Super-resolution and habitat radiomics based computed tomography machine-learning model for prediction of lung invasive adenocarcinoma: a multi-centre study

Introduction

Lung cancer remains a leading cause of cancer-related mortality worldwide (1), with lung adenocarcinoma being the most prevalent histological subtype (2). Timely and accurate identification of invasive adenocarcinoma (IAC) is critical for optimizing treatment strategies (3), and ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) positron emission tomography-computed tomography (PET-CT) serves this goal by refining target-volume definition, detecting occult lesions, and guiding dose adjustments in locally advanced non-small-cell lung cancer (4). The 5-year overall survival rates highlight a striking disparity between non-IAC and IAC lesions, at 100% vs. 78.4% (P<0.001), underscoring the profound prognostic importance of this histological distinction (5). CT is indispensable for non-IAC evaluation of pulmonary nodules (6); however, conventional CT often suffers from limited resolution, hindering the detection of subtle features indicative of IAC (7). Super-resolution imaging has emerged as a valuable adjunct in medical imaging, effectively enhancing CT resolution and potentially revealing critical diagnostic details that are invisible on standard scans (8).

Radiomics, a technique for decoding tumor heterogeneity by extracting high-dimensional quantitative features from medical images (9), traditionally relies on coarse-grained texture analysis (10) that may overlook subtle micro-environmental variations critical for predicting tumor invasiveness (11,12). Habitat maps CT-derived clusters to specific histological subregions of IAC, including lepidic, invasive, and stromal components, representing a refined sub-regional radiomics approach that transcends unsupervised texture-based segregation to enhance the biological and clinical significance of radiomic feature extraction. A previous study integrating intra-tumoral and peri-tumoral habitat analysis have demonstrated that such models can effectively predict lung adenocarcinoma invasiveness, achieving area under the curve (AUC) of 0.947 in the validation set and 0.800–0.936 in the external testing sets (13). A critical unmet need in existing habitat radiomics research is the limitation of conventional CT resolution (1–2 mm) to capture microscale tumor heterogeneity predictive of IAC invasiveness. To address this, we integrated generative adversarial network (GAN) based super-resolution reconstruction (quadrupling through-plane resolution to 0.25 mm) with pathologically guided habitat radiomics, a combination never previously reported for IAC stratification. This framework enables capture of microstructural features (e.g., fine invasive foci, stromal-tumor interfaces) that are undetectable on conventional CT, directly linking imaging resolution to biological invasiveness. The integration of the Internet of Things (loT) has emerged as a transformative force in medical practice as highlighted by Mulita et al. [2022] (14) and aligns with the technological evolution driving innovations like our super resolution and habitat radiomics based CT machine learning model for predicting IAC which leverages advanced data integration and analytical capabilities similar to the core strengths of IoT in enhancing precision and accessibility in cancer care.

Here, we introduce an integrative machine-learning framework that combines super-resolution enhanced CT imaging with spatially resolved habitat radiomics signatures to predict IAC invasiveness. Our methodology not only enables precise lesion boundary characterization through resolution recovery but also reveals habitat-specific radiomics signatures that correlate with histopathologically validated invasive mechanisms. This approach holds promise for delivering a clinically actionable and interpretable platform for multi-scale risk profiling, thereby guiding personalized therapeutic decisions and optimizing longitudinal surveillance protocols. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1405/rc).

Methods

Ethical statement

This multi-cohort, retrospective investigation was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by institutional review boards of Zhejiang Provincial People’s Hospital (No. 2026 [009]); Jianyang People’s Hospital (No. JYL202335) and The First People’s Hospital of Tongxiang (No. 2025-003-01), and individual consent for this retrospective analysis was waived.

