Intraoperative lung cancer precise operation guidance system based on radiomics and frozen section multidisciplinary diagnosis

Abstract

The invention discloses an intraoperative lung cancer precise operation guidance system based on radiomics and frozen section multidisciplinary diagnosis. According to the method, a thin-layer CT image of the chest of a patient is taken as a research object, and the cytological type and tissue subtype of each focus are obtained according to a postoperative pathological result, so that a real label of a high-grade component of the invasive lung adenocarcinoma is obtained; open source software 3D Slicer is used for sketching lesions manually, then a plug-in Slicer-Radiomics of the 3D Slicer is used for extracting image omics characteristics of each lesion according to set parameters, a data set is formed in combination with real labels, the lesion image omics characteristics in the training data set serve as input, lung high-level components based on pathological diagnosis results serve as the real labels, and the lung high-level components in the lung high-level components of the pathological diagnosis results serve as the real labels. A radiomics model R of lung adenocarcinoma high-level component prediction is trained, and an R-FS parallel diagnosis model is constructed in combination with an intra-operative frozen section pathology (FS) result so as to guide final operative selection.

Description

Technical Field

[0001] The present invention relates to an intraoperative precise lung cancer surgery guidance system based on radiomics and frozen section multidisciplinary diagnosis, belonging to the technical field of biomedicine. Background Art

[0002] Lung cancer has always been one of the most serious diseases threatening human health worldwide, among which lung adenocarcinoma (LUAD) accounts for the highest proportion. With the development of screening technology, especially the popularization of low-dose computed tomography technology, many lesions can be discovered earlier and receive surgical treatment opportunities, thus effectively reducing the mortality rate of lung cancer.

[0003] However, the choice of surgical treatment remains a focus of medical attention. In recent years, researchers have focused on the differences between surgical methods for different early-stage non-small cell lung cancer (NSCLC). For early-stage lung cancer (T1aN0-7th edition T / N / M, tumor size ≤2cm, or even stage I), a large number of clinical trials have confirmed that sublobar resection (SLR, including wedge resection or segmental resection) is not only not inferior to lobectomy in terms of disease-free survival, but also can preserve more lung function, reduce perioperative complications, and reduce the mortality rate of lung cancer patients. This finding challenges lobectomy, the standard surgical method for treating T1N0 since 1995.

[0004] It is worth noting that even though most patients are diagnosed with LUAD, its classification, especially histological subtypes, is constantly being refined and improved. LUAD is a histologically heterogeneous disease with complex histological subtypes and growth patterns. The International Association for the Study of Lung Cancer (IASLC), the American Thoracic Society (ATS), and the European Respiratory Society (ERS) have developed a new multidisciplinary classification for LUAD, including lepidic, acinar, papillary, micropapillary, and solid types (recorded in units of 5% growth ratio). Different growth patterns, especially high-grade patterns such as micropapillary, solid, and cribriform, represent different biological behaviors and clinical significance, thus affecting clinical decision-making. High-grade patterns are usually accompanied by more frequent hypermethylation, KRAS mutations, worse prognosis, a high proportion of recurrence, and more frequent lymph node micrometastasis, leading to the need for more aggressive treatment strategies, such as lobectomy, systematic mediastinal lymph node resection, and adjuvant chemotherapy.

[0005] There is an urgent need for a reliable method to obtain high-level histological information from tissue samples in a timely and accurate manner. In addition to the final pathological examination, intraoperative frozen sections (FS) provide a practical approach. Although FS shows high consistency in distinguishing invasive adenocarcinoma, glandular precursor lesions, and microinvasive adenocarcinoma, it performs poorly in high-grade growth pattern (HGP). Studies have found that sampling and interpretation errors and poor quality of frozen sections are the main causes of misdiagnosis. Radiomics can provide histological information in a timely, simple, and non-invasive manner, avoid sampling errors, comprehensively consider the entire tumor tissue, and reduce the impact of human intervention.

