A method, apparatus and program product for predicting adrenal lesions based on chest ct

Abstract

The application relates to the field of intelligent medical treatment, in particular to a method, equipment and program product for predicting adrenal gland lesions based on chest CT. The method comprises the following steps: S1, acquiring chest CT data; S2, inputting the chest CT data into a segmentation network to obtain a region of interest; and S3, inputting the region of interest into a classifier to obtain a prediction result of adrenal gland normality or abnormality. The application can predict the adrenal gland normality or abnormality through the chest CT, provides a new and convenient way, and has good clinical value.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical care, and specifically to a method, device, program product, and computer-readable storage medium for predicting adrenal lesions based on chest CT. Background Art

[0002] Adrenal disease accounts for approximately 5% of patients with hypertension. However, with advances in testing technology and increased physician awareness of secondary hypertension, the proportion of adrenal hypertension is now believed to have risen to over 20%, and even exceeds 30% in patients with refractory hypertension. Most adrenal diseases can be prevented, controlled, or even cured through surgical resection. Therefore, early screening for adrenal disease is crucial for reducing complications and improving quality of life.

[0003] However, there is currently no effective screening method for adrenal disease. Due to the small size and deep location of the adrenal glands, ultrasound examinations often have difficulty accurately detecting the adrenal region. Although abdominal enhanced CT has a high detection rate, it is expensive, involves high radiation doses, carries the risk of contrast agent allergy, and the examination procedures are relatively complex. These unfavorable factors result in many cases not being diagnosed in a timely manner, thus affecting the detection rate and treatment efficacy of adrenal disease, and causing patients to often face the serious consequences of hypertension by the time they seek medical treatment. Summary of the Invention

[0004] In response to the above problems, the present invention proposes a method for predicting adrenal lesions based on chest CT. As a routine physical examination item, chest CT scans from the apex to the base of the lungs. In addition to being able to detect lung diseases such as lung nodules, it can also identify adrenal abnormalities. Using chest CT to screen for adrenal nodules can not only reduce radiation dose, but also reduce medical costs, becoming a preferred screening strategy that "kills two birds with one stone". And with the continuous advancement of artificial intelligence technology in the field of medical imaging, the development of specialized software based on deep learning models for screening adrenal nodules in chest CT can effectively reduce medical costs and improve work efficiency, thereby providing timely and accurate medical services to more people. The method for predicting adrenal lesions based on chest CT specifically includes:

[0005] S1. Obtain chest CT data;

[0006] S2. Inputting the chest CT data into a segmentation network for segmentation to obtain a region of interest;

[0007] S3. Inputting the region of interest into a classifier for prediction to obtain a prediction result of whether the adrenal gland is normal or abnormal.

[0008] Furthermore, the chest CT includes one or more of the following: low-dose chest CT, plain scan CT, enhanced CT, and high-resolution CT;

[0009] Optionally, the chest CT is a low-dose CT;

[0010] Optionally, the segmentation includes one or more of the following: a two-dimensional segmentation network, a three-dimensional segmentation network;

[0011] Optionally, the two-dimensional segmentation network includes one or more of the following: UNet, FCN, SegNet, Enet, DeepLabV3+, DSN;

[0012] Optionally, the three-dimensional segmentation network includes one or more of the following: 3D Vnet, 3D UNet, 3D FCN, AtlasNet, DenseVoxNet, nnUNet, 3D ResNet / ResNeXt.

[0013] The S2 is replaced by S2': inputting the chest CT data into a segmentation network for segmentation to obtain a target region, performing region expansion based on the target region to generate a region bounding box, and performing segmentation based on the region bounding box to obtain a region of interest.

[0014] The training process of the three-dimensional segmentation network is:

[0015] Obtain chest CT dataset and adrenal gland region labels;

[0016] Inputting the data set and labels into the three-dimensional segmentation network to be trained to obtain a region recognition result;

[0017] Compare the region recognition results with the adrenal gland region label to obtain the loss function;

[0018] The 3D segmentation network is optimized based on the loss function, and the above steps are repeated until the loss function remains unchanged to obtain the 3D segmentation network.