Study participants

The initial cohort included 1,130 individuals with sub-solid pulmonary nodules (SSNs) identified via Picture Archiving and Communication System screening across Zhejiang Provincial People’s Hospital, The First People’s Hospital of Tongxiang, and Jianyang People’s Hospital from January 2017 to December 2024. The inclusion criteria were as follows: (I) acquisition of preoperative thin-section CT scans prior to 14 days before surgical intervention; (II) availability of comprehensive clinical records, including demographic data, oncological history, tobacco exposure, and other relevant medical history; (III) CT images of diagnostic quality, devoid of motion artifacts or respiratory interference; (IV) maximum diameter of SSNs not exceeding 3.0 cm. Exclusion criteria consisted of: (I) prior administration of neoadjuvant therapies, such as radiation treatment or systemic chemotherapy; (II) documented primary malignancy with metastatic potential; (III) active infectious diseases or a history of specific infectious conditions.

Following rigorous screening, the final analytical cohort consisted of 858 cases (mean age 57.635±12.978 years, 562 females) from Zhejiang Provincial People’s Hospital, which was utilized as a training set for IAC prediction modeling. An independent testing set consisting of 272 cases (mean age: 57.037±11.683 years, 189 females) from The First People’s Hospital of Tongxiang and Jianyang People’s Hospital was designated for external performance assessment. Comprehensive clinical parameters were collected, encompassing demographic characteristics (age, gender), lesion morphology (diameter, location distribution across right upper/middle/lower and left upper/lower lobes), lifestyle factors (smoking and drinking), and medical history (hypertension, diabetes mellitus, and oncological background).

Image acquisition and super-resolution reconstruction

CT examinations were performed using 64- or 128-slice multi-detector CT systems from major manufacturers, namely Siemens Somatom Definition AS, Philips Incisive CT, GE Bright Speed, GE Revolution Maxima, and Canon Aquilion Prime. Standard thoracic imaging protocols were applied, capturing volumetric data from the pulmonary apex to the adrenal region during inspiratory breath-hold. In line with conventional thoracic imaging practices, patients were positioned supine with elevated upper extremities to reduce artifacts. Technical parameters were standardized at 120 kVp with automatic tube current modulation, and both slice thickness and reconstruction intervals were maintained at 1–2 mm.

Before super-resolution analysis, all imaging datasets underwent standardized preprocessing, including isotropic voxel resampling to 1 mm × 1 mm × 1 mm and application of uniform window settings [width: 1,500 Hounsfield units (HU); level: −600 HU] optimized for pulmonary parenchyma assessment. A hybrid approach combining a GAN and deep transfer learning network was utilized to quadruple the through-plane resolution, refining the voxel dimensions from the initial isotropic 1 mm × 1 mm × 1 mm to an optimized 1 mm × 1 mm × 0.25 mm configuration. GANs consist of two competitively trained neural networks: the generator produces synthetic images that mimic real data distributions, while the discriminator learns to differentiate between genuine and computer-generated images (15). The detailed schematic diagram is illustrated in Figure 1. The generator, an encoder-decoder architecture integrated with residual dense blocks, takes 1 mm × 1 mm × 1 mm CT voxels as input and outputs 1 mm × 1 mm × 0.25 mm voxels via sub-pixel convolution. In contrast, a five-layer convolutional discriminator classifies patches as real or fake. Both networks are trained using a combined loss function encompassing L1-pixel loss, Visual Geometry Group (VGG)-perceptual loss, and least-squares adversarial loss, while cycle-consistent feedback constrains downsampled super-resolved images to match the original thick-slice CT scans, thereby suppressing the generation of hallucinatory features.

Figure 1 The schematic diagram of GAN. BN, batch normalization; Conv, convolution; Deconv, deconvolution; GAN, generative adversarial network; HR, high resolution; LR, low resolution; ReLU, rectified linear unit; SR, super resolution.

