Preoperative Prediction of Ki-67 in Stage T1 Lung Adenocarcinoma Based on 2.5D and Ensemble Integrated Models: A Multicenter Study

Preoperative Prediction of Ki-67 in Stage T1 Lung Adenocarcinoma Based on 2.5D and Ensemble Integrated Models: A Multicenter Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Preoperative Prediction of Ki-67 in Stage T1 Lung Adenocarcinoma Based on 2.5D and Ensemble Integrated Models: A Multicenter Study Xiuhua Peng, Xinchao Xu, Jingnan Xue, Pengliang Xu, Yuanbin Li, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8746230/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Objective: In order to set up a predictive model which combines 2.5D deep learning, 2D deep learning, 3D deep learning, and an ensemble fusion model, which could be used for the accurate prediction of preoperative expression levels of ki-67 in patients diagnosed with stage T1 invasive lung adenocarcinoma(LUAD ). Patients and Methods: In total, 503 patients from our own institution and 102 patients with invasive LUAD of T1 stage from two other centers were included retrospectively. The subjects were then divided into a Ki-67 high-expression group and a Ki-67 low-expression group, comprising 254 and 351 subjects, respectively. The subject set from our own institution formed the training set, and the subject sets from the other two centers formed the test set. Three categories of deep learning(DL) models, 2D, 2.5D, and 3D models, and an ensemble fusion model were developed and used to evaluate the performance of each model in predicting Ki-67 expressions in subjects with T1 LUAD. Results: All of these models showed excellent discriminative performance for the training set. The Ensemble model showed the best discriminative performance with an AUC of 0.991 (95% CI: 0.9858–0.9964). The Support Vector Machine 3D model (SVM 3D) showed an AUC of 0.970 (95% CI: 0.9538–0.9866). The Multilayer Perceptron 2D model (MLP 2D; AUC = 0.931; 95% CI: 0.9093–0.9529) and Gradient Boosting Machine 2.5D model (GBM 2.5D; AUC = 0.908; 95% CI: 0.8811–0.9342) also showed excellent discriminative performance. For all of these models, AUC performance declined as expected for the independent test set; however, some of these models showed different degrees of performance declines. The Ensemble model showed excellent robustness to performance declines and showed an AUC of 0.870 (95% CI: 0.7943–0.9447). The GBM 2.5D model showed excellent robustness to performance declines and showed an AUC of 0.843 (95% CI: 0.7405–0.9450); it showed the smallest performance declines of all of the models. The MLP 2D model showed moderate performance and showed an AUC of 0.810 (95% CI: 0.7110–0.9099). On the other hand, it was evident that the performance of the SVM 3D model showed substantial declines; it showed an AUC of 0.745 (95% CI: 0.6338–0.8566); it showed the greatest performance declines of all of the models and possibly suffered from. Conclusion: The Ensemble model demonstrated high performance in the prediction of Ki-67 expression in stage T1 invasive LUAD, which confirms the feasibility of the proposed approach in the prediction of Ki-67. Invasive lung adenocarcinoma deep learning Ensemble Ki-67 2.5D Introduction Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction " Lung adenocarcinoma(LUAD) serves as the leading classification of Non-Small Cell Lung Cancer (NSCLC), which comprises 80% of the global aggregate NSCLC case load" [1] . Stage T1 LUAD, where the size of the tumor does not exceed 3 cm in greatest dimension with no metastasis or in the lymph nodes, can be treated with resection [2, 3] . However, the high rate of recurrence following resection remains a major concern, with the proliferation activity of the cancer cells being the key predictor of the potential malignancy [4] . Ki-67 PI, the gold-standard biomarker that can be used as a predictor of cancer cell proliferation, is present only in the G1, S, G2, or M phase of the cell division process, directly reflecting the proportion of the cancerous cell population that is dividing. Ki-67 PI is closely related with the invasive potential, resistance to therapy, and survival outcome in LUAD. [5] . Studies have proven that the risk of postoperative recurrence is high in patients with stage T1 LUAD and high Ki-67 proliferative activity, suggesting that the patients have the potential to obtain clinical benefit from adjuvant chemotherapy or targeted therapy, while the risk is low in patients with low Ki-67 proliferative activity, allowing the patients to avoid the risk of overtreatment [6] . Though the postoperative pathological examination combined with immunohistochemistry is accurate in determining the status of Ki-67 proliferative activity, there is a considerable time lag in obtaining the test results, by which time the surgical method and initial treatment modalities are already applied and cannot be refined or improved [7] . Moreover, the conventional method of determining the status of Ki-67 by immunohistochemistry is also associated with considerable disadvantages, such as the presence of a high risk of pneumothorax and hemorrhage due to the tumor heterogeneity of the tumor, especially in the case of ground-glass nodules (GGNs) that are commonly seen in the case of T1 tumors, as well as the unavailability of the test in the case of abnormal bleeding or the unwillingness of the patient to undergo the test [8] . A critical imperative emerges to develop a reliable and accurate method to address the disadvantages of the conventional method of determining the status of Ki-67 by immunohistochemistry to improve the decision-making process by providing a reliable method to assess the status of the tumor before the commencement of the operation. However, in the past years, some improvements have been made using conventional radiomics techniques as well as single slice 2D deep learning models in the pre-operative prediction of the proliferative activity of Ki-67 in LUAD, with the AUC ranging from 0.77 to 0.83 [9-11] . Despite the improvements, the conventional radiomics techniques still present some limitations, as they fail to represent the heterogeneity of the tumor, while the single slice 2D models rely solely on the largest slice of the tumor, ignoring the rich 3D spatial structural information, which is important in the portrayal of the tumor behavior [12] . In order to overcome the above problems, 2.5D and 3D DL models were proposed and compared in the present study [13, 14] . In the 2.5D model, cross-dimensional information fusion is realized by fusing multi-plane features from the coronal, sagittal, and axial views. In other words, sampling bias is avoided [15, 16] . On the contrary, the three-dimensional (3D) model is based on the whole volume containing the tumor, so that the stereoscopic spatial distribution and heterogeneity of the tumor are fully preserved. This method is also more consistent with the real biological characteristics of the tumor [17] . Moreover, to make the most of the advantages of models with varying dimensions and architectures, the current study introduces an ensemble fusion model (Ensemble Model) [18, 19] . A "meta-learning" method is used to combine the outputs of the predictions of various heterogeneous models (such as GBM 2.5D, MLP 2D, and SVM 3D), so that the fusion of multi-scale features and the improvement of the robustness of decision-making are effectively achieved. A comprehensive comparison of the above models is made to achieve a more comprehensive, accurate, and effective technical support for the preoperative, non-invasive appraisal of the Ki-67 proliferation index of T1-stage LUAD from various perspectives. Materials and methods Patient characteristics This study’s protocol received ratification and registration from the Ethics Review Committee of The First Affiliated Hospital of Huzhou Normal University (The First People's Hospital of Huzhou). Due to the nature of the present study, a waiver of informed consent was granted after deliberation and review by the Ethics Review Committee. A retrospective study population of patients diagnosed with T1 stage invasive LUAD undergoing radical surgery at center1 between January 2019 and July 2025, as well as two collaborating medical centers between July 2024 and July 2025, was recruited. The technical framework of the present study is shown in Fig. 1 . Preoperative CT scan imaging data, as well as the clinicopathological information of all recruited patients, was systematically collected. Inclusion criteria:1) Pathologically proven invasive LUAD;2) Maximal tumor diameter ≤ 3 cm as measured by preoperative CT;3) Availability of preoperative CT imaging data and data acquisition within 1 month prior to surgery;4) No distant metastases as determined by preoperative clinical evaluation, which may have included the use of imaging and laboratory techniques. Exclusion Criteria:1) History of other malignant tumors;2) Provision of neoadjuvant therapy pre-surgically;3) Presence of several pulmonary nodules based on the results obtained in the patient's preoperative CT 4) Incomplete clinical or pathological data; 5) Pathological diagnosis of non-invasive lung adenocarcinoma (Fig. 2 ) In the current research, a total of 605 patients have been included. Among them, 503 eligible cases have been obtained from our hospital, including 201 with high levels of Ki-67 and 293 with low levels of Ki-67. A validation set consisting of 102 samples from two other institutions included 44 with high levels and 58 with low levels of Ki-67. The baseline clinical and demographic information used for the purpose of the analysis included the patients’ age, sex, location of the tumor, and size of the tumor in the form of the diameter of the lesions. There were also other pathological aspects of the tumors that were noted in the patients, including the invasion of the nerves, the invasion of the pleura, the involvement of the lymph nodes in the metastasis of the tumors, and the involvement of the airspaces in the metastasis of the tumors in the patients. Histopathological evaluation The post-operative pathological immunohistochemical examination findings were used as a gold standard to evaluate Ki-67 expression. The Ki-67 expression status of all enrolled cases was established. For IHC assays of all enrolled cases, tissue samples were immobilized in 10% neutral buffered formalin and processed for paraffin sectioning were subjected to IHC examination using Ki-67 monoclonal antibody. The IHC examination findings were assessed by two senior pathologists using a double-blind system. For microscopic examination of Ki-67 expression levels, regions of sections showing the maximum number of tumor cells with Ki-67 positivity were identified under a high-power field of ×400 magnification. No fewer than 1,000 cells were counted. By virtue of preceding studies [ 20 ] , cases with Ki-67 proliferative activity > 10% were classified as high proliferative activity cases; cases with Ki-67 proliferative activity ≤ 10% were classified as low proliferative activity cases. This classification of cases as low or high expression was used as a label to train the model. All detection procedures were performed under uniform laboratory conditions to exclude any inter-batch detection errors. CT Imaging Protocol. All registered study participants underwent chest spiral computed tomography (CT) imaging. CT protocols were site-specific: Center 1: The imaging technique used in this center consisted of a Siemens CT scan using a Somatom Definition AS and a Somatom Perspective scanner from Siemens. The scanning parameters used were a tube voltage of 120 kV and a tube current of 40 mAs, a pitch of 0.758, a collimation of 0.6 x 64 mm, a slice thickness of 1 mm, and a 512 x 512 matrix size. Center 2: The imaging modality was the GE Revolution APEX, from GE Healthcare, operating at the following conditions: tube voltage: 120 kV; tube current: 40 mAs; pitch: 0.758; collimation: 0.6 × 16 mm; slice: 0.6/1 mm; matrix: 512 × Center 3: The scanner used in this center was a GE Optima 540 scanner from GE Healthcare. The settings of this scanner were similar to those of Center 2: tube voltage of 120 kV and tube current of 40 mAs; also similar were the settings of the scanner's pitch and collimation at 0.758 and 0.6 x 16 mm respectively; similar also Conventional radiomics ROI segmentation and feature extraction. Feature Extraction, Selection and Modeling ROI Image Segmentation and Feature Extraction In the image preprocessing phase, all images are resampled and transformed into a standardized voxel size