOIs). Subsequently, two radiologists with 8 and 15 years of work experience respectively performed layer-by-layer independent correction on the VOIs. The inter-observer DSC was 0.873. For regions with high inter-observer consistency (DSC ≥ 0.700), the segmentation results were adopted from the radiologist with more extensive experience (15 years of practice).
Radiomic feature extraction
Before radiomic feature extraction, CT images underwent standardized preprocessing, which included the following steps: signal intensity normalization through gray-level discretization (bin width: 25); spatial resampling of voxels to an isotropic resolution of 1 mm × 1 mm × 1 mm using B-spline interpolation; and window setting normalization with a window width of 1500 HU and a level of −600 HU. Fourteen imaging filters were applied to the images for post-processing. Features were subsequently extracted from both the original and filtered image sets. A total of 2264 radiomic features were obtained from each VOI, comprising 104 original features, 432 gray-level statistical features, and 1728 textural features.
Density-based subregion analysis
The original CT images in NIfTI format were loaded using SimpleITK. The window width and level were normalized to 1500 HU and −600 HU, respectively. Subsequently, the voxel values were scaled to the range [0, 1] using max-min normalization. The subregion segmentation was conducted in two steps. First, simple linear iterative clustering was applied to the voxel values within each VOI to perform supervoxel segmentation at the individual level (n_segments = 100). Second, a Gaussian mixture model was employed to cluster supervoxels with similar imaging phenotypes across the cohort, thereby identifying distinct subregions. The optimal number of subregions was determined based on cluster validity indices. The same criteria and quantitative procedures were applied to both the validation and independent test cohorts to obtain consistent subregion segmentation results.
Subsequently, an MSI matrix was constructed to characterize the spatial interaction heterogeneity by recording the co-occurrence relationships among subregions. First-order and second-order statistical features were then derived from the MSI matrix. The methodology for constructing the MSI matrix and the interpretation of the corresponding features are described in the S2 Text and S3 Fig.
Feature selection
To address the issue of class imbalance, the synthetic minority oversampling technique (SMOTE) was applied exclusively to the training cohort to generate synthetic samples for the minority class, with concurrent down-sampling of the majority class (parameters: perc. over = 200, perc. under = 155).
All feature processing and selection steps described below were performed strictly within the training cohort. To eliminate scale differences, both the MSI and radiomic features were standardized separately using Z-score normalization. For each feature set, recursive feature elimination was then applied to remove redundant and non-discriminative features, followed by the application of Least Absolute Shrinkage and Selection Operator (LASSO) regression to select the most predictive subset. During LASSO modeling, 10-fold cross-validation was used to determine the optimal regularization parameter α based on the minimal cross-validation error for each feature set. Univariate and multivariate logistic regression analyses of clinical-radiological characteristics were performed to identify predictors independently associated with grading.
Model construction
All models were trained and their hyperparameters were tuned using a uniform pipeline, and these processes were performed exclusively in the training cohort. Given the unknown distributional characteristics of the dataset—such as non-normality, outliers, and scale variations—and considering that different classifiers rely on distinct data assumptions, we adopted a comprehensive modeling strategy. This approach combined four preprocessing methods (Box-Cox transformer, MaxAbs scaling, Quantile transformer, and Yeo-Johnson transformer) with five classifiers (Random Forest (RF), Logistic Regression (LR), Quadratic Discriminant Analysis (QDA), Support Vector Machine (SVM), and Decision Tree (DT)) to identify the optimal model for the current dataset. Based on the selected clinical-radiological, MSI, and radiomic features, we constructed three distinct models (clinical-radiological, MSI, and radiomics) and two integrated models: an MSI-combined model (incorporating clinical-radiological and MSI features) and a radiomics-combined model (incorporating clinical-radiological and radiomic features).