Volume of interest (VOI) segmentation and sub-region clustering

Two radiologists with extensive expertise in CT imaging interpretation (10 and 15 years, respectively), manually segmented the intra-tumor VOIs on preprocessed CT datasets (Figure 2). Their initial annotations underwent a collaborative review process, during which iterative refinements were made to reconcile discrepancies and reach an expert consensus. All VOI segmentation tasks were carried out using ITK-SNAP software (Version 3.8.0), ensuring accurate volumetric delineation of the targeted anatomical structures.

Figure 2 The VOI segmentation of intra-tumor and sub-region. VOI, volume of interest.

Subsequently, the VOIs were computationally partitioned into sub-regions by integrating HU clusters with spatial heterogeneity metrics obtained from textural analysis. A 3×3×3 volumetric kernel was utilized for local feature extraction, followed by the application of K-means clustering to divide each VOI into three distinct subdomains (Figure 2). The Calinski-Harabasz index (CHI) was adopted as the key criterion for determining the optimal number of clusters (16), with systematic evaluation conducted across a range of 3 to 9 cluster configurations (Figure 3).

Figure 3 The CHI of 3 to 9 clusters. CHI, Calinski-Harabasz index.

Feature extraction and selection

Radiomics features were categorized into three primary domains, encompassing geometric attributes, intensity distributions, and texture. Geometric features characterized the three-dimensional (3D) shape of tumors, offering insights into their spatial structural organization. Intensity features quantified the first-order statistical distributions of voxels within lesions, facilitating the assessment of internal homogeneity and heterogeneity. Texture features, meanwhile, described the second- and higher-order spatial distribution of intensity values, including gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level dependence matrix (GLDM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). Using in-house feature analysis software implemented with PyRadiomics (http://pyradiomics.readthedocs.io), a total of 107 radiomics features were extracted from each intra-tumor region or sub-region. This comprehensive set included 18 first-order features, 14 shape features, 24 GLCM features, 16 GLRLM features, 14 GLDM features, 16 GLSZM features, and 5 NGTDM features.

Feature selection in the training set was executed through a five-step approache. First, radiomics features were normalized by Z-score (17). Second, the method of synthetic minority oversampling technique (SMOTE) was applied to balance the sample distribution between IAC and non-IAC groups (18). Third, statistical tests including the t-test, Mann-Whitney U test, or Chi-squared test (P<0.01) were employed to identify features with significant differences between the IAC and non-IAC groups. Fourth, correlation analysis was performed to address multi-collinearity; features with correlation coefficients ≤0.9 were excluded, retaining only those with optimal diagnostic utility. Finally, the maximal redundancy minimal relevance (mRMR) algorithm combined with least absolute shrinkage and selection operator (LASSO) regression were used to further eliminate redundant features. The optimal regularization parameter λ was determined via 5-fold cross-validation, based on minimum criteria (19).

Model construction and evaluation

After selecting the radiomics features from both intra-tumor and sub-regions, the optimal features identified in the training set were used to develop logistic regression machine-learning models, which were subsequently validated using the radiomics features from the testing set. Specifically, models customized for intra-tumor radiomics features (Intra-Model) and habitat radiomics features (Habitat-Model) were established for the prediction of IAC invasiveness. Subsequently, the prediction probability values of Intra-Model and Habitat-Model were integrated with significant clinical risk factors to develop a comprehensive prediction nomogram. The diagnostic efficacy of these models was systematically evaluated using the AUC with 95% confidence interval (CI) of receiver operating characteristics (ROC) using DeLong test for statistical comparison, along with multiple classification metrics including accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), determined at optimal thresholds derived from the Youden index criterion. The overall study design and workflow are illustrated in Figure 4.

Figure 4 The overall design and workflow of this study. CT, computed tomography; GAN, generative adversarial network; GGN, ground-glass nodules; IAC, invasive adenocarcinoma; LASSO, least absolute shrinkage and selection operator; mRMR, maximal redundancy minimal relevance; ROC, receiver operating characteristics; SMOTE, synthetic minority oversampling technique; VOI, volume of interest.