[0019] The three-dimensional segmentation network performs region of interest segmentation through 3D Vnet, which includes an encoder and a decoder. The chest CT data is downsampled N times by the encoder and the feature map obtained by each downsampling is output, where N is a natural number greater than 1. The feature map obtained by the Nth downsampling is input to the decoder, and the decoder performs N upsampling to obtain the region of interest, and the downsampled feature map is input to the decoder through a jump connection.

[0020] The classifier makes predictions through a pre-trained deep learning network;

[0021] Optionally, the pre-trained deep learning network includes one or more of the following: Med3D, MTCNN, DenseVNet, DGCNN, VoteNet, ShellNet;

[0022] Optionally, the pre-trained deep learning network is a Med3D deep learning network;

[0023] Optionally, the decoding module of the Med3D deep learning network is replaced by L convolution modules and fully connected layers, where L is a natural number greater than 1, and the convolution module includes a maximum pooling layer and a convolution layer;

[0024] Optionally, the training process of the pre-trained deep learning network is:

[0025] Obtaining a set of regions of interest and labels of whether the adrenal glands are normal or abnormal;

[0026] The set of regions of interest and labels of normal or abnormal adrenal glands are input into a pre-trained deep learning network to be trained to obtain a pre-trained deep learning network. The pre-trained deep learning network is parameter-adjusted and weight-optimized based on the input data.

[0027] The method further includes data preprocessing, wherein the chest CT data is preprocessed to obtain processed data, and the processed data is input into a three-dimensional segmentation network;

[0028] Optionally, the data preprocessing includes: data enhancement, data resampling, and data standardization.

[0029] The object of the present invention is to provide a computer product, including a computer program or instructions, which are executed by a processor to implement the above-mentioned method for predicting adrenal lesions based on chest CT.

[0030] The object of the present invention is to provide a computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-mentioned method for predicting adrenal lesions based on chest CT.

[0031] The present invention is directed to a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the above-mentioned method for predicting adrenal lesions based on chest CT.

[0032] Advantages of the present invention:

[0033] 1. Chest CT is routinely used for health checkups and lung cancer screening in the community, and its scan area covers the adrenal glands. This invention utilizes chest CT to detect incidental adrenal lesions, improving the detection rate of adrenal abnormalities without increasing medical costs or radiation dose.

[0034] 2. This invention uses artificial intelligence deep learning technology to achieve automatic segmentation and classification of chest CT scans, assisting clinicians in automatically assessing the adrenal gland region after the scan. When abnormalities are detected, it prompts the doctor to conduct further examinations, reducing the doctor's workload and improving work efficiency.

[0035] 3. Adrenal disease can lead to secondary hypertension and refractory hypertension, but some patients fail to receive further treatment due to missed diagnosis, leading to serious complications. This invention aims to improve the detection rate of adrenal abnormalities, enabling early diagnosis and treatment, thereby reducing long-term complications and has significant clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 A schematic flow chart of a method for predicting adrenal lesions based on chest CT provided by an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of a system for predicting adrenal lesions based on chest CT provided by an embodiment of the present invention;

[0039] Figure 3 A schematic diagram of a device for predicting adrenal lesions based on chest CT provided by an embodiment of the present invention;