Model performance evaluation and interpretability analysis
The validation cohort was used solely to select the optimal model, while the independent test cohort was employed strictly to evaluate the model’s generalization performance. Receiver Operating Characteristic (ROC) curves were plotted, and the Area Under the Curve (AUC) was calculated to evaluate the models’ ability to discriminate between different grades. Calibration curves were generated along with the Brier score to assess the agreement between predicted probabilities and actual outcomes. Decision Curve Analysis (DCA) was applied to quantify the net clinical benefit. DeLong’s test was used to compare AUC values among different models.
The model’s decision-making process was interpreted using SHapley Additive exPlanations (SHAP; version 0.46.0) to visualize feature contributions and enhance clinical usability.
Statistical analysis
Normality of continuous variables was assessed using the Kolmogorov-Smirnov test. Data conforming to a normal distribution are presented as mean ± standard deviation (Mean ± SD), and group comparisons were performed using the independent samples t-test. Non-normally distributed data are expressed as median (interquartile range) [M (Q1, Q3)], and the Mann-Whitney U test was used for intergroup comparisons. Categorical variables are summarized as frequency (percentage) [n (%)], and group differences were assessed using the chi-square test. The statistical significance level was set at p < 0.05. All statistical analyses were conducted using the uAI Research Portal and SPSS statistical software (version 26.0; IBM, Armonk, NY, USA).
Results
Clinical-radiological characteristics and clinical-radiological model construction
This study included a total of 355 patients with IPA, among whom 66 (18.6%) were classified into the high-grade group. The clinical-radiological characteristics of the three cohorts are summarized in Table 1. Most baseline characteristics showed no significant differences across the cohorts, with the exception of clinical stages (p < 0.001). Multivariable logistic regression analysis identified nodule/mass attenuation (p < 0.001) and lobulation (p = 0.025) as independent radiologic characteristics (Table 2). Among the clinical-radiological models constructed based on the above features, the combination of Box-Cox transformation and LR demonstrated the best predictive performance on the validation cohort and was therefore selected as the final model (S4 Table). This model achieved AUC values of 0.790, 0.633, and 0.720 in the training, validation, and independent test cohorts, respectively (Table 3; Figs 2a-2c).
a Training cohort, b validation cohort, and c independent test cohort.
Radiomic feature and radiomics model construction
A total of 10 radiomic features were retained after selection (S5 Table). In the validation cohort, the combination of Yeo-Johnson transformation and LR yielded the best performance (S4 Table) and was therefore selected as the final model. This model achieved AUC values of 0.989, 0.694, and 0.684 in the training, validation, and independent test cohorts, respectively (Table 3).
Subregion segmentation
The optimal number of clusters was determined to be 3 based on the elbow method (Figs 3a-3c). The three density-based subregions exhibited distinct mean CT attenuation values. Subregion 1 exhibited a mean CT value of −72.6 HU, representing the component with the highest relative attenuation. This region is characterized by closely packed tumor cells, thickened alveolar septa, and minimal air content, corresponding to the denser, “sub-solid” portion of the tumor. Subregion 2 showed a mean CT value of −827.25 HU, representing the component with the lowest attenuation and predominantly consisting of air-filled alveolar structures and areas of sparse cellularity. Subregion 3 had a mean CT value of −424.05 HU, corresponding to the intermediate-attenuation component that consists of a mixture of “sub-solid” tissue and aerated structures with variable cellularity. The mean CT values of Subregion 1 (for which the p-value was 0.822), Subregion 2 (p = 0.255), and Subregion 3 (p = 0.838) did not differ significantly across the training, validation, and independent test cohorts, and this result supports the consistent mapping of CT density-based subregions across cohorts.
(a) patient-level supervoxel generation using simple linear iterative clustering, followed by (b) cohort-level phenotypic clustering with a Gaussian mixture model. The optimal number of clusters (k = 3) was determined using the elbow method based on the rate of change in the sum of squared errors. The cohort-level clustering results are visualized in (c). Color coding: red indicates Subregion 1, green represents Subregion 2, and blue corresponds to Subregion 3.