Statistical analysis

Statistical analyses were performed using software of IBM SPSS (Version 24.0.0), R software package (Version 3.4.1), and Python (Version 3.7.12). A two-tailed P value <0.05 was defined as the threshold for statistical significance. Continuous variables were presented as the mean ± standard deviation (SD), while categorical variables were presented as frequencies (percentages). The normality of continuous variables was evaluated using the Kolmogorov-Smirnov test. For parametric data, the t-test was applied, whereas the Mann-Whitney U test was utilized for non-parametric distributions. Categorical variables were analyzed using either Pearson’s Chi-squared test or Fisher’s exact test, depending on the expected cell frequencies.

Results

Patients’ population and clinical characteristics

The study cohort comprised 858 patients, with 501 assigned to the non-IAC group and 357 to the IAC group. All these patients were recruited from Center 1 and were used to construct the logistic machine-learning model. Subsequently, the model’s performance was validated using an external testing set of 272 patients, sourced from Centers 2 and 3. The detailed data of the training and testing sets are presented in Table 1. In the training set, several clinical factors demonstrated statistically significant differences, including age (P<0.001), gender (P<0.001), diameter (P<0.001), smoking status (P<0.001), alcohol consumption history (P=0.001), hypertension (P<0.001), and diabetes (P=0.027). Following this, a multi-variable logistic regression analysis was conducted, identifying age and diameter as the key clinical predictors. These factors were selected for an in-depth integrative study aimed at developing a nomogram that combined clinical, intra-tumor radiomics, and habitat radiomics features.

Performance of Intra-Model and Habitat-Model

Following feature selection, 24 intra-tumor radiomics features and 35 sub-region habitat radiomics features were retained for model construction. The Intra-Model achieved AUCs of 0.853 (95% CI: 0.828–0.878) in the training set and 0.875 (95% CI: 0.829–0.921) in the testing set. Conversely, the Habitat-Model demonstrated favorable discriminatory power (Figure 5), with AUCs of 0.893 (95% CI: 0.871–0.915) in the training set and 0.882 (95% CI: 0.833–0.930) in the testing set. Comprehensive efficacy metrics (Table 2), including accuracy, sensitivity, specificity, PPV, and NPV, were calculated to further evaluate model performance. Decision curve analysis (DCA) was employed to assess the clinical utility of the predictive models, providing insights into their potential net benefits across different threshold probabilities (Figure 6).

Figure 5 The comparison of AUCs in Intra-Model and Habitat-Model. AUC, area under the curve.

Figure 6 The comparison of DCAs between Intra-Model, Habitat-Model, and combined nomogram. DCA, decision curve analysis.

Construction of combined nomogram

The combined nomogram (Figure 7) for differentiating IAC from non-IAC was developed by integrating key clinical factors, namely age and tumor diameter, with the prediction probability values from both the Intra-Model and Habitat-Model. This integrative model boasted an AUC of 0.907 (95% CI: 0.887–0.927) in the training set and 0.895 (95% CI: 0.849–0.941) in the testing set, outperforming each individual counterpart (Table 2) and highlighting the synergy-driven predictive improvement. DCA results, depicted in Figure 6, further validated the nomogram’s clinical utility. Compared to single-modality models, the combined approach achieved a 28–34% net benefit improvement at critical risk thresholds (10–25%), solidifying its position as a more effective tool for clinical decision-making.

Figure 7 The combined nomogram including clinical factors, intra-tumor radiomics, and habitat radiomics.