[0040] Figure 4Schematic diagram of non-enhanced CT adrenal abnormalities provided in an embodiment of the present invention, with abnormal areas in the adrenal region indicated by arrows. (A) Adrenal adenoma: a local soft tissue density lesion in the adrenal region, roughly circular in shape, with clear and uniform boundaries, and an average CT value of 0-10HU. (B) Non-specific mass or nodule in the adrenal region: a local soft tissue lesion in the adrenal region, with clear or unclear boundaries, and a CT value of 10-40HU. (C) Adrenal myelolipoma: a local fat density lesion in the adrenal region, with a CT value ranging from -20 to -100HU. (D) Cystic lesion in the adrenal region: a local water density lesion in the adrenal region, roughly circular in shape, with clear boundaries, and a CT value of approximately 0HU. (E) Adrenal calcification: a single or multiple high-density lesions in the adrenal region, usually small, appearing as dots, lines or plaques, with clear boundaries, and may involve the entire adrenal gland, with a CT value exceeding 100HU. (F) Postoperative changes in the adrenal area: The normal adrenal glands are not clearly visible, and short linear high-density objects can be seen locally, which are postoperative metal sutures or clips. (G) Adrenal hemorrhage: The acute phase manifests as diffuse enlargement and increased density of the adrenal glands, with a CT value of 50-80HU. (H) Diffuse adrenal hyperplasia: The adrenal glands are diffusely enlarged, which can be an increase in the volume and thickness of the entire adrenal gland, or an increase in the volume of the medial or lateral limbs. The boundaries are clear, there is no focal abnormal density, and no masses or nodules are formed. (I) Adrenal nodular hyperplasia: One or more nodules appear in the adrenal glands, with equal density or slightly lower density. The shape of the adrenal glands is irregular and can occur in any part of the adrenal glands. They can exist in a single adrenal gland or both adrenal glands;

[0041] Figure 5 A data flow diagram provided for an embodiment of the present invention;

[0042] Figure 6 Evaluation indicators of the segmentation model provided in the embodiment of the present invention, A represents the DSC results of the test set, internal validation set and external validation set. The first row shows the DSC value of the left adrenal gland, which is divided into a normal group and an abnormal group. The second row provides the DSC value of the right adrenal gland, using the same classification method; the third row shows the DSC values ​​of all adrenal glands; the first column lists the DSC values ​​of the left and right adrenal glands in the test set; the second column details the DSC values ​​in the internal validation cohort; the third column shows the DSC values ​​of all adrenal glands; B represents volume similarity (VS), and C represents the average Hausdorff distance (HD) statistic; In Val: internal validation; Exter Val: external validation; DSC: Dice similarity coefficient; VS: volume similarity; Hd: Hausdorff distance;

[0043] Figure 7Model training results provided by an embodiment of the present invention; AC shows the confusion matrix of the AI ​​model in the test set, internal validation set, and external validation set. DF shows the confusion matrix of the radiology report in the test set, internal validation set, and external validation set. GI shows the ROC curves of the AI ​​model and radiology report in the test set, internal validation set, and external validation set. JL shows the PR curves of the AI ​​model and radiology report in the test set, internal validation set, and external validation set. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0045] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as S101, S102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, nor do they limit "first" and "second" to be different types.

[0046] Figure 1 A schematic diagram of a method for predicting adrenal lesions based on chest CT provided in an embodiment of the present invention specifically includes:

[0047] S1: Acquire chest CT data;

[0048] In one embodiment, the chest CT includes one or more of the following: low-dose chest CT, plain scan CT, enhanced CT, and high-resolution CT.

[0049] In one embodiment, the chest CT is a low-dose CT.

[0050] In one embodiment, the concept of low-dose CT is: when other conditions remain unchanged, the tube current is reduced to reduce radiation, while ensuring the image diagnosis requirements. Its radiation dose is about 1 / 5 of that of ordinary CT, about 1 mSv. Low-dose CT is not a special machine, but is made by changing the parameters of an ordinary CT machine (generally a 64-row CT). While reducing the radiation dose, it also loses some clarity, so it is mainly used for screening people at high risk of lung cancer. LDCT, also known as low-dose spiral CT, has a radiation dose of 1 / 4-1 / 3 of that of ordinary CT, so it has a higher safety factor and is more suitable for physical examinations. Cheap: LDCT is relatively cheaper than ordinary CT. Low resolution: The image resolution of LDCT is lower than that of ordinary CT.

