A non-invasive detection method and device for Ki67 expression level in lung adenocarcinoma based on deep imaging omics

Abstract

The present invention discloses a non-invasive detection method and device for the expression level of Ki67 in lung adenocarcinoma based on deep imaging omics. The following scheme is proposed, which includes the following steps: S1: data preprocessing: obtaining basic clinical information and CT images of patients with lung adenocarcinoma; S2: extracting deep learning features of CT images: extracting deep imaging omics features from CT images, feeding back the region of interest into the network, and adaptively learning potential features in a supervised manner; S3: extracting imaging omics features: extracting imaging omics features from tumor images using a hand-crafted radiomic signature method (HCR); S4: constructing a prediction model using imaging features fused with multi-source features: fusing deep learning features, hand-crafted imaging omics features and clinical features, connecting and fusing using three FC layers, and the softmax function will output the probability of the Ki67 state. The present invention can non-invasively detect the state of Ki67 in lung adenocarcinoma, and assist clinicians in making clinical decisions for patients with lung adenocarcinoma.

Description

Technical Field

[0001] The present invention relates to a method and device for non-invasive detection of Ki67 expression level in lung adenocarcinoma based on deep imaging omics. Background Art

[0002] Lung adenocarcinoma is one of the most common histopathological diseases and a logical subtype of non-small cell lung cancer (NSCLC), accounting for 85% of all lung cancers. As a marker of tumor cell proliferation, Ki67 has been widely used in cancer research and clinical applications.

[0003] Increasing evidence shows the clinical value of Ki67 as a predictor of efficacy and is expected to become an important molecular marker for the diagnosis and prediction of tumors. It has been reported that high Ki67 expression is an indicator of poor prognosis in patients with non-small cell lung cancer. Therefore, the Ki67 index can be used clinically to measure the degree of tumor proliferation and predict survival and recurrence. However, it is usually necessary to sample tumor tissue to determine the Ki67 index, which is invasive for patients. Therefore, it is necessary to design a non-invasive identification method for the Ki67 index. More and more studies have shown that computed tomography (CT), a conventional and non-invasive cancer diagnostic method technology, can provide potential features associated with molecular markers of lung adenocarcinoma. However, there are few studies on the prediction of Ki67 status in lung adenocarcinoma, and the concept behind Ki67 has been found in other cancer studies or other applications.

[0004] Ki67 is a promising molecular target for diagnosing lung adenocarcinoma. However, conventional methods for determining the Ki67 index often require sampling of tumor tissue, which is invasive for patients. In this study, we proposed a hybrid radiomics signature, namely HC-DLR, to non-invasively predict Ki67 status by fusing multi-source features, including deep learning radiomics (DLR) features, handcrafted radiomics (HCR) features, and clinical features. Specifically, DLR features were adaptively extracted from CT images using a pre-trained MobileNetV3 model. HCR features were refined from a large number of handcrafted radiomic features extracted from the gross tumor volume on CT images. Finally, multi-source features were concatenated and fused to construct an HC-DLR model to predict Ki67 status in an automatic and non-invasive manner. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a non-invasive detection method and device for the Ki67 expression level of lung adenocarcinoma based on deep imaging genomics.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A non-invasive detection method and device for Ki67 expression level in lung adenocarcinoma based on deep imaging omics, comprising the following steps:

[0008] S1: Data preprocessing: Obtain basic clinical information and CT images of patients with lung adenocarcinoma;

[0009] S2: Extract deep learning features from CT images: Extract deep radiomics features from CT images, feed the regions of interest back into the network, and adaptively learn latent features in a supervised manner;

[0010] S3: Extract radiomic features: A handcrafted radiomic signature method (HCR) was used to extract radiomic features from tumor images;

[0011] S4: Constructing a prediction model using imaging features fused from multiple sources: Deep learning features, handcrafted radiomics features, and clinical features are integrated and connected using three FC layers, and the softmax function outputs the probability of Ki67 status.

[0012] Preferably, in S1, the original CT image and the corresponding image ROI mask are used to extract radiation features, and the adjacent CT selects a slice centered on the largest tumor area and resizes it to a shape of 224×224×3 to extract deep radiation features.

[0013] Preferably, in S1, the extracted data is randomly divided into two parts, 80% as a training set and 20% as a test set.

[0014] Preferably, in S2, 10% of the training data is randomly divided into a validation set for guiding hyperparameter selection.

[0015] Preferably, in S2, the Adam optimizer is used to initialize the decay learning rate at 1e-4 to optimize the model parameters. Finally, the cross entropy is used as the loss function to calculate the loss between the output and the label. In order to alleviate overfitting and category imbalance, online data enhancement techniques such as random horizontal and vertical flipping, random cropping, etc. are used, and each category is input into the network with the same probability.