of 1×1×1 mm. Finally, the Z-score normalization is used as a technique of standardizing the images. The process of ROI segmentation was conducted independently by two experienced radiologists under blinded conditions to avoid bias: An attending physician specializing in diagnosis, manually segmented the ROI slice by slice using an open-source tool called ITK-SNAP Version 3.8.0. An Associate Chief Physician specializing in diagnosis, checked and validated all the manually segmented ROI images generated by Radiologist A on a per-case basis.Moreover, the definition of ROI varied among CNN models of different dimensionality. In the 2D CNN model, the ROI included the cross-section where the maximum cross-sectional area of the tumor was found. In the 2.5D CNN model, the ROI included 3D information of the tumor in the coronal, sagittal, and axial views. In the 3D CNN model, the ROI included a bounding box that enclosed the whole volume of the tumor, thus covering all the components of the tumor. The regions of interest of the tumor were delineated by each of the radiologists individually using fixed lung window settings that are defined as follows: window level = − 450 HU and window width = 1500 HU. Before providing input to the network, the original images were resampled to have a uniform voxel size of 1 mm x 1 mm x 1 mm. The original images of the tumor were normalized to have units of Hounsfield Units (HU). This was performed by using slope and intercept parameters obtained from the DICOM headers of each of the images. A threshold was also applied to remove interference due to extreme value effects. The mean and variance of each individual 3D tumor image within the training cohort were computed, and all images were normalized via the Z-score method. Additional training images were created by translating the bounding box of each of the images by several voxel sizes in different directions to account for variability during the delineation process. In the context of deep learning-based feature extraction, the ResNet-50 model was used as the basic CNN model to avoid gradient vanishing problems when deep networks were being trained. After the pre-processing of the input images, the images were inputted to the model to conduct the initial feature extraction through the 7x7 convolution layer and the max pooling layer of size 3x3, as well as the hierarchical extraction of deep features of different scales through the four residual units. Global pooling was conducted to reduce the dimensionality of the features, while the feature mapping was conducted through the fully connected layer to achieve the classification/prediction results according to the research purposes. Feature Selection and Model Construction Through the above process of feature extraction, the 2D, 2.5D, and 3D sets of DL features were obtained separately. Prior to the process of feature selection, the extracted feature variables inside the training cohort were standardized with the intention of normalizing the features with differing dimensional scales while minimizing the disparities that are dimensional in nature. A three-step process was then followed with the intention of filtering the extracted feature variables inside the training cohort: 1) Preliminary feature screening through the Mann-Whitney U test was undertaken, retaining radiomic features with a statistically significant difference, i.e., P < 0.05. For highly repetitive features after the screening, the inter-feature correlation was also calculated using the Spearman rank correlation coefficient. Redundant features, as indicated by an inter-feature correlation coefficient greater than 0.9, were removed using a greedy recursive elimination approach, where the feature with the greatest redundancy was removed at each iteration. 2) A LASSO regression model was developed on the exploratory dataset and used to develop a radiomic signature. The optimal value of the regularization parameter, denoted by λ and corresponding to minimum cross-validation error, was computed using 10-fold cross-validation. The non-zero coefficient features were selected and then used to develop the regression model. 3) A radiomics score (RS) for each patient was calculated as a weighted linear combination of retained features based on their corresponding model coefficients. Based on the results of the feature selection, the 2D, 2.5D, and 3D sets of DL features were developed. Using the Onekey tool, multiple machine learning-based prediction models were developed for each of the sets of features using the training data. To improve the performance of the models, the stability of the predictions, and the ability of the models to generalize, the study used the ensemble strategy of bagging to combine three sets of deep learning-based models with imaging features of different dimensions—2D, 2.5D, and 3D. The ability of the models to generalize was evaluated. Statistical Methods For the statistical analysis, the SPSS Statistics 27.0 package and Onekey were used. The ROC curve was deployed to gauge the operational performance of each distinct model, and the area beneath the curve was instituted to ascertain the predictive capacity of the models. The DeLong test was harnessed to appraise the functional efficacy of the individual models. The calibration curves was employed to verify the calibration validity of the individual models, and the decision curve analysis was enlisted to evaluate the clinical applicability of the individual models. Where the resultant data yielded a statistical significance level of less than 0.05, the findings were deemed statistically meaningful. Results Baseline characteristics of the patients Table 1 demonstrates the clinical characteristics of the patients included in the two independent data sets. For the intergroup comparisons, the characteristics included age, gender, lesion location, and lesion size. The findings revealed that there were no significant differences in the characteristics of the patients included in the training set andthe external test set. Table 1 Patients' clinicopathological findings. Feature name Label all Label train Label test Label train P value Age 63.65 ± 9.95 63.72 ± 9.88 63.28 ± 10.30 63.72 ± 9.88 0.848 Diameter 1.66 ± 0.62 1.62 ± 0.61 1.86 ± 0.63 1.62 ± 0.61 < 0.001 Gender 0.305 Male 325(53.72) 265(52.68) 60(58.82) 265(52.68) Female 280(46.28) 238(47.32) 42(41.18) 238(47.32) Location 0.194 RUL 185(30.58) 161(32.01) 24(23.53) 161(32.01) RML 63(10.41) 54(10.74) 9(8.82) 54(10.74) RLL 104(17.19) 80(15.90) 24(23.53) 80(15.90) LUL 172(28.43) 139(27.63) 33(32.35) 139(27.63) LLL 81(13.39) 69(13.72) 12(11.76) 69(13.72) Nerve Invasion 0.075 NO 596(98.51) 498(99.01) 98(96.08) 498(99.01) Yes 9(1.49) 5(0.99) 4(3.92) 5(0.99) Pleural Invasion 0.054 NO 530(87.60) 447(88.87) 83(81.37) 447(88.87) Yes 75(12.40) 56(11.13) 19(18.63) 56(11.13) Lymph Node Metastasis 0.076 NO 571(94.38) 479(95.23) 92(90.20) 479(95.23) Yes 34(5.62) 24(4.77) 10(9.80) 24(4.77) STAS 0.177 NO 428(70.74) 362(71.97) 66(64.71) 362(71.97) Yes 177(29.26) 141(28.03) 36(35.29) 141(28.03) Data were presented as mean ± SD, or n(%) unless otherwise stated. RUL, right upper lobe;RML, right middle lobe; RLL, right lower lobe; LUL, left upper lobe; LLL, left lower lobe. Construction and Performance Validation of Three Deep Learning Models and Fusion Models Analysis of the ROC curve results obtained from the training and test datasets revealed that all the models had high discriminability within the training dataset. Among all the models, the ensemble model had the highest AUC at 0.991 (95% CI: 0.986–0.996), followed by the SVM 3D model with an AUC of 0.970 (95% CI: 0.954–0.987). In the test dataset, the AUC values obtained by all the models decreased as expected. However, the rate of decrease in the AUC values differed. Among all the models, the gradient boosting machine 2.5-dimensional model had the most robust results with an AUC of 0.843 (95% CI: 0.740–0.945). Among all the models, the ensemble model had the highest AUC at 0.870 (95% CI: 0.794–0.945). However, the SVM 3D model, which had high discriminability within the training dataset, had a very low AUC of 0.745 (95% CI: 0.634–0.857). Such a result revealed a high level of over-fitting. The results obtained by the multilayer perceptron two-dimensional model had a moderate AUC of 0.810 (95% CI: 0.711–0.910). The results obtained by the model had a high level of stability as revealed by the results presented in Table 2 . Table 2 Comparison of predictive performance of different models for Ki-67 in T1-stage LUAD Signature Accuracy AUC 95%CI Sensitivity Specificity PPV NPV Cohort GBM 2.5D 0.835 0.908 0.8811–0.9342 0.833 0.837 0.787 0.873 Train MLP 2D 0.857 0.931 0.9093–0.9529 0.900 0.826 0.790 0.919 Train SVM 3D 0.938 0.970 0.9538–0.9866 0.957 0.924 0.901 0.967 Train Ensemble 0.948 0.991 0.9858–0.9964 0.976 0.927 0.907 0.982 Train GBM 2.5D 0.787 0.843 0.7405–0.9450 0.821 0.769 0.657 0.889 Test MLP 2D 0.787 0.810 0.7110–0.9099 0.821 0.769 0.657 0.889 Test SVM 3D 0.725 0.745 0.6338–0.8566 0.643 0.769 0.600 0.800 Test Ensemble 0.800 0.870 0.7943–0.9447 0.929 0.731 0.650 0.950 Test Significant differences in the accuracy of the predicted probability were observed in the test set based on the calibration curve analysis performed on the models. The ensemble model (Ensemble) had a calibration curve that was closer to the ideal line than the other models, indicating its potential in providing accurate risk probability values. A similar observation was noted in the gradient boosting machine 2.5-dimensional model (GBM 2.5D). The curve was closer to the ideal line with only a few fluctuations in the localized region of the curve. Conversely, the support vector machine three-dimensional model (SVM 3D) had a poor calibration curve with a number of deviations from the ideal line in the low-to-moderate predicted probability region, indicating a tendency to overestimate the predicted probability in this region, as the predicted probability was substantially larger than the observed proportions in the test set. A degree of overestimation was also observed in the case of the multilayer perceptron two-dimensional model in the moderate-to-high predicted probability region. The overestimation in the case of the SVM 3D was confirmed based on the AUC value drop in the test set compared to the train set; this indicates a considerable degree of overfitting in the case of the SVM 3D model (Fig. 3 b,e). Appraisal of clinical applicability via decision curve analysis ascertained that within the testing cohort, all models exhibited superior performance with respect to net benefit relative to the “treat-all” and “treat-none” strategies approaches in the entire spectrum of the principal decision threshold probabilities, ranging from 0.1 to 0.6. However, the proposed ensemble model’s net benefit was consistently the best in the entire spectrum, surpassing that of the other models, especially in the clinically common threshold probability spectrum of 0.2–0.5, where its advantage was maximized in clinical decision-making contexts. The proposed gradient boosting machine 2.5-dimensional model’s net benefit was also high, ranking second after the proposed ensemble model in the entire spectrum of threshold probabilities, while the proposed support vector machine three-dimensional model’s net benefit curve revealed a noticeable decrease in its clinical utility, especially at high threshold probabilities (> 0.4), where its clinical utility was limited, while the proposed multilayer perceptron two-dimensional model’s net benefit curve revealed a mid-range performance between the best-performing proposed models, as shown in Fig. 3 (c, f). On the three aspects of evaluation, the discriminative performance of the ensemble model (AUC), the probability calibration accuracy, and the clinical decision net benefit, the ensemble model was ranked the best among the three models. The ensemble model still enjoys the best generalization ability with relatively accurate risk probability results, showing the highest net benefit for clinical decision-making to aid individualized patient care for those with low to moderate risk of the disease. Figure 3 shows the results of the evaluation of the three models. Figure 4 delineates the outcomes yielded by the DeLong test pertaining to the training and validation cohorts of the three models. Figure 5 showcases the visualization outputs corresponding to the three deep learning models. Discussion In recent years, considerable potential for preoperative prediction of the Ki-67 index in LUAD was demonstrated through the