MSI feature and MSI model construction
A total of 33 first-order and 5 second-order statistical features were derived from the MSI matrix. The high-level group exhibited a substantially larger proportion of “Subregion 1” (median values of 0.667, 0.426, and 0.776 across the three cohorts), with broader interfaces to adjacent regions and the background matrix. In contrast, “Subregion 2” occupied a notably smaller proportion (median values of 0 across all cohorts) and demonstrated more limited boundary contact (S6 Table) (Figs 4a-4b). Ten MSI features were retained to construct the MSI model (S7 Table). The combination of Box-Cox transformation and QDA demonstrated the best performance in the validation cohort (S4 Table) and was therefore selected as the final model. This model achieved AUC values of 0.851, 0.807, and 0.806 in the training, validation, and independent test cohorts, respectively. In the independent test cohort, the MSI model exhibited superior predictive performance, significantly outperforming both the clinical-radiological model (p = 0.002) and the radiomics model (p = 0.010) (Table 4). Furthermore, the calibration curve indicated good agreement between predicted and actual probabilities, with a Brier score of 0.179 (Figs 5a-5c). DCA further confirmed that the MSI model provided higher net clinical benefit across a range of risk thresholds (Figs 5d-5f).
Patient A (high-grade group): Subregion 1: 73.74%, Subregion 2: 0%, Subregion 3: 26.26% (a). Patient B (low-grade group): Subregion 1: 16.16%, Subregion 2: 6.06%, Subregion 3: 77.78% (b).
We further integrated the independent radiological characteristics with the selected MSI and radiomic features to develop combined models. For the radiomics-combined model, the combination of MaxAbs Scaling and QDA demonstrated the best predictive performance (S4 Table), achieving AUC values of 0.953, 0.707, and 0.717 in the training, validation, and independent test cohorts, respectively. For the MSI-combined model, the optimal performance was attained using Box-Cox transformation combined with QDA (S4 Table), with corresponding AUC values of 0.827, 0.810, and 0.782 across the three cohorts. No statistically significant difference was observed between the predictive performance of the MSI-combined model and the MSI model (p > 0.05) (Table 4).
Interpretability analysis of the MSI model
At the global level, the feature importance ranking (Fig 6a) revealed that MSI_border_proportion_2_3 was the most influential feature. This feature represents the relative border proportion between subregion 2 and 3. The beeswarm plot (Fig 6b) shows that higher values of MSI_border_proportion_2_3 correspond to negative SHAP values, indicating a lower likelihood of high-grade classification.
At the individual level, the SHAP force plots visually demonstrate the decision-making rationale of the MSI model for two representative patients (Figs 6c-6d).
Subgroup analysis of the MSI model in different tumor sizes
The MSI model was employed to assess predictive performance across tumor size subgroups stratified as ≤20 mm, > 20 to ≤30 mm, and >30 mm. This classification is based on TNM staging thresholds and also ensures adequate subgroup sample sizes for comparative analysis. In the independent test cohort, the MSI model exhibited the highest diagnostic performance for IPA graded as >20 to ≤30 mm, with an AUC of 0.833 (95% CI: 0.690–0.976). This was followed by the diagnostic performance for IPA > 30 mm, which had an AUC of 0.822 (95% CI: 0.610–1.000). In contrast, the model showed relatively lower diagnostic performance for IPA ≤ 20 mm, with an AUC of 0.658 (95% CI: 0.363–0.954).
Discussion
This study integrated CT density-based subregion analysis and the MSI matrix to quantify the spatial heterogeneity of IPA, and further explored this heterogeneity’s association with IASLC grading. Results showed that the MSI model exhibited the optimal performance; SHAP analysis revealed that MSI_border_proportion_2_3 (relative border proportion of subregion 2 to subregion 3) was the most important feature. From the perspective of spatial heterogeneity, this study initially achieved preoperative prediction of IASLC grading, which can provide pathologists with auxiliary diagnostic evidence.