[0016] Preferably, in S3, the extracted radiation features include first-order features, shape features, grayscale co-occurrence matrix features, grayscale run length matrix features, grayscale size area matrix features, adjacent grayscale difference matrix features and grayscale correlation matrix.

[0017] Preferably, in S4, independent samples are used for evaluation of continuous variables and categorical variables, and t-test and chi-square test are performed respectively.

[0018] Preferably, in S4, these features are used to establish a clinical model for detecting Ki67 status.

[0019] Preferably, in S4, the image feature set of lung adenocarcinoma patients with unknown Ki67 levels is input into the model designed by us to perform Ki67 level detection.

[0020] Beneficial effects of the present invention:

[0021] (1) The present invention proposes a non-invasive detection method and device for the Ki67 expression level in lung adenocarcinoma based on deep imaging genomics. The DLR model is established through HCR, and its predictive performance is studied;

[0022] (2) A method and device for non-invasive detection of Ki67 expression level in lung adenocarcinoma based on deep imaging omics proposed in the present invention, multi-source finite element analysis derived from the fusion of features of DLR, HCR and demographics to establish a HC-DLR model to improve the prediction performance;

[0023] (3) The present invention proposes a non-invasive detection method and device for Ki67 expression level in lung adenocarcinoma based on deep imaging genomics. A large data set of 661 patients was collected to evaluate and verify the effectiveness of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of the workflow of a non-invasive detection method and device for Ki67 expression level in lung adenocarcinoma based on deep imaging omics proposed in the present invention;

[0025] Figure 2 The patient data statistical characteristics of a non-invasive detection method and device for Ki67 expression level in lung adenocarcinoma based on deep imaging omics proposed in the present invention;

[0026] Figure 3 A schematic diagram showing the performance comparison between a non-invasive detection method and device for Ki67 expression level in lung adenocarcinoma based on deep imaging omics proposed in the present invention and other detection methods;

[0027] Figure 4 A schematic diagram of the importance of the first 20 radiomics features calculated using a random forest algorithm in a method and device for non-invasive detection of Ki67 expression level in lung adenocarcinoma based on deep radiomics proposed by the present invention;

[0028] Figure 5 Schematic diagram of the ROC curves of the clinical model, HCR, DLR and HC-DLR models of the non-invasive detection method and device of the Ki67 expression level of lung adenocarcinoma based on deep imaging omics proposed in the present invention;

[0029] Figure 6A schematic diagram of the signature score distribution of a non-invasive detection method and device for Ki67 expression level in lung adenocarcinoma based on deep imaging omics proposed in the present invention; DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0031] Reference Figure 1-6 , a non-invasive method for detecting the Ki67 expression level of lung adenocarcinoma, comprising the following steps:

[0032] S1: Data preprocessing: Extract the basic clinical information and CT images of patients with lung adenocarcinoma, as well as the Ki67 index result data of postoperative pathological examination; Finally, 661 patients diagnosed with lung adenocarcinoma were used for this study. Among all these patients, 229 patients had Ki67 index, and 432 patients did not. The summary of these data is as follows Figure 2 All regions of interest (ROIs) were determined by the clinician using IBEX software at the slice level in the axial view.

[0033] S2: Extract deep learning features from CT images: Extract deep radiomic features from CT images, feed the region of interest back into the network, and adaptively learn the latency to present features in a supervised manner; this study uses a transfer learning network MobileNetV3smallpre, trained on Imagenet, especially the penultimate FC layer with 1024 neurons is used as deep radiomic features. Finally, these features are input into three FC layers and a softmax layer for predicting Ki67 status. During training, a rectangular ROI is input into the network to update the model parameters through back propagation, and the batch size is set to 32 and the maximum number of epochs is set to 1000.

[0034] S3: Extraction of radiomic features: Handcrafted radiomics (HCR) has been widely used in many different precision medicine applications, such as cancer diagnosis and prognosis. In this study, HCR features were extracted using an in-house developed software. We mainly extracted seven radiomic features, including first-order features, shape-based features, gray-level co-occurrence matrix (GLCM) features, gray-level run length matrix (GLRLM) features, gray-level size zone matrix (GLSZM) features, neighboring gray-level difference matrix (NGTDM) features, and gray-level correlation matrix (GLDM) features. Finally, a total of 2153 quantitative HCR features were extracted from each total tumor volume. To reduce the feature dimensionality, we used common feature reduction methods such as the Mann-Whitney statistical test, variance threshold, and least absolute shrinkage and selection operator (LASSO) to select 44 important features. These features were also used to construct the HCR signature by using a support vector machine (SVM) classifier for Ki67 prediction.