[0006] The future development direction is to integrate radiomics and frozen section pathological diagnosis to provide a rapid and accurate multidisciplinary diagnostic method for patients with early lung cancer to guide surgical decisions and optimize patients' treatment strategies. Summary of the invention

[0007] The purpose of the present invention is: in view of the shortcomings of the prior art, in order to solve the technical problem of how to accurately identify the benign and malignant nature of the resected lung tissue and identify whether the resected lung adenocarcinoma tissue contains high-grade components, the present invention provides an intraoperative lung cancer precision surgery guidance system based on radiomics and frozen section multidisciplinary diagnosis. The present invention improves the accuracy of diagnosis by establishing and comparing three models for predicting T1 invasive lung adenocarcinoma high-grade pattern (HGP) before or during surgery, and combining the R-FS joint diagnosis model of radiomics and frozen section pathological diagnosis; the system of the present invention is based on the constructed R-FS model, which can provide practical and powerful decision support for clinical surgery and ensure the timeliness and effectiveness of treatment strategies.

[0008] In order to achieve the above object, the present invention provides a method for constructing a high-grade pattern prediction model for lung adenocarcinoma based on radiomics and frozen section multidisciplinary diagnosis, the method comprising the following steps:

[0009] S1. Using preoperative thin-layer chest CT images as the object, the lesions were manually segmented using 3D Slicer, and the cytological type and tissue subtype of each lesion were obtained based on the postoperative pathological results, thereby obtaining the true label of whether invasive lung adenocarcinoma has high-grade components;

[0010] S2. Use the open source software 3D Slicer to manually outline the lesions, and then use the 3D Slicer plug-in Slicer-Radiomics to extract the radiomic features of each lesion according to the set parameters, and combine them with the real labels to form a training data set;

[0011] S3. Using the radiomic features of lesions in the training dataset as input and the high-grade lung components based on pathological diagnosis results as true labels, a radiomics model R for predicting high-grade components of lung adenocarcinoma was trained.

[0012] S4. Combining the results of intraoperative frozen section pathology (FS) with the radiomics model, a parallel diagnostic model (R-FS) was constructed to identify high-grade patterns of lung adenocarcinoma.

[0013] Preferably, in step S1, the thin-layer chest CT image refers to a CT lung window image including a layer thickness of ≤1 mm.

[0014] Preferably, in step S1, in cases cytologically diagnosed as invasive lung adenocarcinoma, based on the histological subtypes included, lesions including microinvasive type, solid type, cribriform type, and complex glandular type are considered to be lesions containing high-grade components, and lesions including only lepidic type, acinar type, and papillary type are considered to be non-high-grade lesions; lung adenocarcinomas containing high-grade components and those without high-grade components are used as true labels and prediction targets for training classification models.

[0015] Preferably, in step S2, the set parameters are: (a) Manual Customization: √; (b) Toggle Features: All Features √; (c) Resampled voxel size: 1, 1, 1; (d) Wavelet-based features: √; (e) Bin Width: 25; e) Enforce Symmetrical GLCM: √.

[0016] Preferably, in step S2, the finally extracted radiomics features include 6 feature categories and 9 image types; wherein,

[0017] The six feature categories are based on the first-order shape, the gray-level co-occurrence matrix GLCM, the gray-level correlation matrix GLDM, the gray-level run length matrix GLRLM, the adjacent gray-level tone difference matrix NGTDM, and the gray-level size zone matrix GLSZM;

[0018] The nine image types are respectively the original image and the other eight images after wavelet transformation.

[0019] Preferably, in step S3, the training process of the radiomics model R is:

[0020] Use variance filtering, mutual information method, minimum absolute value convergence and selection operator (Lasso) to select the best features for model training;

[0021] Randomly select 8 to 10 groups of data as training sample modeling features to train a random forest (RF) model, and then use the output of the random forest model as the initial value to train the gradient boosting tree (GBDT) model;

[0022] A 5-fold cross validation and Bayesian optimization algorithm were used, and the area under the curve of the validation set was used as the optimization target.

[0023] Preferably, in step S4, if the intraoperative frozen section determines the presence of high-grade components or detects high-risk pathological features highly correlated with them, such as poor differentiation, pleural metastasis, vascular cancer thrombus, and dissemination through the air cavity, the frozen section is considered to be successfully predicted and marked as 1 (with a probability of 100%), otherwise it is marked as 0 (with a probability of 0).

[0024] Preferably, in step S4, the recognition principle of the parallel diagnosis model is that if one of the results of radiomics or intraoperative freezing indicates that the lesion contains high-grade components, then the lesion is considered to contain high-grade components.