application of both radiomics and DL techniques [ 21 – 23 ] . Nevertheless, most previous studies have focused on single-dimensional imaging data and/or adopted a conventional feature fusion method, which is subject to considerable limitations in terms of robustness and clinical utility [ 24 ] . In the current investigation, three different imaging modeling methods, namely [ 12 , 14 , 25 ] 2D, 2.5D, and 3D, have been evaluated and an ensemble learning-facilitated approach for feature integration is also presented to enhance the performance and clinical interpretability and decision utility of the model. The results show that the proposed ensemble model has superior performance compared to other models in all performance metric criteria [ 19 , 26 ] . The superior performance of the proposed model may be due to a number of reasons, which include the following: first, the proposed model combines a number of base models with different architectures as well as feature dimensions, thus creating a heterogeneous ensemble model that has the ability to extract tumor-related features from different perspectives, hence reducing errors in prediction that are associated with structural biases as well as feature limitations that are associated with each of the models [ 27 , 28 ] . Second, the proposed model has the advantage of creating an ensemble that has a regularizing effect, thus having the ability to prevent the occurrence of overtraining, which is a common problem that occurs in complex models when dealing with a limited number of training samples.The proposed model has the advantage of adapting the outputs from each of the constituent models during the integration process, thus having the ability to improve the reliability of the final outputs that are produced during the prediction process. In the comparative evaluation of the performance of the models on an individual level, the proposed model achieved superior performance in terms of generalization capability as well as calibration performance [ 14 , 29 ] . It is reasonable to assume that the performance of the proposed model can be related to the mechanism of feature construction employed by the model. The proposed model utilizes the mechanism of the 2.5D model, where features from the coronal, sagittal, and axial planes of the tumor volume are simultaneously constructed. Hence, the model is able to take advantage of the high computational efficiency as well as the high architectural stability of the 2D models while incorporating the advantages of the 3D model as well [ 30 – 32 ] . Compared to the deep-learning-based models, the proposed model utilizes fewer parameters as well as lesser data requirements, thus being more suitable to the current clinical environment where the sample sizes of the images obtained from the patients are low. Unlike the traditional models of the 2D type, the proposed model is able to more comprehensively deal with the spatial distribution characteristics of the tumor, thus preventing prediction bias due to information loss [ 33 , 34 ] . The study not only examines the discriminative ability of the models but also their applicability to real-world settings [ 35 ] . DCA reveals the ensemble model’s net benefit to be superior to the “treat all” and “treat none” strategies across the range of decision thresholds, confirming the utility of the ensemble model for decision-making. The calibration curves reveal the correspondence of the observed outcomes with the predicted results of the ensemble model, suggesting the utility of the risk probability results for risk stratification of the patient population, thereby increasing the usability of the results obtained with the ensemble model [ 36 , 37 ] . Limitations of This Study This research incorporated a set of inherent limitations.First, although data were obtained from various centers, the total sample size was limited, and this could potentially impact the consistency in the assessment of the performance of the models. Second, the retrospective study design was potentially limited to selection biases; hence, further validation in large numbers in a prospective study design would be desirable to confirm the generalizability and clinical applicability of the model. The model was based only on imaging characteristics and did not account for clinical information, pathomics, or genomic information. It would therefore be desirable to explore the construction of a multimodal fusion model to enrich the predictive system. Conclusion In summary, this study assessed and compared various dimensions of imaging modeling strategies and established an ensemble learning-based fusion prediction model. This model had excellent performance in preoperative and noninvasive prediction of Ki-67 proliferation indices of T1 lung adenocarcinomas and outperformed other individual models in discriminability and had exceptional clinical decision net benefit and probability. Declarations Ethics Approval and Consent to Participate This multicenter retrospective study was conducted in strict accordance with the Declaration of Helsinki (World Medical Association, 2013 revision) and the Measures for the Ethical Review of Biomedical Research Involving Humans issued by the National Health Commission of the People's Republic of China. The study protocol was reviewed and formally approved by the Medical Research and Clinical Trial Ethics Committee of The First Affiliated Hospital of Huzhou Normal University (also known as Huzhou First People's Hospital), with the ethics approval number: 2024KYLL085-01. Funding: This work was supported by grants from the Science and Technology Project of Huzhou City, Zhejiang Province(2024GY41) and the Health Science and Technology Program(2025KY1555). Author Contribution X.P. (Xiuhua Peng)：Study design, data collection and analysis, manuscript drafting, funding acquisitionX.X. (Xinchao Xu)：Imaging data processing, statistical analysis.J.X. (Jingnan Xue)：Case data collation, figure and table preparation.P.X. (Pengliang Xu)：Methodological guidance, result validation. funding acquisitionY.L. (Yuanbin Li)：Literature search, initial manuscript revision.M.Y. (Mei Yang)：Ethics application, patient informed consent acquisition.L.Q. (Luying Qi)：Data verification, supplementary material organization.J.Z. (Jiacheng Zhao)：Imaging technical support.F.X. (Fenglin Xiang)：Study conception, protocol optimization.Y.S. (Yanyan Shen)：Manuscript polishing, language proofreading.S.F. (Shenyun Fang)：Data entry, result visualization.H.Z. (Hongxing Zhao)：Academic supervision.J.J. (Jianping Jiang, corresponding author)：Study oversight, final manuscript review and submission. References Chen H, Liu L, Zhang M, et al. Correlation of LOXL2 expression in non-small cell lung cancer with immunotherapy[J]. Int J Clin Exp Pathol. 2024;17(9):268–86. Dong H, Li Y, Zhao L, et al. Consensus clustering based on CT radiomics has potential for risk stratification of patients with clinical T1 stage lung adenocarcinoma[J]. BMC Med Imaging. 2025;25(1):231. Shi W, Hu Y, Wang H, et al. The value of relative CT attenuation in predicting invasiveness in patients with T1-stage lung adenocarcinoma[J]. Quant imaging Med Surg. 2025;15(10):9294–308. Herbst RS, Morgensztern D, Boshoff C. The biology and management of non-small cell lung cancer[J]. Nature. 2018;553(7689):446–54. Warth A, Cortis J, Soltermann A, et al. Tumour cell proliferation (Ki-67) in non-small cell lung cancer: a critical reappraisal of its prognostic role[J]. Br J Cancer. 2014;111(6):1222–9. Liu Z, Feng H, Ma S, et al. Clinicopathological characteristics of peripheral clinical stage IA lung adenocarcinoma with high Ki-67 expression[J]. Translational cancer Res. 2021;10(1):152–61. Gomà C, Mollà M, Oses G, et al. Preoperative Radiation Therapy-Induced Molecular and Immune Modulation in Early-Stage Breast Cancer: Results From the YOUNGSTER trial[J]. Int J Radiat Oncol Biol Phys. 2026;124(1):122–32. Wang X, Zhang H, Zhang X. Analysis of Influencing Factors of Complications for CT-guided ?Percutaneous Lung Biopsy[J]. Zhongguo fei ai za zhi = Chinese. J lung cancer. 2024;27(3):179–86. Zhou B, Xu J, Tian Y, et al. Correlation between radiomic features based on contrast-enhanced computed tomography images and Ki-67 proliferation index in lung cancer: A preliminary study[J]. Thorac cancer. 2018;9(10):1235–40. Gu Q, Feng Z, Liang Q, et al. Machine learning-based radiomics strategy for prediction of cell proliferation in non-small cell lung cancer[J]. Eur J Radiol. 2019;118:32–7. Truhn D, Schrading S, Haarburger C, et al. Radiomic versus Convolutional Neural Networks Analysis for Classification of Contrast-enhancing Lesions at Multiparametric Breast MRI[J]. Radiology. 2019;290(2):290–7. Yang G, Nie P, Zhao L, et al. 2D and 3D texture analysis to predict lymphovascular invasion in lung adenocarcinoma[J]. Eur J Radiol. 2020;129:109111. Huang T, Ke M, Liu Q, et al. A computed tomography-based deep learning model for non-invasively predicting World Health Organization (WHO)/International Society of Urological Pathology (ISUP) pathological grades of clear cell renal cell carcinoma (ccRCC): a multicenter cohort study[J]. Translational Androl Urol. 2025;14(7):2018–28. Kim S, Choi M, Han S, et al. Comparison of 2D, 2.5D, and 3D landmark localization networks for 3D cephalometry in CT images[J]. BMC Oral Health. 2025;25(1):1843. Wang W, Ren M, Ren J, et al. Predicting Radiation Pneumonitis Integrating Clinical Information, Medical Text, and 2.5D Deep Learning Features in Lung Cancer[J]. Int J Radiat Oncol Biol Phys. 2026;124(1):194–205. Cheng X, Li H, Li C et al. Attention-based deep learning network for predicting World Health Organization meningioma grade and Ki-67 expression based on magnetic resonance imaging[J]. Eur Radiol, 2025:10–1007. Park GE, Kim SH, Nam Y, et al. 3D Breast Cancer Segmentation in DCE-MRI Using Deep Learning With Weak Annotation[J]. J Magn Reson imaging: JMRI. 2024;59(6):2252–62. Gunasekaran H, Ramalakshmi K, Swaminathan DK et al. GIT-Net: An Ensemble Deep Learning-Based GI Tract Classification of Endoscopic Images[J]. Bioengineering (Basel, Switzerland), 2023,10(7):809. Sui C, Chen K, Ding E, et al. (18)F-FDG PET/CT-based intratumoral and peritumoral radiomics combining ensemble learning for prognosis prediction in hepatocellular carcinoma: a multi-center study[J]. BMC Cancer. 2025;25(1):300. Zhu M, Yang Z, Zhao W, et al. Predicting Ki-67 labeling index level in early-stage lung adenocarcinomas manifesting as ground-glass opacity nodules using intra-nodular and peri-nodular radiomic features[J]. Cancer Med. 2022;11(21):3982–92. Song C, Chen J, Zhao C, et al. Prediction of Ki-67 Expression in HIV-Associated Lung Adenocarcinoma Patients Using Multiple Machine Learning Models Based on CT Imaging Radiomics[J]. Cancer Manage Res. 2025;17:881–92. Zhu M, Yang Z, Zhao W, et al. Predicting Ki-67 labeling index level in early-stage lung adenocarcinomas manifesting as ground-glass opacity nodules using intra-nodular and peri-nodular radiomic features[J]. Cancer Med. 2022;11(21):3982–92. Zhu C, Liang Y, Li Y, et al. Prediction of the Ki-67 proliferation index in lung adenocarcinoma using an interpretable CT-based deep learning radiomics model: a two-center study[J]. BMC Pulm Med. 2025;25(1):536. Alabi RO, Elmusrati M, Leivo I, et al. Artificial Intelligence-Driven Radiomics in Head and Neck Cancer: Current Status and Future Prospects[J]. Int J Med Informatics. 2024;188:105464. Li M, Gu H, Xue T, et al. CT-based radiomics nomogram for the pre-operative prediction of lymphovascular invasion in colorectal cancer: a multicenter study[J]. Br J Radiol. 2023;96(1141):20220568. Tanveer MA, Khan MJ, Sajid H, et al. Convolutional neural networks ensemble model for neonatal seizure detection[J]. J Neurosci Methods. 2021;358:109197. Li Y, Chou H, Lin P, et al. A novel deep learning-based algorithm combining histopathological features with tissue areas to predict colorectal cancer survival from whole-slide images[J]. J translational Med. 2023;21(1):731. Eckhart L, Lenhof K, Rolli L, et al. A comprehensive benchmarking of machine learning algorithms and dimensionality reduction methods for drug sensitivity prediction[J]. Brief Bioinform. 2024;25(4):bbae242. Yoganathan SA, Paul SN, Paloor S, et al. Automatic segmentation of magnetic resonance images for high-dose-rate cervical cancer brachytherapy using deep learning[J]. Med Phys. 2022;49(3):1571–84. Rosas-Gonzalez S, Birgui-Sekou T, Hidane M, et al. Asymmetric Ensemble of Asymmetric U-Net Models for Brain Tumor Segmentation With Uncertainty Estimation[J]. Front Neurol. 