[0035] S4: Constructing a prediction model using imaging features fused with multi-source features: Deep features, manual features, and clinical features are connected and fused by using three FC layers, and the softmax function will output the probability of Ki67 status; The purpose of fusing multi-source feature representations is to improve the prediction performance of radiomic features. In particular, deep features, manual features, and clinical features are connected and fused by using the following three FC layers, which are composed of 512, 64, and 2 neurons. Finally, the softmax function will output the probability of Ki67 status.

[0036] Experimental results show that the proposed method has an AUC of 0.78, an accuracy of 0.77, a sensitivity of 0.72, and a specificity of 0.79, and has potential clinical value.

[0037] To illustrate the superiority, we compare our proposed method with other deep transfer learning methods, such as Figure 3 As shown in Figure 3, we can see that the prediction performance of our proposed HC-DLR model is better than other models. Although these comparative models can adaptively learn features related to Ki67 status from CT images, they do not consider the use of multi-source heterogeneous data. The encouraging performance of our HC-DLR indicates that multi-source features play an important role in predicting Ki67 status.

[0038] Although our study and comparative experiments showed that our HC-DLR model achieved better prediction performance, there are some limitations in this study in predicting Ki67 status. First, the data used in this study originated from a single institution, which may be biased in the selection of patients. In the future, we will collect external validation datasets from multiple institutions to evaluate the generalization performance. Second, CT images with 5-mm slice thickness were used in this study. CT images with thin slices or different slice thicknesses deserve further study in predicting Ki67 status. Third, our study only focused on Ki67-positive and Ki67-negative predictions. Next, we will further study the prediction of Ki67-high and Ki67-low expression. In summary, we proposed a hybrid radiomics feature to predict Ki67 status by using multi-source feature representation in an automatic and non-invasive manner.

[0039] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0040] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A non-invasive detection method for Ki67 expression level in lung adenocarcinoma based on deep imaging omics, characterized in that: The following steps are involved: S1: Data preprocessing: Obtain basic clinical information and CT images of patients with lung adenocarcinoma; S2: Extract deep learning features from CT images: Extract deep radiomics features from CT images, feed the regions of interest back into the network, and adaptively learn latent features in a supervised manner; S3: Extraction of radiomic features: A handcrafted radiomic signature method (HCR) was used to extract radiomic features from tumor images, including first-order features, shape features, gray-level co-occurrence matrix features, gray-level run length matrix features, gray-level size region matrix features, adjacent gray-level difference matrix features, and gray-level correlation matrix; S4: Constructing a prediction model using imaging features fused from multiple sources: Deep learning features, handcrafted radiomics features, and clinical features are integrated and connected using three FC layers, and the softmax function outputs the probability of Ki67 status.

2. The non-invasive detection method of Ki67 expression level in lung adenocarcinoma based on deep imaging omics according to claim 1, characterized in that: In S1, the original CT image and the corresponding image ROI mask are used to extract radiation features, and the adjacent CT selects a slice centered on the largest tumor area and resizes it to a shape of 224×224×3 to extract deep radiation features.

3. The non-invasive detection method of Ki67 expression level in lung adenocarcinoma based on deep imaging omics according to claim 1, characterized in that: In S1, the extracted data is randomly divided into two parts, 80% as a training set and 20% as a test set.

4. The non-invasive detection method of Ki67 expression level in lung adenocarcinoma based on deep imaging omics according to claim 1, characterized in that: In S2, 10% of the training data is randomly divided into a validation set for guiding hyperparameter selection.

5. The non-invasive detection method of Ki67 expression level in lung adenocarcinoma based on deep imaging omics according to claim 1, characterized in that: In S2, the Adam optimizer is used to initialize the decaying learning rate at 1e-4 to optimize the model parameters. Finally, the cross entropy is used as the loss function to calculate the loss between the output and the label.

6. The non-invasive detection method of Ki67 expression level in lung adenocarcinoma based on deep imaging omics according to claim 1, characterized in that: In S4, continuous variables and categorical variables were evaluated using independent samples, using t-test and chi-square test, respectively.

7. The non-invasive detection method of Ki67 expression level in lung adenocarcinoma based on deep imaging omics according to claim 1, characterized in that: In S4, these features were used to build a clinical model for predicting Ki67 status.

8. The non-invasive detection method for Ki67 expression level in lung adenocarcinoma based on deep imaging omics according to claim 1, characterized in that: In S4, the image feature set of unknown Ki67 levels of lung adenocarcinoma patients is input into the designed model, and the Ki67 levels are detected.

Images