IPA is a tumor with high heterogeneity, and its typical characteristic often involves a mixture of two or more histological subtypes [28]. However, traditional radiomics methods typically extract features by treating the entire tumor as a uniform region, thereby overlooking the differences between intratumoral subregions [11]. In fact, a tumor is a complex system comprising multiple subregions with distinct microenvironmental differences. Each subregion may correspond to different environmental selection pressures and cellular evolutionary paths. Independent analysis of these subregions allows for more precise characterization of the tumor’s heterogeneity, thereby revealing its biological behavioral characteristics more authentically [29]. Previous studies [30,31] have employed subregion analysis (e.g., via K-means clustering) to predict treatment response in cancers like nasopharyngeal carcinoma and NSCLC. However, these approaches have primarily focused on features extracted from individual subregions, with limited exploration of how these subregions are adjacency in three-dimensional space. In this study, we therefore introduced the MSI matrix, which aims to complement current methods by explicitly quantifying the adjacency and interaction patterns between subregions. Given that 3 subregions were identified in the study, a 4 × 4 MSI matrix (including 1 background region) was constructed. In the MSI matrix, diagonal elements reflect subregion volume, while off-diagonal elements quantify inter-subregion adjacency. We found a higher mean value of off-diagonal elements in high-grade tumors, indicating a greater overall quantity of interfaces between different subregions. This aligns with the pathological understanding that active proliferation and destructive infiltration in high-grade IPA disrupt the tissue architecture, resulting in a complex mixture of pathological components and blurred interregional borders [32]. The effectiveness of this method has been preliminarily verified in previous studies: Wu et al. [24] identified three subregions in breast cancer perfusion MR images and constructed an MSI matrix similarly. By using the derived features of the MSI matrix to perform risk stratification on patients, they found that this stratification was independently associated with RFS, confirming that imaging spatial heterogeneity has significant prognostic value.
Across the three subregions, Subregion 1 exhibited the highest mean CT attenuation, corresponding to the denser tissue component within the tumor. Previous studies [33,34] have linked elevated CT values to higher cellular density and desmoplastic stroma, which are histologic hallmarks of aggressive tumors. Therefore, Subregion 1 may correspond to more invasive pathological patterns, such as solid or micropapillary subtypes. Consistent with this, the higher proportion of Subregion 1 in our high-grade group aligns with the established pathological knowledge that high-grade IPA contains a greater proportion of these aggressive components. Notably, despite its association with denser tissue, the mean CT value of Subregion 1 was negative (−72.6 HU), which was lower than that of pure soft tissue (typically >0 HU). This phenomenon may be attributed to the influence of partial-volume averaging at tissue interfaces. Furthermore, in the MSI model, SHAP analysis identified two key predictive features: MSI_border_proportion_2_3 and MSI_border_proportion_0_2. Lower values of these features were associated with a higher likelihood of high-grade IPA. This observation is grounded in the pathological correlates of these subregions. Subregion 2, with very low CT attenuation, predominantly corresponds to air-containing structures or lepidic growth patterns, which represent indolent, non-invasive components. Subregion 3 exhibits intermediate attenuation, reflecting a mixture of denser tissue and aerated structures. Critically, the feature MSI_border_proportion_2_3 directly measures the extent of contact between these two components. A higher proportion implies greater connectivity between the lepidic/air-rich areas (Subregion 2) and the mixed-density tissue (Subregion 3), suggesting a tumor architecture that is more organized and less solid, a characteristic of a lower-grade, less invasive phenotype. Conversely, a lower proportion indicates that Subregion 3 has little direct contact with Subregion 2. In high-grade IPA, aggressive components expand as consolidated masses, pushing and displacing pre-existing lepidic structures, thereby minimizing this interface [35]. Consequently, our model captures a biomarker that reflects how tumor growth remodels internal architecture, yielding insights beyond simple measurements of component presence or volume.