2021;12:609646. Huang C, Li E, Hu J, et al. Enabling Early Identification of Malignant Vertebral Compression Fractures Through 2.5D Convolutional Neural Network Model With CT Image Analysis[J]. Spine. 2025;50(24):1728–36. Kumar A, Jiang H, Imran M, et al. A flexible 2.5D medical image segmentation approach with in-slice and cross-slice attention[J]. Comput Biol Med. 2024;182:109173. Karimzadeh M, Seyedarabi H, Jodeiri A, et al. Enhanced Brain Stroke Lesion Segmentation in MRI Using a 2.5D Transformer Backbone U-Net Model[J]. Brain Sci. 2025;15(8):778. Gao F, Hu Z, Xian J, et al. Mixed U-Net: Segmentation of focal liver lesions using a hybrid 2D and 3D model[J]. J Appl Clin Med Phys. 2026;27(1):e70408. Alonso S, Bra AI, Loredo M, et al. Performance of Disease Activity Indices Used in Axial Spondyloarthritis in Real-World Clinical Settings[J]. J Rhuematol. 2025;52(5):444–9. Vetsch T, Huber M, Wuethrich PY, et al. Preoperative prediction of severe short-term complications in patients with bladder cancer undergoing radical cystectomy[J]. Surg Oncol. 2025;61:102253. Chen S, Deng T, Yang Q, et al. Development and validation of an explainable machine learning model for predicting postoperative pulmonary complications after lung cancer surgery: a machine learning study[J]. EClinicalMedicine. 2025;86:103386. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 20 Apr, 2026 Reviews received at journal 13 Apr, 2026 Reviewers agreed at journal 08 Apr, 2026 Reviewers agreed at journal 07 Apr, 2026 Reviews received at journal 01 Apr, 2026 Reviewers agreed at journal 26 Mar, 2026 Reviewers invited by journal 09 Mar, 2026 Editor assigned by journal 05 Feb, 2026 Submission checks completed at journal 05 Feb, 2026 First submitted to journal 30 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8746230","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":602877610,"identity":"4a8f5a07-4288-4bff-b897-bb47cef6cdf6","order_by":0,"name":"Xiuhua Peng","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Xiuhua","middleName":"","lastName":"Peng","suffix":""},{"id":602877611,"identity":"d37d04a9-e742-457d-8452-94d8bdd64996","order_by":1,"name":"Xinchao Xu","email":"","orcid":"","institution":"The Sixth Affiliated Hospital of South China University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Xinchao","middleName":"","lastName":"Xu","suffix":""},{"id":602877612,"identity":"165ad8f0-23d1-4064-a936-d40df9d5a4fc","order_by":2,"name":"Jingnan Xue","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Jingnan","middleName":"","lastName":"Xue","suffix":""},{"id":602877613,"identity":"0efc40cf-5ce3-4a46-81b3-69d28aa0beda","order_by":3,"name":"Pengliang Xu","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Pengliang","middleName":"","lastName":"Xu","suffix":""},{"id":602877614,"identity":"8d682aa8-b05f-4299-a4c2-4de006e3f495","order_by":4,"name":"Yuanbin Li","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Yuanbin","middleName":"","lastName":"Li","suffix":""},{"id":602877615,"identity":"e8bb4aaa-7b61-4936-acdf-616acd39d553","order_by":5,"name":"Mei Yang","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Mei","middleName":"","lastName":"Yang","suffix":""},{"id":602877617,"identity":"f0bc20bc-5411-4eb5-8657-6056936a3389","order_by":6,"name":"Luying Qi","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Luying","middleName":"","lastName":"Qi","suffix":""},{"id":602877618,"identity":"4c9fef22-ca50-4db5-9630-6fbd463fee05","order_by":7,"name":"Jiacheng Zhao","email":"","orcid":"","institution":"Tokyo Denki University","correspondingAuthor":false,"prefix":"","firstName":"Jiacheng","middleName":"","lastName":"Zhao","suffix":""},{"id":602877620,"identity":"db541808-e5a6-4fd4-b4b2-03f28d1dd906","order_by":8,"name":"Fenglin Xiang","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Fenglin","middleName":"","lastName":"Xiang","suffix":""},{"id":602877621,"identity":"93017a9c-76e8-475b-abd7-6cf33a2b7636","order_by":9,"name":"Yanyan Shen","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Yanyan","middleName":"","lastName":"Shen","suffix":""},{"id":602877622,"identity":"a614f324-db25-4d8f-84ef-69091e56ca9d","order_by":10,"name":"Shenyun Fang","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Shenyun","middleName":"","lastName":"Fang","suffix":""},{"id":602877623,"identity":"25ec2327-fdb0-4025-88f3-5a5b03e5fc45","order_by":11,"name":"Hongxing Zhao","email":"","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":false,"prefix":"","firstName":"Hongxing","middleName":"","lastName":"Zhao","suffix":""},{"id":602877624,"identity":"9274a37d-9c28-4bf1-bc44-e0a57f6e38dd","order_by":12,"name":"Jianping Jiang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA2UlEQVRIie3RsQrCMBCA4ZRCXQ7r2CLoEwgHgiCKDr5IRNClg6ObiqviqtCHEATnC4FOuncVwalg3B2sm06Jm2D+OR+54xiz2X4wnxEnhe3purAgMxLO6Sw246GzXSbcjKAUFwlKOrs0QsPJEskloHSRonuasU6lNtMIJ59HxDjyQjrtmzEb1BukIW5w5JRhC4pidSgDo/5BR7zqTRGgGzAJVzOSP+ICsIelBDwzErzIBoc8XHr1ZowGu3SJBko92tyvXs5pNulUtOTzSzA9zRv5VthsNttf9ATwD01Zw34xPAAAAABJRU5ErkJggg==","orcid":"","institution":"The First Affiliated Hospital of Huzhou University (The First People's Hospital of Huzhou)","correspondingAuthor":true,"prefix":"","firstName":"Jianping","middleName":"","lastName":"Jiang","suffix":""}],"badges":[],"createdAt":"2026-01-31 04:08:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8746230/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8746230/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104557398,"identity":"a4a8dbc5-d063-4b79-a653-9ad4fb58edbf","added_by":"auto","created_at":"2026-03-13 09:27:32","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":401050,"visible":true,"origin":"","legend":"\u003cp\u003eRadiomics analysis implementation pathway.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8746230/v1/c3de9a651882cc69144289cb.png"},{"id":104557119,"identity":"cce52af1-7f0d-4262-a3a3-99825053213d","added_by":"auto","created_at":"2026-03-13 09:26:23","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":210087,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of patient selection.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8746230/v1/6ee9b9119e94eb46959f920e.png"},{"id":104557153,"identity":"c7fdff85-0d8c-4a44-96d9-796be8db95fb","added_by":"auto","created_at":"2026-03-13 09:26:34","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":262366,"visible":true,"origin":"","legend":"\u003cp\u003edepicts in detail the corresponding subgraphs of performance assessment outcomes for all models in the following configuration: ROC curves are illustrated for the (a) training cohort and (d) testing cohort; calibration curves are delineated for the (b) training cohort and (e) testing cohort; and DCA curves are showcased for the (c) training cohort and (f) testing cohort.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8746230/v1/e1c516425f1d31c9fa380045.png"},{"id":104557219,"identity":"ff0ffa75-e960-4217-89c8-90d4b60bc364","added_by":"auto","created_at":"2026-03-13 09:27:00","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":84440,"visible":true,"origin":"","legend":"\u003cp\u003eDeLong test results for the training set (a) and test set (b)\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8746230/v1/c45c864a41c0cd9c1751b43d.png"},{"id":104557171,"identity":"f5abdcab-6b8e-4257-933d-997b4a2fca18","added_by":"auto","created_at":"2026-03-13 09:26:38","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":191025,"visible":true,"origin":"","legend":"\u003cp\u003eGradient-Weighted Class Activation Mapping (Grad-CAM) plots of the three DL models.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8746230/v1/7de09ae7d7fe55e5bc11ba16.png"},{"id":104557557,"identity":"24f366e3-924b-4656-93df-f38005e0b844","added_by":"auto","created_at":"2026-03-13 09:28:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1844255,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8746230/v1/6bbd62de-db89-438d-aecd-506a3d31f396.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Preoperative Prediction of Ki-67 in Stage T1 Lung Adenocarcinoma Based on 2.5D and Ensemble Integrated Models: A Multicenter Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003e\" Lung adenocarcinoma(LUAD) serves as the leading classification of Non-Small Cell Lung Cancer (NSCLC), which comprises 80% of the global aggregate NSCLC case load\" \u0026nbsp;\u003csup\u003e[1]\u003c/sup\u003e. Stage T1 LUAD, where the size of the tumor does not exceed 3 cm in greatest dimension with no metastasis or in the lymph nodes, can be treated with resection\u003csup\u003e[2, 3]\u003c/sup\u003e. However, the high rate of recurrence following resection remains a major concern, with the proliferation activity of the cancer cells being the key predictor of the potential malignancy\u003csup\u003e[4]\u003c/sup\u003e . Ki-67 PI, the gold-standard biomarker that can be used as a predictor of cancer cell proliferation, is present only in the G1, S, G2, or M phase of the cell division process, directly reflecting the proportion of the cancerous cell population that is dividing. Ki-67 PI is closely related with the invasive potential, resistance to therapy, and survival outcome in LUAD.\u003csup\u003e[5]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eStudies have proven that the risk of postoperative recurrence is high in patients with stage T1 LUAD and high Ki-67 proliferative activity, suggesting that the patients have the potential to obtain clinical benefit from adjuvant chemotherapy or targeted therapy, while the risk is low in patients with low Ki-67 proliferative activity, allowing the patients to avoid the risk of overtreatment\u0026nbsp;\u003csup\u003e[6]\u003c/sup\u003e. Though the postoperative pathological examination combined with immunohistochemistry is accurate in determining the status of Ki-67 proliferative activity, there is a considerable time lag in obtaining the test results, by which time the surgical method and initial treatment modalities are already applied and cannot be refined or improved\u0026nbsp;\u003csup\u003e[7]\u003c/sup\u003e. Moreover, the conventional method of determining the status of Ki-67 by immunohistochemistry is also associated with considerable disadvantages, such as the presence of a high risk of pneumothorax and hemorrhage due to the tumor heterogeneity of the tumor, especially in the case of ground-glass nodules (GGNs) that are commonly seen in the case of T1 tumors, as well as the unavailability of the test in the case of abnormal bleeding or the unwillingness of the patient to undergo the test\u0026nbsp;\u003csup\u003e[8]\u003c/sup\u003e. A critical imperative emerges to develop a reliable and accurate method to address the disadvantages of the conventional method of determining the status of Ki-67 by immunohistochemistry to improve the decision-making process by providing a reliable method to assess the status of the tumor before the commencement of the operation.\u003c/p\u003e\n\u003cp\u003eHowever, in the past years, some improvements have been made using conventional radiomics techniques as well as single slice 2D deep learning models in the pre-operative prediction of the proliferative activity of Ki-67 in LUAD, with the AUC ranging from 0.77 to 0.83\u003csup\u003e[9-11]\u003c/sup\u003e. Despite the improvements, the conventional radiomics techniques still present some limitations, as they fail to represent the heterogeneity of the tumor, while the single slice 2D models rely solely on the largest slice of the tumor, ignoring the rich 3D spatial structural information, which is important in the portrayal of the tumor behavior\u003csup\u003e[12]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn order to overcome the above problems, 2.5D and 3D DL models were proposed and compared in the present study\u003csup\u003e[13, 14]\u003c/sup\u003e. In the 2.5D model, cross-dimensional information fusion is realized by fusing multi-plane features from the coronal, sagittal, and axial views. In other words, sampling bias is avoided\u003csup\u003e[15, 16]\u003c/sup\u003e. On the contrary, the three-dimensional (3D) model is based on the whole volume containing the tumor, so that the stereoscopic spatial distribution and heterogeneity of the tumor are fully preserved. This method is also more consistent with the real biological characteristics of the tumor\u003csup\u003e[17]\u003c/sup\u003e. Moreover, to make the most of the advantages of models with varying dimensions and architectures, the current study introduces an ensemble fusion model (Ensemble Model)\u0026nbsp;\u003csup\u003e[18, 19]\u003c/sup\u003e. A \"meta-learning\" method is used to combine the outputs of the predictions of various heterogeneous models (such as GBM 2.5D, MLP 2D, and SVM 3D), so that the fusion of multi-scale features and the improvement of the robustness of decision-making are effectively achieved. A comprehensive comparison of the above models is made to achieve a more comprehensive, accurate, and effective technical support for the preoperative, non-invasive appraisal of the Ki-67 proliferation index of T1-stage LUAD from various perspectives.