Discussion

The GAN based super-resolution framework for images were adopted for analysis, as it is a well-validated method whose ability to enhance clinically critical CT details was established in You et al.’s landmark study (20), with additional support for the clinical value of high-resolution CT provided by Zhang et al.’s prospective research (21). Our study reveals that habitat radiomics analysis derived from super-resolution CT images outperforms traditional intra-tumor radiomics models in predicting IAC invasiveness. The Habitat-Model showed incremental discriminative performance compared to the Intra-Model (training AUC: 0.893 vs. 0.853; testing AUC: 0.882 vs. 0.875), and the combined nomogram demonstrated improved performance over individual modalities. This aligns with emerging research indicating that sub-region radiomics signatures, which quantify tumor micro-environment heterogeneity, can serve as invasion biomarkers (22). The 35 retained habitat features, especially those capturing shape features (23) and texture features of GLCM/GLSZM (24), likely reflect pathobiological and micro-environment processes such as angiogenesis-driven niche formation and stromal remodeling (25), which are hallmarks of invasive progression (26). Our findings build upon the work of Lee et al. (27), who introduced habitat radiomics for glioblastoma prognosis prediction and demonstrated its utility in identifying radio-resistant areas in soft tissue sarcoma for genomically adjusted radiation therapy (28). The Habitat-Model’s consistent performance (training and testing AUCs >0.880) suggests that super-resolution enhancement effectively addresses scanner-induced feature variability, a common limitation in conventional radiomics (29). Notably, in external validation, the model balanced sensitivity (0.86) and specificity (0.82) (Table 2), surpassing some deep learning methods that often prioritize high AUCs at the cost of clinical interpretability (30). Our DCA result in the testing set further underscored the clinical value of habitat radiomics. Compared to the Intra-Model, it offered greater net benefits at a 10% threshold probability, outperforming both “treat-all” and “treat-none” strategies. This highlights its potential to reduce unnecessary surgeries for indolent lesions while focusing on high-risk pulmonary nodules.

The integration of clinical factors (age and tumor diameter) with intra-tumor and habitat radiomics signature through a combined nomogram represents a significant advancement in the prediction of IAC, achieving exceptional discriminative performance (AUC: 0.904 training set, 0.895 testing set), outperforming standalone Intra-Model and Habitat-Model. This outcome aligns with the emerging trend towards multi-modal prediction frameworks (31), as the synergy between clinical data and radiomics signatures effectively overcomes the limitations of imaging-only analyses (32). At a 20% risk threshold, the testing set nomogram provided 24% higher net benefit than the Intra-Model and Habitat-Model alone, suggesting its potential to avoid approximately 20 unnecessary lobectomies per 100 patients while preserving diagnostic sensitivity. This finding resonates well with the Fleischer Society’s evidence-based recommendations, which advocate for composite models to optimize oncological surveillance and mitigate overtreatment risks in sub-solid nodule management (33). This tripartite integration echoes the “biological bridge” concept introduced by Balagurunathan et al. (34), where clinical and radiomics data collectively map tumor evolutionary trajectories. While the magnitude of AUC improvement (1.3–2.2% in the testing set) is modest, the nomogram’s enhanced net benefit (28–34%) at clinically relevant thresholds (10–25% risk) supports its utility in reducing overtreatment. In summary, this study first integrates GAN-driven super-resolution that quadruples through-plane resolution to 0.25 mm for pulmonary SSNs, histopathology-informed spatial habitat clustering that merges CT density with 3D heterogeneity metrics, and a three-input nomogram combining clinical data, conventional intra-tumor radiomics and super-resolution habitat features for IAC stratification.

Despite these advances, three limitations warrant attention. Firstly, the current combined nomogram does not incorporate emerging biomarkers, such as liquid biopsy markers. These markers have the potential to further enhance the precision of risk stratification for IAC, and their exclusion represents an area for future improvement. Secondly, the retrospective design of tumor data collection inherently introduces selection bias. To address this, future research should focus on two key aspects. Prospective validation of standardized measurement protocols is essential to ensure the generalizability of the findings. Additionally, exploring the development of dynamic nomograms that integrate longitudinal imaging biomarkers could provide a more comprehensive understanding of tumor progression, thereby improving the predictive capabilities of the model in real-world clinical settings. Thirdly, GAN based super-resolution was applied to all images to ensure consistent capture of microstructural features across the cohort, which prevents direct assessment of the incremental value of the super-resolution step itself. While our framework demonstrates the potential of high-resolution habitat radiomics for IAC stratification, the specific contribution of super-resolution beyond conventional thin-slice CT in extracting pathologically relevant micro-heterogeneity such as fine invasive foci and stromal-tumor interfaces remains unquantified.