This study also analyzed the overall spatial interaction heterogeneity of subregion distribution from the second-order statistics. Results showed that the high-grade group exhibited lower MSI_contrast and higher MSI_correlation. In our analysis, MSI_contrast quantifies the local differences in the strength of spatial interactions between adjacent subregions. A lower value indicates that the strength of spatial interactions between subregions is relatively homogeneous throughout the tumor. This likely mirrors the pathological feature of high-grade IPA, where high-grade components display a destructive, coalescent growth pattern [36]. This growth pattern blurs and intermingles the boundaries between different intra-tumoral “subregions”, rather than maintaining clearly separated compartments, and this characteristic is quantified as reduced contrast in the MSI matrix. MSI_correlation measures the linear dependency between row and column element values in the MSI matrix, which can be understood as the regularity or predictability of spatial interaction patterns across different local directions within the tumor. A higher value implies that, despite the intrinsic structural complexity of high-grade tumors, the spatial interaction patterns between subregions exhibit a certain degree of systematic and ordered arrangement in three-dimensional space. This observation is supported by prior research on tumor imaging phenotypes. In a study of NSCLC, Sujit et al. [25] similarly reported that a more aggressive disease state (high recurrence risk) was correlated with lower contrast and higher correlation. This further supports the close association between imaging spatial heterogeneity and the malignant behavior of tumors.
Among clinical-radiological characteristics, this study identified lobulation and attenuation of nodule/mass as independent predictors for grading, findings which are consistent with previous studies [16,37]. Lobulation sign and solid density are common invasive imaging features in high-grade IPA. However, when these CT features were further incorporated into the MSI model, the predictive performance of the model was not significantly improved. This result suggests that the features derived from the MSI matrix may have integrated key imaging information related to grading, which can reduce reliance on physicians’ subjective morphological interpretation to a certain extent.
This study has several limitations. First, the enrolled cohort had a relatively low rate of high-grade cases (18.6%), which may affect the stability of estimates for some metrics (e.g., yielding a wide confidence interval for AUC in subgroup analyses and a modest positive predictive value). Although we addressed class imbalance in the training cohort using SMOTE, the model’s performance in the independent test cohort mirrors this real-world challenge. Future studies with cohorts actively enriched for high-grade cases or large-scale multi-center collaborations are needed to further improve the precision of the model’s predictive performance. Second, subregion analysis was restricted to non-enhanced CT images—largely because non-enhanced CT was the most consistent and standardized protocol across all sites in this retrospective multi-center study. We note that contrast-enhanced CT could provide complementary information on tumor vascularity and heterogeneity; thus, while the current framework demonstrates the independent value of MSI features, it fails to capture all relevant tumor biological characteristics obtainable via multi-parametric imaging (e.g., contrast-enhanced CT, PET). Future work will integrate such multi-parametric data into subregion clustering to enhance the model. Third, translational application of the MSI model to routine clinical workflows faces practical challenges: (1) workflow integration: the automated segmentation/analysis pipeline (including subregion analysis and MSI matrix construction) needs embedding as a seamless plugin in existing PACS/RIS systems for efficient “one-click” DICOM processing; (2) clinical decision-making: the model’s probability output requires a clinically validated threshold (e.g., for guiding resection extent), which should be established via multidisciplinary consensus and net clinical benefit assessment; (3) clinician adoption: an intuitive reporting system that integrates quantitative predictions with visually clear, interpretable insights—for example, by highlighting key SHAP-based features—is essential to help clinicians correlate these outputs with conventional imaging assessment. In future studies, prospective multi-center validation and the development of a user-friendly, DICOM-compliant application will be undertaken to further confirm the model’s utility across diverse clinical settings.
Conclusions
This study demonstrates that the CT-based MSI model can predict histological grading of IPA by decoding the spatial interaction heterogeneity of different subregions in the tumor. By leveraging SHAP analysis to elucidate feature contributions, we enhanced the model’s interpretability and clinical application. These findings offer a valuable supplement to pathological diagnosis, providing objective imaging evidence to guide preoperative risk stratification and personalized treatment planning.
Supporting information
S1 Table. CT scan parameters in center 1, center 2, and center 3.
https://doi.org/10.1371/journal.pone.0341163.s001
(DOCX)
S2 Text. MSI matrix construction and feature interpretation.