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e \u003c/p\u003e\n\u003ch3\u003ePatient characteristics\u003c/h3\u003e\n\u003cp\u003e This study\u0026rsquo;s protocol received ratification and registration from the Ethics Review Committee of The First Affiliated Hospital of Huzhou Normal University (The First People's Hospital of Huzhou). Due to the nature of the present study, a waiver of informed consent was granted after deliberation and review by the Ethics Review Committee. A retrospective study population of patients diagnosed with T1 stage invasive LUAD undergoing radical surgery at center1 between January 2019 and July 2025, as well as two collaborating medical centers between July 2024 and July 2025, was recruited. The technical framework of the present study is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Preoperative CT scan imaging data, as well as the clinicopathological information of all recruited patients, was systematically collected.\u003c/p\u003e \u003cp\u003eInclusion criteria:1) Pathologically proven invasive LUAD;2) Maximal tumor diameter\u0026thinsp;\u0026le;\u0026thinsp;3 cm as measured by preoperative CT;3) Availability of preoperative CT imaging data and data acquisition within 1 month prior to surgery;4) No distant metastases as determined by preoperative clinical evaluation, which may have included the use of imaging and laboratory techniques.\u003c/p\u003e \u003cp\u003eExclusion Criteria:1) History of other malignant tumors;2) Provision of neoadjuvant therapy pre-surgically;3) Presence of several pulmonary nodules based on the results obtained in the patient's preoperative CT 4) Incomplete clinical or pathological data; 5) Pathological diagnosis of non-invasive lung adenocarcinoma (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eIn the current research, a total of 605 patients have been included. Among them, 503 eligible cases have been obtained from our hospital, including 201 with high levels of Ki-67 and 293 with low levels of Ki-67. A validation set consisting of 102 samples from two other institutions included 44 with high levels and 58 with low levels of Ki-67.\u003c/p\u003e \u003cp\u003eThe baseline clinical and demographic information used for the purpose of the analysis included the patients\u0026rsquo; age, sex, location of the tumor, and size of the tumor in the form of the diameter of the lesions. There were also other pathological aspects of the tumors that were noted in the patients, including the invasion of the nerves, the invasion of the pleura, the involvement of the lymph nodes in the metastasis of the tumors, and the involvement of the airspaces in the metastasis of the tumors in the patients.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eHistopathological evaluation\u003c/h2\u003e \u003cp\u003eThe post-operative pathological immunohistochemical examination findings were used as a gold standard to evaluate Ki-67 expression. The Ki-67 expression status of all enrolled cases was established. For IHC assays of all enrolled cases, tissue samples were immobilized in 10% neutral buffered formalin and processed for paraffin sectioning were subjected to IHC examination using Ki-67 monoclonal antibody. The IHC examination findings were assessed by two senior pathologists using a double-blind system. For microscopic examination of Ki-67 expression levels, regions of sections showing the maximum number of tumor cells with Ki-67 positivity were identified under a high-power field of \u0026times;400 magnification. No fewer than 1,000 cells were counted. By virtue of preceding studies\u003csup\u003e[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/sup\u003e, cases with Ki-67 proliferative activity\u0026thinsp;\u0026gt;\u0026thinsp;10% were classified as high proliferative activity cases; cases with Ki-67 proliferative activity\u0026thinsp;\u0026le;\u0026thinsp;10% were classified as low proliferative activity cases. This classification of cases as low or high expression was used as a label to train the model. All detection procedures were performed under uniform laboratory conditions to exclude any inter-batch detection errors.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCT Imaging Protocol.\u003c/b\u003e \u003c/p\u003e \u003cp\u003eAll registered study participants underwent chest spiral computed tomography (CT) imaging. CT protocols were site-specific:\u003c/p\u003e \u003cp\u003eCenter 1: The imaging technique used in this center consisted of a Siemens CT scan using a Somatom Definition AS and a Somatom Perspective scanner from Siemens. The scanning parameters used were a tube voltage of 120 kV and a tube current of 40 mAs, a pitch of 0.758, a collimation of 0.6 x 64 mm, a slice thickness of 1 mm, and a 512 x 512 matrix size.\u003c/p\u003e \u003cp\u003eCenter 2: The imaging modality was the GE Revolution APEX, from GE Healthcare, operating at the following conditions: tube voltage: 120 kV; tube current: 40 mAs; pitch: 0.758; collimation: 0.6 \u0026times; 16 mm; slice: 0.6/1 mm; matrix: 512 \u0026times;\u003c/p\u003e \u003cp\u003eCenter 3: The scanner used in this center was a GE Optima 540 scanner from GE Healthcare. The settings of this scanner were similar to those of Center 2: tube voltage of 120 kV and tube current of 40 mAs; also similar were the settings of the scanner's pitch and collimation at 0.758 and 0.6 x 16 mm respectively; similar also Conventional radiomics ROI segmentation and feature extraction.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eFeature Extraction, Selection and Modeling\u003c/h3\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eROI Image Segmentation and Feature Extraction\u003c/h2\u003e \u003cp\u003eIn the image preprocessing phase, all images are resampled and transformed into a standardized voxel size of 1\u0026times;1\u0026times;1 mm. Finally, the Z-score normalization is used as a technique of standardizing the images.\u003c/p\u003e \u003cp\u003eThe process of ROI segmentation was conducted independently by two experienced radiologists under blinded conditions to avoid bias: An attending physician specializing in diagnosis, manually segmented the ROI slice by slice using an open-source tool called ITK-SNAP Version 3.8.0. An Associate Chief Physician specializing in diagnosis, checked and validated all the manually segmented ROI images generated by Radiologist A on a per-case basis.Moreover, the definition of ROI varied among CNN models of different dimensionality. In the 2D CNN model, the ROI included the cross-section where the maximum cross-sectional area of the tumor was found. In the 2.5D CNN model, the ROI included 3D information of the tumor in the coronal, sagittal, and axial views. In the 3D CNN model, the ROI included a bounding box that enclosed the whole volume of the tumor, thus covering all the components of the tumor.\u003c/p\u003e \u003cp\u003eThe regions of interest of the tumor were delineated by each of the radiologists individually using fixed lung window settings that are defined as follows: window level\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;450 HU and window width\u0026thinsp;=\u0026thinsp;1500 HU. Before providing input to the network, the original images were resampled to have a uniform voxel size of 1 mm x 1 mm x 1 mm. The original images of the tumor were normalized to have units of Hounsfield Units (HU). This was performed by using slope and intercept parameters obtained from the DICOM headers of each of the images. A threshold was also applied to remove interference due to extreme value effects. The mean and variance of each individual 3D tumor image within the training cohort were computed, and all images were normalized via the Z-score method. Additional training images were created by translating the bounding box of each of the images by several voxel sizes in different directions to account for variability during the delineation process.\u003c/p\u003e \u003cp\u003eIn the context of deep learning-based feature extraction, the ResNet-50 model was used as the basic CNN model to avoid gradient vanishing problems when deep networks were being trained. After the pre-processing of the input images, the images were inputted to the model to conduct the initial feature extraction through the 7x7 convolution layer and the max pooling layer of size 3x3, as well as the hierarchical extraction of deep features of different scales through the four residual units. Global pooling was conducted to reduce the dimensionality of the features, while the feature mapping was conducted through the fully connected layer to achieve the classification/prediction results according to the research purposes.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eFeature Selection and Model Construction\u003c/h3\u003e\n\u003cp\u003eThrough the above process of feature extraction, the 2D, 2.5D, and 3D sets of DL features were obtained separately. Prior to the process of feature selection, the extracted feature variables inside the training cohort were standardized with the intention of normalizing the features with differing dimensional scales while minimizing the disparities that are dimensional in nature. A three-step process was then followed with the intention of filtering the extracted feature variables inside the training cohort:\u003c/p\u003e \u003cp\u003e1) Preliminary feature screening through the Mann-Whitney U test was undertaken, retaining radiomic features with a statistically significant difference, i.e., P\u0026thinsp;\u0026lt;\u0026thinsp;0.05. For highly repetitive features after the screening, the inter-feature correlation was also calculated using the Spearman rank correlation coefficient. Redundant features, as indicated by an inter-feature correlation coefficient greater than 0.9, were removed using a greedy recursive elimination approach, where the feature with the greatest redundancy was removed at each iteration.\u003c/p\u003e \u003cp\u003e2) A LASSO regression model was developed on the exploratory dataset and used to develop a radiomic signature. The optimal value of the regularization parameter, denoted by λ and corresponding to minimum cross-validation error, was computed using 10-fold cross-validation. The non-zero coefficient features were selected and then used to develop the regression model.\u003c/p\u003e \u003cp\u003e3) A radiomics score (RS) for each patient was calculated as a weighted linear combination of retained features based on their corresponding model coefficients.\u003c/p\u003e \u003cp\u003eBased on the results of the feature selection, the 2D, 2.5D, and 3D sets of DL features were developed. Using the Onekey tool, multiple machine learning-based prediction models were developed for each of the sets of features using the training data. To improve the performance of the models, the stability of the predictions, and the ability of the models to generalize, the study used the ensemble strategy of bagging to combine three sets of deep learning-based models with imaging features of different dimensions\u0026mdash;2D, 2.5D, and 3D. The ability of the models to generalize was evaluated.\u003c/p\u003e\n\u003ch3\u003eStatistical Methods\u003c/h3\u003e\n\u003cp\u003eFor the statistical analysis, the SPSS Statistics 27.0 package and Onekey were used. The ROC curve was deployed to gauge the operational performance of each distinct model, and the area beneath the curve was instituted to ascertain the predictive capacity of the models. The DeLong test was harnessed to appraise the functional efficacy of the individual models. The calibration curves was employed to verify the calibration validity of the individual models, and the decision curve analysis was enlisted to evaluate the clinical applicability of the individual models. Where the resultant data yielded a statistical significance level of less than 0.05, the findings were deemed statistically meaningful.