[0025] Preferably, in step S4, the R-FS model is constructed by the model fusion (soft voting) method using the R model prediction probability and the intraoperative frozen section probability, and the same weight is given to the two. According to the parallel diagnosis principle, 0.25 is used as the boundary value, and the final P>=0.25 is considered that the R-FS model predicts a positive result for the lesion.

[0026] The present invention also provides an intraoperative lung cancer precision surgery guidance system based on radiomics and frozen section multidisciplinary diagnosis, comprising:

[0027] Preoperative imaging data acquisition module: used to obtain radiomics data through imaging scanning methods such as CT or MRI before surgery as one of the input data sources of the model;

[0028] Intraoperative frozen section diagnosis module: used to obtain tissue samples from the patient's tumor site during surgery for frozen section (FS) analysis to obtain real-time pathological information;

[0029] R-FS model prediction module: used to integrate the pathological diagnosis results of radiomics and frozen sections through the R-FS model to provide accurate diagnostic support for the prediction of high-grade pattern (HGP) of lung adenocarcinoma before and during surgery; the R-FS model model is a high-grade pattern prediction model of lung adenocarcinoma constructed by the above-mentioned construction method.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) Diagnostic accuracy: By integrating radiomics with frozen section pathological diagnosis, the advantages of both methods are combined to improve the accuracy of preoperative and intraoperative identification of high-grade pattern (HGP) of lung adenocarcinoma, laying the foundation for more accurate clinical decision-making;

[0032] (2) Real-time diagnosis: Frozen section diagnosis provides real-time pathological information during surgery. Combined with the R-FS model, it can provide immediate feedback on diagnostic results during surgery, providing guidance for surgeons to adjust surgical strategies in a timely manner.

[0033] (3) Multidisciplinary approach: The model combines two complementary diagnostic tools, radiomics and frozen sections, to achieve multidimensional and comprehensive pathological diagnosis and provide personalized surgical treatment plans for patients with lung adenocarcinoma;

[0034] (4) Optimizing surgical plans: By accurately identifying high-grade components of tissue samples, surgeons can more reasonably determine the extent of lobectomy and the method of lymph node dissection, thereby maximizing the preservation of the patient's lung function and reducing the incidence of postoperative complications;

[0035] (5) Non-invasive prediction: Radiomics, as a non-invasive imaging analysis method, can evaluate tumor characteristics holistically, avoid the influence of sampling errors, and provide important information for comprehensive diagnosis;

[0036] (6) Improve prognosis: With the accurate predictive information provided by the model, surgeons can formulate more aggressive and effective surgical and adjuvant treatment plans, thereby improving the patient's survival rate and prognosis;

[0037] (7) Reduce the misdiagnosis rate: Use the model comprehensive diagnostic method to reduce the sampling and interpretation errors in frozen section diagnosis, improve the reliability of diagnosis, and reduce the occurrence of intraoperative misdiagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flowchart for constructing a multidisciplinary diagnostic model based on radiomics and frozen section (FS) pathological diagnosis;

[0039] Figure 2 It is a flow chart of the implementation of the present invention before and during surgery. DETAILED DESCRIPTION

[0040] In order to make the present invention more clearly understood, preferred embodiments are described in detail below with reference to the accompanying drawings.

[0041] Example

[0042] This embodiment provides a high-grade pattern (HGP) prediction model and method for lung adenocarcinoma based on radiomics and frozen section (FS) pathological diagnosis, which is used to guide the selection of lung cancer surgical procedures:

[0043] I. Patients and Analytical Methods

[0044] (1) Patients and data: A retrospective analysis was performed on patients who underwent radical resection for lung adenocarcinoma at Zhongshan Hospital Affiliated to Fudan University from January to April 2019. Preoperative, intraoperative, and postoperative data were included, including clinical data, pathological results, CT images, etc. The dataset was randomly divided into a training set and a test set in a ratio of 8:2.

[0045] (2) CT examination and pathological analysis: Patients underwent chest CT scans before surgery, and thin-section lung window CT images were selected for analysis. Pathological sections were interpreted by junior pathologists and reviewed by senior pathologists to determine whether the lesions were HGP or non-HGP. Frozen sections were used for comparison with the radiomics model and run in parallel in the R-FS model.