Conclusions

In summary, the habitat radiomics model, leveraging spatially resolved information, surpasses traditional intra-tumor radiomics models in discriminative power for IAC. Building upon this advancement, a combined nomogram framework integrates intra-tumor radiomics, habitat radiomics, and key clinical biomarkers achieved favorable performance in IAC risk stratification. This integrated approach provides a robust tool for precision therapeutic decision-making in the management of pulmonary nodules.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1405/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1405/dss

Funding: This study was supported by the Medical and Health Research Projects of Health Commission of Zhejiang Province (Nos. 2025HY0135, 2022KY040 and 2023KY472), 2025 Medical Education Research Projects by the Medical Education Branch of the Chinese Medical Association and the National Center for Medical Education Development (No. 2025B83), and Zhejiang Provincial Natural Science Foundation of China (No. LTGY24H180017).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1405/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This multi-cohort, retrospective investigation was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by institutional review boards of Zhejiang Provincial People’s Hospital (No. 2026 [009]); Jianyang People’s Hospital (No. JYL202335) and The First People’s Hospital of Tongxiang (No. 2025-003-01), and individual consent for this retrospective analysis was waived.

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

References

1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Wadowska K, Błasiak P, Rzechonek A, Bil-Lula I, Śliwińska-Mossoń M. New Insights on Old Biomarkers Involved in Tumor Microenvironment Changes and Their Diagnostic Relevance in Non-Small Cell Lung Carcinoma. Biomolecules 2021;11:1208. [Crossref] [PubMed]
3. Wang X, Li Q, Cai J, Wang W, Xu P, Zhang Y, Fang Q, Fu C, Fan L, Xiao Y, Liu S. Predicting the invasiveness of lung adenocarcinomas appearing as ground-glass nodule on CT scan using multi-task learning and deep radiomics. Transl Lung Cancer Res 2020;9:1397-406. [Crossref] [PubMed]
4. Mulita A, Valsamaki P, Bekou E, Anevlavis S, Nanos C, Zisimopoulos A, Giatromanolaki A, Koukourakis MI. Benefits from (18)F-FDG PET-CT-Based Radiotherapy Planning in Stage III Non-Small-Cell Lung Cancer: A Prospective Single-Center Study. Cancers (Basel) 2025;17:1969. [Crossref] [PubMed]
5. Takahashi M, Shigematsu Y, Ohta M, Tokumasu H, Matsukura T, Hirai T. Tumor invasiveness as defined by the newly proposed IASLC/ATS/ERS classification has prognostic significance for pathologic stage IA lung adenocarcinoma and can be predicted by radiologic parameters. J Thorac Cardiovasc Surg 2014;147:54-9. [Crossref] [PubMed]
6. Ichinose J, Kawaguchi Y, Nakao M, Matsuura Y, Okumura S, Ninomiya H, Oikado K, Nishio M, Mun M. Utility of Maximum CT Value in Predicting the Invasiveness of Pure Ground-Glass Nodules. Clin Lung Cancer 2020;21:281-7. [Crossref] [PubMed]
7. Chen X, Yao B, Li J, Liang C, Qi R, Yu J. Feasibility of Using High-Resolution Computed Tomography Features for Invasiveness Differentiation of Malignant Nodules Manifesting as Ground-Glass Nodules. Can Respir J 2022;2022:2671772. [Crossref] [PubMed]