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eBaseline characteristics of the patients\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e demonstrates the clinical characteristics of the patients included in the two independent data sets. For the intergroup comparisons, the characteristics included age, gender, lesion location, and lesion size. The findings revealed that there were no significant differences in the characteristics of the patients included in the training set andthe external test set.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatients' clinicopathological findings.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLabel all\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLabel train\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLabel test\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLabel train\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAge\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e63.65\u0026thinsp;\u0026plusmn;\u0026thinsp;9.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e63.72\u0026thinsp;\u0026plusmn;\u0026thinsp;9.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e63.28\u0026thinsp;\u0026plusmn;\u0026thinsp;10.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e63.72\u0026thinsp;\u0026plusmn;\u0026thinsp;9.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.848\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDiameter\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.66\u0026thinsp;\u0026plusmn;\u0026thinsp;0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.62\u0026thinsp;\u0026plusmn;\u0026thinsp;0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.86\u0026thinsp;\u0026plusmn;\u0026thinsp;0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.62\u0026thinsp;\u0026plusmn;\u0026thinsp;0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGender\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.305\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMale\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e325(53.72)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e265(52.68)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e60(58.82)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e265(52.68)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFemale\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e280(46.28)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e238(47.32)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e42(41.18)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e238(47.32)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLocation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.194\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRUL\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e185(30.58)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e161(32.01)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24(23.53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e161(32.01)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRML\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e63(10.41)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e54(10.74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9(8.82)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e54(10.74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRLL\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e104(17.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e80(15.90)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24(23.53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e80(15.90)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLUL\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e172(28.43)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e139(27.63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e33(32.35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e139(27.63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLLL\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e81(13.39)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e69(13.72)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12(11.76)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e69(13.72)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNerve Invasion\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.075\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNO\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e596(98.51)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e498(99.01)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e98(96.08)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e498(99.01)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eYes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9(1.49)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5(0.99)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4(3.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5(0.99)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePleural Invasion\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.054\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNO\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e530(87.60)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e447(88.87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e83(81.37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e447(88.87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eYes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e75(12.40)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e56(11.13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e19(18.63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e56(11.13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLymph Node Metastasis\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.076\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNO\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e571(94.38)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e479(95.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e92(90.20)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e479(95.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eYes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e34(5.62)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24(4.77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10(9.80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e24(4.77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSTAS\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.177\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNO\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e428(70.74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e362(71.97)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e66(64.71)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e362(71.97)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eYes\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e177(29.26)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e141(28.03)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e36(35.29)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e141(28.03)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eData were presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD, or n(%) unless otherwise stated. RUL, right upper lobe;RML, right middle lobe; RLL, right lower lobe; LUL, left upper lobe; LLL, left lower lobe.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eConstruction and Performance Validation of Three Deep Learning Models and Fusion Models\u003c/h3\u003e\n\u003cp\u003eAnalysis of the ROC curve results obtained from the training and test datasets revealed that all the models had high discriminability within the training dataset. Among all the models, the ensemble model had the highest AUC at 0.991 (95% CI: 0.986\u0026ndash;0.996), followed by the SVM 3D model with an AUC of 0.970 (95% CI: 0.954\u0026ndash;0.987). In the test dataset, the AUC values obtained by all the models decreased as expected. However, the rate of decrease in the AUC values differed. Among all the models, the gradient boosting machine 2.5-dimensional model had the most robust results with an AUC of 0.843 (95% CI: 0.740\u0026ndash;0.945). Among all the models, the ensemble model had the highest AUC at 0.870 (95% CI: 0.794\u0026ndash;0.945). However, the SVM 3D model, which had high discriminability within the training dataset, had a very low AUC of 0.745 (95% CI: 0.634\u0026ndash;0.857). Such a result revealed a high level of over-fitting. The results obtained by the multilayer perceptron two-dimensional model had a moderate AUC of 0.810 (95% CI: 0.711\u0026ndash;0.910). The results obtained by the model had a high level of stability as revealed by the results presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of predictive performance of different models for Ki-67 in T1-stage LUAD\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSignature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95%CI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eCohort\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGBM 2.5D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.835\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.908\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8811\u0026ndash;0.9342\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.837\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.787\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.873\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTrain\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMLP 2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.857\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.931\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9093\u0026ndash;0.9529\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.900\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.826\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.790\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.919\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTrain\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM 3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.938\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.970\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9538\u0026ndash;0.9866\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.957\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.924\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.901\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.967\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTrain\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnsemble\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.948\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.991\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9858\u0026ndash;0.9964\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.976\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.927\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.907\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.982\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTrain\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGBM 2.5D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.787\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.843\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7405\u0026ndash;0.9450\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.657\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.889\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMLP 2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.787\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.810\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7110\u0026ndash;0.9099\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.657\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.889\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM 3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.725\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.745\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6338\u0026ndash;0.8566\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.643\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.600\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnsemble\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.870\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7943\u0026ndash;0.9447\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.929\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.731\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.650\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.950\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eSignificant differences in the accuracy of the predicted probability were observed in the test set based on the calibration curve analysis performed on the models. The ensemble model (Ensemble) had a calibration curve that was closer to the ideal line than the other models, indicating its potential in providing accurate risk probability values. A similar observation was noted in the gradient boosting machine 2.5-dimensional model (GBM 2.5D). The curve was closer to the ideal line with only a few fluctuations in the localized region of the curve. Conversely, the support vector machine three-dimensional model (SVM 3D) had a poor calibration curve with a number of deviations from the ideal line in the low-to-moderate predicted probability region, indicating a tendency to overestimate the predicted probability in this region, as the predicted probability was substantially larger than the observed proportions in the test set. A degree of overestimation was also observed in the case of the multilayer perceptron two-dimensional model in the moderate-to-high predicted probability region. The overestimation in the case of the SVM 3D was confirmed based on the AUC value drop in the test set compared to the train set; this indicates a considerable degree of overfitting in the case of the SVM 3D model (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb,e).