[0046] (3) Radiomics analysis: Tumors in patient CT images were segmented and reconstructed, 852 features were extracted and selected, and a radiomics model (R model) was constructed using machine learning methods. Model interpretation, including global and local interpretation, was performed to reveal the importance of features.

[0047] Second, the construction method of the high-grade pattern (HGP) prediction model of lung adenocarcinoma based on radiomics (R) and frozen section (FS) pathological diagnosis is used for rapid intraoperative diagnosis of high-grade tissue components of lung malignancies, such as Figure 1 As shown, the method comprises the following steps:

[0048] S1. Using thin-layer chest CT images as the object, the pulmonary nodule lesions are manually segmented layer by layer in the open source software 3D Slicer to form a 3D mask. The specific tissue subtype of each lung adenocarcinoma is obtained according to the pathological biopsy results, and then the true label of the high-level classification is obtained.

[0049] S2. Then use the 3D Slicer plug-in Slicer-Radiomics to extract the radiomics features of each lesion according to the set parameters, and combine them with the real labels to form a training data set. Note: 3D Slicer is an open source medical image analysis and visualization software platform that can perform image segmentation, reconstruction, registration, marker selection, measurement and other operations; pyradiomics is an open source python software package that can extract radiomics radiomics features from medical images; SlicerRadiomics is an extension of 3D Slicer, which includes pyradiomics functions for calculating various radiomics features.

[0050] In step S2, the setting parameters are: (a) Manual Customization: √; (b) Toggle Features: All Features √; (c) Resampled voxel size: 1, 1, 1; (d) Wavelet-based features: √; (e) Bin Width: 25; e) Enforce Symmetrical GLCM: √

[0051] In step S2, the open source Pyradiomics software library (for details, see the literature: van Griethuysen, JJM, Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, RGH, Fillion-Robin, JC, Pieper, S., Aerts, HJWL (2017). Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Research, 77(21), e104–e107. https: / / doi.org / 10.1158 / 0008-5472.CAN-17-0339) is used to automatically extract the above-mentioned radiomic features of pulmonary nodules with clear pathological diagnosis.

[0052] Specifically, taking a lung nodule lesion as an example, the original three-dimensional image and the corresponding binary mask of the lesion are used as the volume and segmentation columns respectively, and finally a series of radiomic features of the lesion are obtained. The formulas and definitions of all features can be referred to the link https: / / pyradiomics.readthedocs.io / en / latest / index.html. The binary mask of each lung nodule is multiplied by the original three-dimensional image to accurately obtain the image area belonging only to this lung nodule. The first-order features (statistics) describe the distribution of voxel density of the target object in the image, such as mean, standard deviation, skewness, abundance, energy, entropy, etc. Shape features describe the size and shape of the target object. These features are independent of voxel density, such as volume, surface area, sphericity, maximum diameter, minimum diameter, etc. Shape features are available in three-dimensional and two-dimensional forms. Please refer to the above documents and Table 1 for the remaining gray-level co-occurrence matrix features, gray-level long-distance matrix features, gray-level size region matrix, adjacent gray-level tone difference matrix, and gray-level dependency matrix. The nine image types are the original image and eight wavelet transform (wavelet) images after the original image has undergone three high-frequency (H) or low-frequency (L) filterings, respectively represented by HHH, HHL, HLH, HLL, LHH, LHL, LLH, and LLL. Assuming that there are a total of N lung nodules in the data set, and each lung nodule is planned to extract 852 features, after this step, a feature data set of N×852 will be obtained under each resolution CT image.

[0053] Table 1 Radiomics characteristics

[0054]

[0055]

[0056]

[0057]

[0058] S3. Using the lesion imaging features in the training dataset as input and the high-grade lung components based on the pathological diagnosis results as the true labels, a lung adenocarcinoma high-grade component prediction model was trained.

[0059] In the sklearn library, the variance filtering method was used to screen out 658 features with P < 0.05, and the mutual information method was used to screen out 749 features with values ​​greater than 0. Then, the intersection of the two was selected to retain 635 features.