This file provides a detailed, step-by-step description of the methodology for constructing the MSI matrix, along with the specific mathematical interpretations of each derived MSI feature used in the MSI model.
https://doi.org/10.1371/journal.pone.0341163.s002
(DOCX)
S3 Fig. Schematic illustration of the MSI matrix construction.
Shown in sequence: (a) The mapping of the three density-based subregions (and the background region) to a 4 × 4 matrix; (b) The meaning of diagonal/off-diagonal elements; (c) Types and quantities of the extracted MSI features.
https://doi.org/10.1371/journal.pone.0341163.s003
(TIF)
S4 Table. The performance of the combination of model preprocessors and classifiers.
https://doi.org/10.1371/journal.pone.0341163.s004
(DOCX)
S5 Table. Radiomics features selected by LASSO and their coefficients.
https://doi.org/10.1371/journal.pone.0341163.s005
(DOCX)
S6 Table. MSI features of patients in cohorts.
https://doi.org/10.1371/journal.pone.0341163.s006
(DOCX)
S7 Table. MSI features selected by LASSO and their coefficients.
https://doi.org/10.1371/journal.pone.0341163.s007
(DOCX)
S8 Dataset. Anonymized raw clinical-radiological data.
This CSV file contains the de-identified dataset of 355 patients, with age reported in groups.
https://doi.org/10.1371/journal.pone.0341163.s008
(XLSX)
S10 Dataset. Anonymized raw radiomics features.
https://doi.org/10.1371/journal.pone.0341163.s010
(XLSX)
S11 Text. Data dictionary for S8 Dataset.
This text file explains all variable names, categorical codes, and the anonymization procedure.
https://doi.org/10.1371/journal.pone.0341163.s011
(TXT)
References
- 1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74(3):229–63. pmid:38572751
- 2. Borczuk AC. Updates in grading and invasion assessment in lung adenocarcinoma. Mod Pathol. 2022;35(Suppl 1):28–35. pmid:34615984
- 3. Kuhn E, Morbini P, Cancellieri A, Damiani S, Cavazza A, Comin CE. Adenocarcinoma classification: patterns and prognosis. Pathologica. 2018;110(1):5–11. pmid:30259909
- 4. Moreira AL, Ocampo PSS, Xia Y, Zhong H, Russell PA, Minami Y, et al. A grading system for invasive pulmonary adenocarcinoma: a proposal from the International association for the study of lung cancer pathology committee. J Thorac Oncol. 2020;15(10):1599–610. pmid:32562873
- 5. Hou L, Wang T, Chen D, She Y, Deng J, Yang M, et al. Prognostic and predictive value of the newly proposed grading system of invasive pulmonary adenocarcinoma in Chinese patients: a retrospective multicohort study. Mod Pathol. 2022;35(6):749–56. pmid:35013526
- 6. E H, Wu J, Ren Y, Xia L, Xu L, Li S, et al. The IASLC grading system for invasive pulmonary adenocarcinoma: a potential prognosticator for patients receiving neoadjuvant therapy. Ther Adv Med Oncol. 2023;15:17588359221148028. pmid:36643658
- 7. Zhang Y, Zhang Y, Hu Y, Zhang S, Zhu M, Hu B, et al. Validation of the novel International association for the study of lung cancer grading system and prognostic value of filigree micropapillary and discohesive growth pattern in invasive pulmonary adenocarcinoma. Lung Cancer. 2023;175:79–87. pmid:36481678
- 8. Deng C, Zheng Q, Zhang Y, Jin Y, Shen X, Nie X, et al. Validation of the novel international association for the study of lung cancer grading system for invasive pulmonary adenocarcinoma and association with common driver mutations. J Thorac Oncol. 2021;16(10):1684–93. pmid:34302987
- 9. Jeon YJ, Lee J, Shin S, Cho JH, Choi YS, Kim J, et al. Prognostic impact of micropapillary and solid histological subtype on patients undergoing curative resection for stage I lung adenocarcinoma according to the extent of pulmonary resection and lymph node assessment. Lung Cancer. 2022;168:21–9. pmid:35462111