\u003c/p\u003e \u003cp\u003eAppraisal of clinical applicability via decision curve analysis ascertained that within the testing cohort, all models exhibited superior performance with respect to net benefit relative to the \u0026ldquo;treat-all\u0026rdquo; and \u0026ldquo;treat-none\u0026rdquo; strategies approaches in the entire spectrum of the principal decision threshold probabilities, ranging from 0.1 to 0.6. However, the proposed ensemble model\u0026rsquo;s net benefit was consistently the best in the entire spectrum, surpassing that of the other models, especially in the clinically common threshold probability spectrum of 0.2\u0026ndash;0.5, where its advantage was maximized in clinical decision-making contexts.\u003c/p\u003e \u003cp\u003eThe proposed gradient boosting machine 2.5-dimensional model\u0026rsquo;s net benefit was also high, ranking second after the proposed ensemble model in the entire spectrum of threshold probabilities, while the proposed support vector machine three-dimensional model\u0026rsquo;s net benefit curve revealed a noticeable decrease in its clinical utility, especially at high threshold probabilities (\u0026gt;\u0026thinsp;0.4), where its clinical utility was limited, while the proposed multilayer perceptron two-dimensional model\u0026rsquo;s net benefit curve revealed a mid-range performance between the best-performing proposed models, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(c, f).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eOn the three aspects of evaluation, the discriminative performance of the ensemble model (AUC), the probability calibration accuracy, and the clinical decision net benefit, the ensemble model was ranked the best among the three models. The ensemble model still enjoys the best generalization ability with relatively accurate risk probability results, showing the highest net benefit for clinical decision-making to aid individualized patient care for those with low to moderate risk of the disease. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the results of the evaluation of the three models.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e delineates the outcomes yielded by the DeLong test pertaining to the training and validation cohorts of the three models. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e showcases the visualization outputs corresponding to the three deep learning models.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn recent years, considerable potential for preoperative prediction of the Ki-67 index in LUAD was demonstrated through the application of both radiomics and DL techniques \u003csup\u003e[\u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/sup\u003e. Nevertheless, most previous studies have focused on single-dimensional imaging data and/or adopted a conventional feature fusion method, which is subject to considerable limitations in terms of robustness and clinical utility\u003csup\u003e[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e. In the current investigation, three different imaging modeling methods, namely\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/sup\u003e2D, 2.5D, and 3D, have been evaluated and an ensemble learning-facilitated approach for feature integration is also presented to enhance the performance and clinical interpretability and decision utility of the model.\u003c/p\u003e \u003cp\u003eThe results show that the proposed ensemble model has superior performance compared to other models in all performance metric criteria\u003csup\u003e[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/sup\u003e. The superior performance of the proposed model may be due to a number of reasons, which include the following: first, the proposed model combines a number of base models with different architectures as well as feature dimensions, thus creating a heterogeneous ensemble model that has the ability to extract tumor-related features from different perspectives, hence reducing errors in prediction that are associated with structural biases as well as feature limitations that are associated with each of the models\u003csup\u003e[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/sup\u003e. Second, the proposed model has the advantage of creating an ensemble that has a regularizing effect, thus having the ability to prevent the occurrence of overtraining, which is a common problem that occurs in complex models when dealing with a limited number of training samples.The proposed model has the advantage of adapting the outputs from each of the constituent models during the integration process, thus having the ability to improve the reliability of the final outputs that are produced during the prediction process.\u003c/p\u003e \u003cp\u003eIn the comparative evaluation of the performance of the models on an individual level, the proposed model achieved superior performance in terms of generalization capability as well as calibration performance\u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]\u003c/sup\u003e. It is reasonable to assume that the performance of the proposed model can be related to the mechanism of feature construction employed by the model. The proposed model utilizes the mechanism of the 2.5D model, where features from the coronal, sagittal, and axial planes of the tumor volume are simultaneously constructed. Hence, the model is able to take advantage of the high computational efficiency as well as the high architectural stability of the 2D models while incorporating the advantages of the 3D model as well \u003csup\u003e[\u003cspan additionalcitationids=\"CR31\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/sup\u003e. Compared to the deep-learning-based models, the proposed model utilizes fewer parameters as well as lesser data requirements, thus being more suitable to the current clinical environment where the sample sizes of the images obtained from the patients are low. Unlike the traditional models of the 2D type, the proposed model is able to more comprehensively deal with the spatial distribution characteristics of the tumor, thus preventing prediction bias due to information loss\u003csup\u003e[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe study not only examines the discriminative ability of the models but also their applicability to real-world settings\u003csup\u003e[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]\u003c/sup\u003e. DCA reveals the ensemble model\u0026rsquo;s net benefit to be superior to the \u0026ldquo;treat all\u0026rdquo; and \u0026ldquo;treat none\u0026rdquo; strategies across the range of decision thresholds, confirming the utility of the ensemble model for decision-making. The calibration curves reveal the correspondence of the observed outcomes with the predicted results of the ensemble model, suggesting the utility of the risk probability results for risk stratification of the patient population, thereby increasing the usability of the results obtained with the ensemble model \u003csup\u003e[\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eLimitations of This Study\u003c/h2\u003e \u003cp\u003e This research incorporated a set of inherent limitations.First, although data were obtained from various centers, the total sample size was limited, and this could potentially impact the consistency in the assessment of the performance of the models. Second, the retrospective study design was potentially limited to selection biases; hence, further validation in large numbers in a prospective study design would be desirable to confirm the generalizability and clinical applicability of the model. The model was based only on imaging characteristics and did not account for clinical information, pathomics, or genomic information. It would therefore be desirable to explore the construction of a multimodal fusion model to enrich the predictive system.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn summary, this study assessed and compared various dimensions of imaging modeling strategies and established an ensemble learning-based fusion prediction model. This model had excellent performance in preoperative and noninvasive prediction of Ki-67 proliferation indices of T1 lung adenocarcinomas and outperformed other individual models in discriminability and had exceptional clinical decision net benefit and probability.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eEthics Approval and Consent to Participate\u003c/h2\u003e \u003cp\u003e This multicenter retrospective study was conducted in strict accordance with the Declaration of Helsinki (World Medical Association, 2013 revision) and the Measures for the Ethical Review of Biomedical Research Involving Humans issued by the National Health Commission of the People's Republic of China. The study protocol was reviewed and formally approved by the Medical Research and Clinical Trial Ethics Committee of The First Affiliated Hospital of Huzhou Normal University (also known as Huzhou First People's Hospital), with the ethics approval number: 2024KYLL085-01.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis work was supported by grants from the Science and Technology Project of Huzhou City, Zhejiang Province(2024GY41) and the Health Science and Technology Program(2025KY1555).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eX.P. (Xiuhua Peng)：Study design, data collection and analysis, manuscript drafting, funding acquisitionX.X. (Xinchao Xu)：Imaging data processing, statistical analysis.J.X. (Jingnan Xue)：Case data collation, figure and table preparation.P.X. (Pengliang Xu)：Methodological guidance, result validation. funding acquisitionY.L. (Yuanbin Li)：Literature search, initial manuscript revision.M.Y. (Mei Yang)：Ethics application, patient informed consent acquisition.L.Q. (Luying Qi)：Data verification, supplementary material organization.J.Z. (Jiacheng Zhao)：Imaging technical support.F.X. (Fenglin Xiang)：Study conception, protocol optimization.Y.S. (Yanyan Shen)：Manuscript polishing, language proofreading.S.F. (Shenyun Fang)：Data entry, result visualization.H.Z. (Hongxing Zhao)：Academic supervision.J.J. (Jianping Jiang, corresponding author)：Study oversight, final manuscript review and submission.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eChen H, Liu L, Zhang M, et al. Correlation of LOXL2 expression in non-small cell lung cancer with immunotherapy[J]. Int J Clin Exp Pathol. 2024;17(9):268\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDong H, Li Y, Zhao L, et al. Consensus clustering based on CT radiomics has potential for risk stratification of patients with clinical T1 stage lung adenocarcinoma[J]. BMC Med Imaging. 2025;25(1):231.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShi W, Hu Y, Wang H, et al. The value of relative CT attenuation in predicting invasiveness in patients with T1-stage lung adenocarcinoma[J]. Quant imaging Med Surg. 2025;15(10):9294\u0026ndash;308.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHerbst RS, Morgensztern D, Boshoff C. The biology and management of non-small cell lung cancer[J]. Nature. 2018;553(7689):446\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWarth A, Cortis J, Soltermann A, et al. Tumour cell proliferation (Ki-67) in non-small cell lung cancer: a critical reappraisal of its prognostic role[J]. Br J Cancer. 