8. Nam JG, Kang SK, Choi H, Hong W, Park J, Goo JM, Lee JS, Park CM. Sixty-four-fold data reduction of chest radiographs using a super-resolution convolutional neural network. Br J Radiol 2024;97:632-9. [Crossref] [PubMed]
9. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
10. Chen Q, Sun H, Jiang Q, Wang X, Cao J, Li Q, Song J, Liu S, Zhu Q, Xiao Y. Preoperative CT-based artificial intelligence-derived quantitative parameters and imaging features for predicting the invasiveness of histologically confirmed subcentimeter adenocarcinomatous nodules: a two-center study. Quant Imaging Med Surg 2025;15:12593-606. [Crossref] [PubMed]
11. Caruso D, Polici M, Zerunian M, Pucciarelli F, Guido G, Polidori T, Landolfi F, Nicolai M, Lucertini E, Tarallo M, Bracci B, Nacci I, Rucci C, Iannicelli E, Laghi A. Radiomics in Oncology, Part 1: Technical Principles and Gastrointestinal Application in CT and MRI. Cancers (Basel) 2021;13:2522. [Crossref] [PubMed]
12. Gatenby RA, Grove O, Gillies RJ. Quantitative imaging in cancer evolution and ecology. Radiology 2013;269:8-15. [Crossref] [PubMed]
13. Shang Y, Zeng Y, Luo S, Wang Y, Yao J, Li M, Li X, Kui X, Wu H, Fan K, Li ZC, Zheng H, Li G, Liu J, Zhao W. 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:e2431675. [Crossref] [PubMed]
14. Mulita F, Verras GI, Anagnostopoulos CN, Kotis K. A Smarter Health through the Internet of Surgical Things. Sensors (Basel) 2022;22:4577. [Crossref] [PubMed]
15. Wolterink JM, Mukhopadhyay A, Leiner T, Vogl TJ, Bucher AM, Išgum I. Generative Adversarial Networks: A Primer for Radiologists. Radiographics 2021;41:840-57. [Crossref] [PubMed]
16. Youssef A, Moa B, El-Sharkawy YH. A novel visible and near-infrared hyperspectral imaging platform for automated breast-cancer detection. Photodiagnosis Photodyn Ther 2024;46:104048. [Crossref] [PubMed]
17. Fatania K, Frood R, Mistry H, Short SC, O'Connor J, Scarsbrook AF, Currie S. Impact of intensity standardisation and ComBat batch size on clinical-radiomic prognostic models performance in a multi-centre study of patients with glioblastoma. Eur Radiol 2025;35:3354-66. [Crossref] [PubMed]
18. Pahlevani M, Rajabi E, Taghavi M, VanBerkel P. Developing a decision support tool to predict delayed discharge from hospitals using machine learning. BMC Health Serv Res 2025;25:56. [Crossref] [PubMed]
19. Ye K, Zhang L, Zhou H, Mo X, Shi C. Machine learning-based radiomic features of perivascular adipose tissue in coronary computed tomography angiography predicting inflammation status around atherosclerotic plaque: a retrospective cohort study. Ann Med 2025;57:2431606. [Crossref] [PubMed]
20. You C, Li G, Zhang Y, Zhang X, Shan H, Li M, Ju S, Zhao Z, Zhang Z, Cong W, Vannier MW, Saha PK, Hoffman EA, Wang G CT. Super-Resolution GAN Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE). IEEE Trans Med Imaging 2020;39:188-203. [Crossref] [PubMed]
21. Zhang F, Gong L. Investigation on super-resolution reconstruction of lung CT images for COVID-19 based on sequential images. Biomedical Signal Processing and Control 2025;102:107424.
22. Aerts HJ. The Potential of Radiomic-Based Phenotyping in Precision Medicine: A Review. JAMA Oncol 2016;2:1636-42. [Crossref] [PubMed]