- 10. Huang K-Y, Ko P-Z, Yao C-W, Hsu C-N, Fang H-Y, Tu C-Y, et al. Inaccuracy of lung adenocarcinoma subtyping using preoperative biopsy specimens. J Thorac Cardiovasc Surg. 2017;154(1):332-339.e1. pmid:28366548
- 11. Shang Y, Zeng Y, Luo S, Wang Y, Yao J, Li M, et al. Habitat imaging with tumoral and peritumoral radiomics for prediction of lung adenocarcinoma invasiveness on preoperative chest CT: a multicenter study. AJR Am J Roentgenol. 2024;223(4):e2431675. pmid:39140631
- 12. Bach PB, Mirkin JN, Oliver TK, Azzoli CG, Berry DA, Brawley OW, et al. Benefits and harms of CT screening for lung cancer: a systematic review. JAMA. 2012;307(22):2418–29. pmid:22610500
- 13. Fujikawa R, Muraoka Y, Kashima J, Yoshida Y, Ito K, Watanabe H, et al. Clinicopathologic and genotypic features of lung adenocarcinoma characterized by the international association for the study of lung cancer grading system. J Thorac Oncol. 2022;17(5):700–7. pmid:35227909
- 14. Li Y, Liu J, Yang X, Wang A, Zang C, Wang L, et al. An ordinal radiomic model to predict the differentiation grade of invasive non-mucinous pulmonary adenocarcinoma based on low-dose computed tomography in lung cancer screening. Eur Radiol. 2023;33(5):3072–82. pmid:36790469
- 15. Chen M, Copley SJ, Viola P, Lu H, Aboagye EO. Radiomics and artificial intelligence for precision medicine in lung cancer treatment. Semin Cancer Biol. 2023;93:97–113. pmid:37211292
- 16. Yang Z, Cai Y, Chen Y, Ai Z, Chen F, Wang H, et al. A CT-based radiomics nomogram combined with clinic-radiological characteristics for preoperative prediction of the novel IASLC grading of invasive pulmonary adenocarcinoma. Acad Radiol. 2023;30(9):1946–61. pmid:36567145
- 17. Zhu X, Dong D, Chen Z, Fang M, Zhang L, Song J, et al. Radiomic signature as a diagnostic factor for histologic subtype classification of non-small cell lung cancer. Eur Radiol. 2018;28(7):2772–8. pmid:29450713
- 18. Zhao B, Ju SH. Habitat imaging: exploring tumor heterogeneity. Chin J Radiol. 2025;59(4):355–7.
- 19. Zhang X, Su G-H, Chen Y, Gu Y-J, You C. Decoding intratumoral heterogeneity: clinical potential of habitat imaging based on radiomics. Radiology. 2023;309(3):e232047. pmid:38085080
- 20. O’Connor JPB, Rose CJ, Waterton JC, Carano RAD, Parker GJM, Jackson A. Imaging intratumor heterogeneity: role in therapy response, resistance, and clinical outcome. Clin Cancer Res. 2015;21(2):249–57. pmid:25421725
- 21. Zhang J, Sha J, Liu W, Zhou Y, Liu H, Zuo Z. Quantification of intratumoral heterogeneity: distinguishing histological subtypes in clinical T1 stage lung adenocarcinoma presenting as pure ground-glass nodules on computed tomography. Acad Radiol. 2024;31(10):4244–55. pmid:38627129
- 22. Zuo Z, Deng J, Ge W, Zhou Y, Liu H, Zhang W, et al. Quantifying intratumoral heterogeneity within sub-regions to predict high-grade patterns in clinical stage I solid lung adenocarcinoma. BMC Cancer. 2025;25(1):51. pmid:39789523
- 23. Zuo Z, Fan X, Zeng Y, Qi W, Liu W, Zhang J. Multiperspective tumor heterogeneity metrics for preoperative prediction of IASLC grading in clinical stage IA lung adenocarcinomas: a multicenter study. Comput Methods Programs Biomed. 2026;274:109137. pmid:41197251
- 24. Wu J, Cao G, Sun X, Lee J, Rubin DL, Napel S, et al. Intratumoral spatial heterogeneity at perfusion MR imaging predicts recurrence-free survival in locally advanced breast cancer treated with neoadjuvant chemotherapy. Radiology. 2018;288(1):26–35. pmid:29714680
- 25. Sujit SJ, Aminu M, Karpinets TV, Chen P, Saad MB, Salehjahromi M, et al. Enhancing NSCLC recurrence prediction with PET/CT habitat imaging, ctDNA, and integrative radiogenomics-blood insights. Nat Commun. 2024;15(1):3152. pmid:38605064
- 26.