2014;111(6):1222\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Feng H, Ma S, et al. Clinicopathological characteristics of peripheral clinical stage IA lung adenocarcinoma with high Ki-67 expression[J]. Translational cancer Res. 2021;10(1):152\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGom\u0026agrave; C, Moll\u0026agrave; M, Oses G, et al. Preoperative Radiation Therapy-Induced Molecular and Immune Modulation in Early-Stage Breast Cancer: Results From the YOUNGSTER trial[J]. Int J Radiat Oncol Biol Phys. 2026;124(1):122\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang X, Zhang H, Zhang X. Analysis of Influencing Factors of Complications for CT-guided ?Percutaneous Lung Biopsy[J]. Zhongguo fei ai za zhi\u0026thinsp;=\u0026thinsp;Chinese. J lung cancer. 2024;27(3):179\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou B, Xu J, Tian Y, et al. Correlation between radiomic features based on contrast-enhanced computed tomography images and Ki-67 proliferation index in lung cancer: A preliminary study[J]. Thorac cancer. 2018;9(10):1235\u0026ndash;40.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGu Q, Feng Z, Liang Q, et al. Machine learning-based radiomics strategy for prediction of cell proliferation in non-small cell lung cancer[J]. Eur J Radiol. 2019;118:32\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTruhn D, Schrading S, Haarburger C, et al. Radiomic versus Convolutional Neural Networks Analysis for Classification of Contrast-enhancing Lesions at Multiparametric Breast MRI[J]. Radiology. 2019;290(2):290\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang G, Nie P, Zhao L, et al. 2D and 3D texture analysis to predict lymphovascular invasion in lung adenocarcinoma[J]. Eur J Radiol. 2020;129:109111.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang T, Ke M, Liu Q, et al. A computed tomography-based deep learning model for non-invasively predicting World Health Organization (WHO)/International Society of Urological Pathology (ISUP) pathological grades of clear cell renal cell carcinoma (ccRCC): a multicenter cohort study[J]. Translational Androl Urol. 2025;14(7):2018\u0026ndash;28.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim S, Choi M, Han S, et al. Comparison of 2D, 2.5D, and 3D landmark localization networks for 3D cephalometry in CT images[J]. BMC Oral Health. 2025;25(1):1843.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang W, Ren M, Ren J, et al. Predicting Radiation Pneumonitis Integrating Clinical Information, Medical Text, and 2.5D Deep Learning Features in Lung Cancer[J]. Int J Radiat Oncol Biol Phys. 2026;124(1):194\u0026ndash;205.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheng X, Li H, Li C et al. Attention-based deep learning network for predicting World Health Organization meningioma grade and Ki-67 expression based on magnetic resonance imaging[J]. Eur Radiol, 2025:10\u0026ndash;1007.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark GE, Kim SH, Nam Y, et al. 3D Breast Cancer Segmentation in DCE-MRI Using Deep Learning With Weak Annotation[J]. J Magn Reson imaging: JMRI. 2024;59(6):2252\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGunasekaran H, Ramalakshmi K, Swaminathan DK et al. GIT-Net: An Ensemble Deep Learning-Based GI Tract Classification of Endoscopic Images[J]. Bioengineering (Basel, Switzerland), 2023,10(7):809.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSui C, Chen K, Ding E, et al. (18)F-FDG PET/CT-based intratumoral and peritumoral radiomics combining ensemble learning for prognosis prediction in hepatocellular carcinoma: a multi-center study[J]. BMC Cancer. 2025;25(1):300.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu M, Yang Z, Zhao W, et al. Predicting Ki-67 labeling index level in early-stage lung adenocarcinomas manifesting as ground-glass opacity nodules using intra-nodular and peri-nodular radiomic features[J]. Cancer Med. 2022;11(21):3982\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSong C, Chen J, Zhao C, et al. Prediction of Ki-67 Expression in HIV-Associated Lung Adenocarcinoma Patients Using Multiple Machine Learning Models Based on CT Imaging Radiomics[J]. Cancer Manage Res. 2025;17:881\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu M, Yang Z, Zhao W, et al. Predicting Ki-67 labeling index level in early-stage lung adenocarcinomas manifesting as ground-glass opacity nodules using intra-nodular and peri-nodular radiomic features[J]. Cancer Med. 2022;11(21):3982\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu C, Liang Y, Li Y, et al. Prediction of the Ki-67 proliferation index in lung adenocarcinoma using an interpretable CT-based deep learning radiomics model: a two-center study[J]. BMC Pulm Med. 2025;25(1):536.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlabi RO, Elmusrati M, Leivo I, et al. Artificial Intelligence-Driven Radiomics in Head and Neck Cancer: Current Status and Future Prospects[J]. Int J Med Informatics. 2024;188:105464.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi M, Gu H, Xue T, et al. CT-based radiomics nomogram for the pre-operative prediction of lymphovascular invasion in colorectal cancer: a multicenter study[J]. Br J Radiol. 2023;96(1141):20220568.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTanveer MA, Khan MJ, Sajid H, et al. Convolutional neural networks ensemble model for neonatal seizure detection[J]. J Neurosci Methods. 2021;358:109197.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Y, Chou H, Lin P, et al. A novel deep learning-based algorithm combining histopathological features with tissue areas to predict colorectal cancer survival from whole-slide images[J]. J translational Med. 2023;21(1):731.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEckhart L, Lenhof K, Rolli L, et al. A comprehensive benchmarking of machine learning algorithms and dimensionality reduction methods for drug sensitivity prediction[J]. Brief Bioinform. 2024;25(4):bbae242.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoganathan SA, Paul SN, Paloor S, et al. Automatic segmentation of magnetic resonance images for high-dose-rate cervical cancer brachytherapy using deep learning[J]. Med Phys. 2022;49(3):1571\u0026ndash;84.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRosas-Gonzalez S, Birgui-Sekou T, Hidane M, et al. Asymmetric Ensemble of Asymmetric U-Net Models for Brain Tumor Segmentation With Uncertainty Estimation[J]. Front Neurol. 2021;12:609646.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang C, Li E, Hu J, et al. Enabling Early Identification of Malignant Vertebral Compression Fractures Through 2.5D Convolutional Neural Network Model With CT Image Analysis[J]. Spine. 2025;50(24):1728\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar A, Jiang H, Imran M, et al. A flexible 2.5D medical image segmentation approach with in-slice and cross-slice attention[J]. Comput Biol Med. 2024;182:109173.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKarimzadeh M, Seyedarabi H, Jodeiri A, et al. Enhanced Brain Stroke Lesion Segmentation in MRI Using a 2.5D Transformer Backbone U-Net Model[J]. Brain Sci. 2025;15(8):778.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGao F, Hu Z, Xian J, et al. Mixed U-Net: Segmentation of focal liver lesions using a hybrid 2D and 3D model[J]. J Appl Clin Med Phys. 2026;27(1):e70408.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlonso S, Bra AI, Loredo M, et al. Performance of Disease Activity Indices Used in Axial Spondyloarthritis in Real-World Clinical Settings[J]. J Rhuematol. 2025;52(5):444\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVetsch T, Huber M, Wuethrich PY, et al. Preoperative prediction of severe short-term complications in patients with bladder cancer undergoing radical cystectomy[J]. Surg Oncol. 2025;61:102253.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen S, Deng T, Yang Q, et al. Development and validation of an explainable machine learning model for predicting postoperative pulmonary complications after lung cancer surgery: a machine learning study[J]. EClinicalMedicine. 2025;86:103386.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Invasive lung adenocarcinoma, deep learning, Ensemble, Ki-67,2.5D Introduction","lastPublishedDoi":"10.21203/rs.3.rs-8746230/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8746230/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjective:\u003c/strong\u003eIn order to set up a predictive model which combines 2.5D deep learning, 2D deep learning, 3D deep learning, and an ensemble fusion model, which could be used for the accurate prediction of preoperative expression levels of ki-67 in patients diagnosed with stage T1 invasive lung adenocarcinoma(LUAD ).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePatients and Methods:\u003c/strong\u003e In total, 503 patients from our own institution and 102 patients with invasive LUAD of T1 stage from two other centers were included retrospectively. The subjects were then divided into a Ki-67 high-expression group and a Ki-67 low-expression group, comprising 254 and 351 subjects, respectively. The subject set from our own institution formed the training set, and the subject sets from the other two centers formed the test set. Three categories of deep learning(DL) models, 2D, 2.5D, and 3D models, and an ensemble fusion model were developed and used to evaluate the performance of each model in predicting Ki-67 expressions in subjects with T1 LUAD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eAll of these models showed excellent discriminative performance for the training set. The Ensemble model showed the best discriminative performance with an AUC of 0.991 (95% CI: 0.9858–0.9964). The Support Vector Machine 3D model (SVM 3D) showed an AUC of 0.970 (95% CI: 0.9538–0.9866). The Multilayer Perceptron 2D model (MLP 2D; AUC = 0.931; 95% CI: 0.9093–0.9529) and Gradient Boosting Machine 2.5D model (GBM 2.5D; AUC = 0.908; 95% CI: 0.8811–0.9342) also showed excellent discriminative performance. For all of these models, AUC performance declined as expected for the independent test set; however, some of these models showed different degrees of performance declines. The Ensemble model showed excellent robustness to performance declines and showed an AUC of 0.870 (95% CI: 0.7943–0.9447). The GBM 2.5D model showed excellent robustness to performance declines and showed an AUC of 0.843 (95% CI: 0.7405–0.9450); it showed the smallest performance declines of all of the models. The MLP 2D model showed moderate performance and showed an AUC of 0.810 (95% CI: 0.7110–0.9099). On the other hand, it was evident that the performance of the SVM 3D model showed substantial declines; it showed an AUC of 0.745 (95% CI: 0.6338–0.8566); it showed the greatest performance declines of all of the models and possibly suffered from.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eThe Ensemble model demonstrated high performance in the prediction of Ki-67 expression in stage T1 invasive LUAD, which confirms the feasibility of the proposed approach in the prediction of Ki-67.\u003c/p\u003e","manuscriptTitle":"Preoperative Prediction of Ki-67 in Stage T1 Lung Adenocarcinoma Based on 2.5D and Ensemble Integrated Models: A Multicenter Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-13 09:23:29","doi":"10.21203/rs.3.rs-8746230/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-20T17:47:13+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-13T15:23:03+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"62803239699089455983343528591273342484","date":"2026-04-08T07:46:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"305887243621629194125763010451388980390","date":"2026-04-07T13:48:12+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-01T12:20:44+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"312164797269296175648760417184292983929","date":"2026-03-26T13:21:37+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-09T06:27:45+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-05T10:16:21+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-05T10:16:12+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2026-01-31T03:57:58+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b7e5015c-db8a-4f8a-b12b-2a09fcfc9058","owner":[],"postedDate":"March 13th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-01T03:08:16+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-13 09:23:29","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8746230","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8746230","identity":"rs-8746230","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}