[0060] Finally, the LASSO method was used to select 28 optimal features from all radiomics features to form an N×28 two-dimensional data table to enter the modeling process;

[0061] The LASSO hyperparameters are set as follows: model_lassoCV = LassoCV(eps = 0.0001, n_alphas = 500, cv = 5, random_state = 0, max_iter = 3000, tol = 0.001);

[0062] The final 28 features selected are as follows:

[0063]

[0064]

[0065] A random forest classification model (RFC) was fitted using N×28, and then its predicted probability results were used as initial values ​​to train the gradient boosted tree model (GBDT). The parameter adjustment method was Bayesian optimization based on the Hyperopt library, and 5-fold cross validation was used. The final nested model combined the advantages of bagging and boosting models.

[0066] The model parameters are as follows:

[0067]

[0068] In step S4, the parallel diagnosis model training process is:

[0069] Parallel diagnosis principle: If one result of imaging omics or intraoperative freezing indicates that the lesion contains high-grade components, then the lesion is considered to contain high-grade components;

[0070] If the intraoperative frozen section is judged to have high-grade components or high-risk pathological features highly associated with it, such as poor differentiation, pleural metastasis, vascular cancer thrombus, and air cavity dissemination, it is considered that the frozen section is successfully predicted and marked as 1 (probability 100%), otherwise it is marked as 0 (probability 0);

[0071] The R-FS model was constructed by combining the R model prediction probability and the intraoperative frozen section probability with the model fusion (soft voting) method, and the two were given the same weight. According to the parallel diagnosis principle, 0.25 was used as the cutoff value. If the final P>=0.25, the R-FS model was considered to have a positive prediction result for the lesion.

[0072] According to the model, the probability value y of the sample in the R, FS and R-FS models is solved, and y is the probability that the lesion contains a high-grade tissue subtype in the three models (range 0 to 1). The prediction results of the R-FS model on the test set are AUC value, sensitivity, specificity, accuracy and missed diagnosis rate of 0.907, 83.3%, 94.1%, 89.8% and 23.3%, respectively.

[0073] The model constructed in the present invention is used to guide the selection of surgical procedures. The specific process is as follows: Figure 2 As shown in the figure, for patients who need surgery, preoperative CT images are collected, and then 3DSlicer is used to manually outline the lesions and extract the omics parameters as required. Then, the best feature value is input into the R model to obtain the probability P. Then, the intraoperative freezing probability and the imaging omics results are fused into the model to obtain the R-FS model P value to comprehensively consider the surgical method; the specific steps include:

[0074] ① Preoperative imaging data collection: Before surgery, radiomics data are obtained through imaging scanning methods such as CT or MRI as one of the input data sources of the model.

[0075] ② Intraoperative frozen section diagnosis: During the operation, tissue samples are obtained from the patient's tumor site for frozen section (FS) analysis to obtain real-time pathological information.

[0076] ③R-FS model integration: The R-FS model integrates the pathological diagnosis results of radiomics and frozen sections to provide accurate diagnostic support for the prediction of high-grade pattern (HGP) of lung adenocarcinoma before and during surgery.

[0077] ④ Diagnostic information feedback: The model analysis results are fed back to the surgeon in real time to determine the benign or malignant nature of the removed tissue samples and whether there are high-grade components in the lung adenocarcinoma tissue.

[0078] ⑤ Formulate clinical strategies: Based on the comprehensive prediction results of the model, surgeons can adjust the surgical plan in a timely manner, such as determining the scope of lobectomy and the method of lymph node dissection, to ensure the accuracy and effectiveness of the treatment strategy.

[0079] The above description is only a preferred embodiment of the present invention and is not any formal or substantial limitation of the present invention. It should be pointed out that a person skilled in the art can make several improvements and supplements without departing from the present invention, and these improvements and supplements should also be regarded as the protection scope of the present invention.

Claims

1. A method for constructing a high-grade pattern prediction model for lung adenocarcinoma based on multidisciplinary diagnosis of radiomics and frozen sections, the method comprising the following steps: S1. Using preoperative thin-layer chest CT images as the object, the lesions were manually segmented using 3D Slicer, and the cytological type and tissue subtype of each lesion were obtained based on the postoperative pathological results, thereby obtaining the true label of whether invasive lung adenocarcinoma has high-grade components; S2. Use the open source software 3D Slicer to manually outline the lesions, and then use the 3D Slicer plug-in Slicer-Radiomics to extract the radiomic features of each lesion according to the set parameters, and combine them with the real labels to form a training data set; S3. Using the radiomic features of lesions in the training dataset as input and the high-grade lung components based on pathological diagnosis results as true labels, a radiomics model R for predicting high-grade components of lung adenocarcinoma was trained. S4. Combining the results of intraoperative frozen section pathology (FS) with the radiomics model, a parallel diagnostic model (R-FS) was constructed to identify high-grade patterns of lung adenocarcinoma.