23. Yap FY, Varghese BA, Cen SY, Hwang DH, Lei X, Desai B, Lau C, Yang LL, Fullenkamp AJ, Hajian S, Rivas M, Gupta MN, Quinn BD, Aron M, Desai MM, Aron M, Oberai AA, Gill IS, Duddalwar VA. Shape and texture-based radiomics signature on CT effectively discriminates benign from malignant renal masses. Eur Radiol 2021;31:1011-21. [Crossref] [PubMed]
24. Sheen H, Cho W, Kim C, Han MC, Kim H, Lee H, Kim DW, Kim JS, Hong CS. Radiomics-based hybrid model for predicting radiation pneumonitis: A systematic review and meta-analysis. Phys Med 2024;123:103414. [Crossref] [PubMed]
25. Wang Y, Jia J, Wang F, Fang Y, Yang Y, Zhou Q, Yuan W, Gu X, Hu J, Yang S. Pre-metastatic niche: formation, characteristics and therapeutic implication. Signal Transduct Target Ther 2024;9:236. [Crossref] [PubMed]
26. Choi SW, Cho HH, Koo H, Cho KR, Nenning KH, Langs G, Furtner J, Baumann B, Woehrer A, Cho HJ, Sa JK, Kong DS, Seol HJ, Lee JI, Nam DH, Park H. Multi-Habitat Radiomics Unravels Distinct Phenotypic Subtypes of Glioblastoma with Clinical and Genomic Significance. Cancers (Basel) 2020;12:1707. [Crossref] [PubMed]
27. Lee J, Narang S, Martinez JJ, Rao G, Rao A. Associating spatial diversity features of radiologically defined tumor habitats with epidermal growth factor receptor driver status and 12-month survival in glioblastoma: methods and preliminary investigation. J Med Imaging (Bellingham) 2015;2:041006. [Crossref] [PubMed]
28. Naghavi AO, Bryant JM, Kim Y, Weygand J, Redler G, Sim AJ, et al. Habitat escalated adaptive therapy (HEAT): a phase 2 trial utilizing radiomic habitat-directed and genomic-adjusted radiation dose (GARD) optimization for high-grade soft tissue sarcoma. BMC Cancer 2024;24:437. [Crossref] [PubMed]
29. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, Sanduleanu S, Larue RTHM, Even AJG, Jochems A, van Wijk Y, Woodruff H, van Soest J, Lustberg T, Roelofs E, van Elmpt W, Dekker A, Mottaghy FM, Wildberger JE, Walsh S. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
30. Piskorski L, Debic M, von Stackelberg O, Schlamp K, Welzel L, Weinheimer O, Peters AA, Wielpütz MO, Frauenfelder T, Kauczor HU, Heußel CP, Kroschke J. Malignancy risk stratification for pulmonary nodules: comparing a deep learning approach to multiparametric statistical models in different disease groups. Eur Radiol 2025;35:3812-22. [Crossref] [PubMed]
31. Xue T, Zhu L, Tao Y, Ye X, Yu H. Development and validation of an interpretable delta radiomics-based model for predicting invasive ground-glass nodules in lung adenocarcinoma: a retrospective cohort study. Quant Imaging Med Surg 2024;14:4086-97. [Crossref] [PubMed]
32. Su Q, Wang B, Guo J, Nie P, Xu W. CT-based radiomics and clinical characteristics for predicting bone metastasis in lung adenocarcinoma patients. Transl Lung Cancer Res 2024;13:721-32. [Crossref] [PubMed]
33. MacMahon H, Naidich DP, Goo JM, Lee KS, Leung ANC, Mayo JR, Mehta AC, Ohno Y, Powell CA, Prokop M, Rubin GD, Schaefer-Prokop CM, Travis WD, Van Schil PE, Bankier AA. Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017. Radiology 2017;284:228-43. [Crossref] [PubMed]
34. Balagurunathan Y, Gu Y, Wang H, Kumar V, Grove O, Hawkins S, Kim J, Goldgof DB, Hall LO, Gatenby RA, Gillies RJ. Reproducibility and Prognosis of Quantitative Features Extracted from CT Images. Transl Oncol 2014;7:72-87. [Crossref] [PubMed]