Chow SC, Shao J, Wang H, Lokhnygina Y. Sample size calculations in clinical research. Chapman and Hall/CRC; 2017.
- 27. Chen L, Gu D, Chen Y, Shao Y, Cao X, Liu G, et al. An artificial-intelligence lung imaging analysis system (ALIAS) for population-based nodule computing in CT scans. Comput Med Imag Graph. 2021;89:101899. pmid:33761446
- 28. Hegedűs F, Zombori-Tóth N, Kiss S, Lantos T, Zombori T. Prognostic impact of the IASLC grading system of lung adenocarcinoma: a systematic review and meta-analysis. Histopathology. 2024;85(1):51–61. pmid:38485464
- 29. Sagreiya H. Finding the pieces to treat the whole: using radiomics to identify tumor habitats. Radiol Artif Intell. 2024;6(2):e230547. pmid:38416038
- 30. Zhu Y, Zheng D, Xu S, Chen J, Wen L, Zhang Z, et al. Intratumoral habitat radiomics based on magnetic resonance imaging for preoperative prediction treatment response to neoadjuvant chemotherapy in nasopharyngeal carcinoma. Jpn J Radiol. 2024;42(12):1413–24. pmid:39162780
- 31. Caii W, Wu X, Guo K, Chen Y, Shi Y, Chen J. Integration of deep learning and habitat radiomics for predicting the response to immunotherapy in NSCLC patients. Cancer Immunol Immunother. 2024;73(8):153. pmid:38833187
- 32. Yanagawa N, Shiono S, Abiko M, Katahira M, Osakabe M, Ogata S-Y. The clinical impact of solid and micropapillary patterns in resected lung adenocarcinoma. J Thorac Oncol. 2016;11(11):1976–83. pmid:27374456
- 33. Fogg KC, Renner CM, Christian H, Walker A, Marty-Santos L, Khan A, et al. Ovarian cells have increased proliferation in response to heparin-binding epidermal growth factor as collagen density increases. Tissue Eng Part A. 2020;26(13–14):747–58. pmid:32598229
- 34. Lee KH, Goo JM, Park SJ, Wi JY, Chung DH, Go H, et al. Correlation between the size of the solid component on thin-section CT and the invasive component on pathology in small lung adenocarcinomas manifesting as ground-glass nodules. J Thorac Oncol. 2014;9(1):74–82. pmid:24346095
- 35. Yoshizawa A, Motoi N, Riely GJ, Sima CS, Gerald WL, Kris MG, et al. Impact of proposed IASLC/ATS/ERS classification of lung adenocarcinoma: prognostic subgroups and implications for further revision of staging based on analysis of 514 stage I cases. Mod Pathol. 2011;24(5):653–64. pmid:21252858
- 36. Ahn B, Yoon S, Kim D, Chun S-M, Lee G, Kim H-R, et al. Clinicopathologic and genomic features of high-grade pattern and their subclasses in lung adenocarcinoma. Lung Cancer. 2022;170:176–84. pmid:35820357
- 37. Zheng S, Liu J, Xie J, Zhang W, Bian K, Liang J, et al. Differentiating high-grade patterns and predominant subtypes for IASLC grading in invasive pulmonary adenocarcinoma using radiomics and clinical-semantic features. Cancer Imaging. 2025;25(1):42. pmid:40155960