2. The construction method according to claim 1, characterized in that: In step S1, the thin-layer chest CT image refers to a CT lung window image with a layer thickness of ≤1 mm.

3. The construction method according to claim 1, characterized in that: In step S1, in cases diagnosed as invasive lung adenocarcinoma by cytology, lesions including microinvasive type, solid type, cribriform type, and complex glandular type are considered to contain high-grade components according to the histological subtypes included, and lesions including only lepidic type, acinar type, and papillary type are considered to be non-high-grade lesions; lung adenocarcinomas containing high-grade components and those without high-grade components are used as true labels and prediction targets for training classification models.

4. The construction method according to claim 1, characterized in that: In the step S2, the set parameters are: (a) Manual Customization: √; (b) Toggle Features: All Features √; (c) Resampled voxel size: 1, 1, 1; (d) Wavelet-based features: √; (e) Bin Width: 25; e) Enforce Symmetrical GLCM: √.

5. The construction method according to claim 1, characterized in that: In step S2, the finally extracted radiomics features include 6 feature categories and 9 image types; among which, The six feature categories are based on the first-order shape, the gray-level co-occurrence matrix GLCM, the gray-level correlation matrix GLDM, the gray-level run length matrix GLRLM, the adjacent gray-level tone difference matrix NGTDM, and the gray-level size zone matrix GLSZM; The nine image types are respectively the original image and the other eight images after wavelet transformation.

6. The construction method according to claim 1, characterized in that: In step S3, the training process of the radiomics model R is: Use variance filtering, mutual information method, minimum absolute value convergence and selection operator (Lasso) to select the best features for model training; Randomly select 8 to 10 groups of data as training sample modeling features to train a random forest (RF) model, and then use the output of the random forest model as the initial value to train the gradient boosting tree (GBDT) model; A 5-fold cross validation and Bayesian optimization algorithm were used, and the area under the curve of the validation set was used as the optimization target.

7. The construction method according to claim 1, characterized in that: In step S4, if the intraoperative frozen section determines the presence of high-grade components or detects high-risk pathological features highly correlated with them, such as poor differentiation, pleural metastasis, vascular cancer thrombus, and dissemination through the air cavity, the frozen section is considered to be successfully predicted and marked as 1 (with a probability of 100%), otherwise it is marked as 0 (with a probability of 0).

8. The construction method according to claim 1, characterized in that: In step S4, the recognition principle of the parallel diagnosis model is that if there is a result of imaging omics or intraoperative freezing indicating that the lesion contains high-grade components, then the lesion is considered to contain high-grade components.

9. The construction method according to claim 1, characterized in that: In step S4, the R-FS model is constructed by combining the R model prediction probability and the intraoperative frozen section probability with a model fusion method, and the two are given the same weight. According to the parallel diagnosis principle, 0.25 is used as the cutoff value. If the final P>=0.25, it is considered that the R-FS model predicts a positive result for the lesion.

10. A precise surgical guidance system for lung cancer during surgery based on radiomics and frozen section multidisciplinary diagnosis, characterized in that: include: Preoperative imaging data acquisition module: used to obtain radiomics data through imaging scanning methods such as CT or MRI before surgery as one of the input data sources of the model; Intraoperative frozen section diagnosis module: used to obtain tissue samples from the patient's tumor site during surgery for frozen section (FS) analysis to obtain real-time pathological information; R-FS model prediction module: used to integrate the pathological diagnosis results of radiomics and frozen sections through the R-FS model to provide accurate diagnostic support for the prediction of high-grade pattern (HGP) of lung adenocarcinoma before and during surgery; the R-FS model is a high-grade pattern prediction model of lung adenocarcinoma constructed by the construction method described in any one of claims 1 to 9.

Images