[0051] In one embodiment, the method further comprises data preprocessing, wherein the chest CT data is preprocessed to obtain processed data, and the processed data is input into a three-dimensional segmentation network;

[0052] Optionally, the data preprocessing includes: data enhancement, data resampling, and data standardization.

[0053] In one specific embodiment, the data of the present invention include four different cohorts. Cohort A is from Hospital 1 and includes chest LDCT scans collected between March 2018 and December 2021. Given the limited number of abnormal cases in Cohort A, in order to increase the sample size of patients with adrenal abnormalities, Cohort B was also included, which included patients from Hospital 1 who underwent abdominal CT scans from October 2021 to December 2021 and whose radiology reports showed adrenal abnormalities. The data from Cohorts A and B were used for the development of the AI ​​model. In addition, Cohort C from Hospital 1 includes consecutive chest LDCT scans from July 2022 to April 2023. Cohort D is from Hospital 2 and has the same inclusion criteria as Cohort C. Cohorts C and D are designated for internal and external validation of the AI ​​model.

[0054] Exclusion criteria included cases where CT images could not be uploaded, duplicate patient examinations, significant upper abdominal artifact obscuring visualization, and cases where the adrenal glands could not be identified on CT images. Incomplete coverage of the adrenal glands on chest CT images was also excluded. For inclusion in this study, patients were required to have complete visualization of both adrenal glands within a maximum slice thickness of 5 mm. Overall, this study included 2574, 1205, 3681, and 779 patients in the four cohorts, respectively, with no overlap between them.

[0055] In one embodiment, all CT images were acquired using 12 scanners from five different manufacturers. The model development cohort was acquired from scanners from four vendors, and the internal validation cohort was acquired from scanners from three vendors. The external validation cohort was acquired from scanners from two vendors. The model development dataset had a tube current of 133.0 (IQR: 101.0-300.0) mA, a peak tube voltage of 100 (IQR: 100-120) kV, and a slice thickness of 5.00 (IQR: 5.00-5.00) mm. For the internal validation cohort, the tube current was 98.0 (IQR: 77.0-120.0) mA, the peak tube voltage was 100 (IQR: 100-100) kV, and the slice thickness was 5.00 (IQR: 5.00-5.00) mm. In the external validation cohort, the tube current was 78.0 (IQR: 65.0-92.5) mA, the peak tube voltage was 120 (IQR: 110-120) kV, and the slice thickness was 1.50 (IQR: 1.50-1.50) mm. All scanning parameters were statistically significant between the model development cohort, the internal validation cohort, and the external validation cohort (all P < 0.001).

[0056] In one embodiment, two experienced urological radiologists independently read and analyze the CT images to determine whether there are abnormalities in the adrenal glands. The evaluation criteria for adrenal abnormalities focus on whether there are isolated or combined density and morphological abnormalities in the adrenal glands, such as Figure 4 Abnormal adrenal density included calcification, hemorrhage, cysts, fatty infiltration, or density lower than normal adrenal tissue. Morphological abnormalities included diffuse adrenal enlargement, multiple nodules, solitary nodules or masses, and changes suggestive of previous surgical intervention. When the two experts disagreed, consensus was reached through mutual consultation.

[0057] S2: Inputting the chest CT data into a segmentation network for segmentation to obtain a region of interest;

[0058] In one embodiment, the segmentation includes one or more of the following: a two-dimensional segmentation network, a three-dimensional segmentation network.

[0059] In one embodiment, the two-dimensional segmentation network includes one or more of the following: UNet, FCN, SegNet, Enet, DeepLabV3+, and DSN.

[0060] In one embodiment, the training process of the two-dimensional segmentation network is:

[0061] Obtain chest CT dataset and adrenal gland region labels;

[0062] Inputting the data set and labels into the two-dimensional segmentation network to be trained to obtain a region recognition result;

[0063] Compare the region recognition results with the adrenal gland region label to obtain the loss function;

[0064] Optimize the two-dimensional segmentation network based on the loss function, repeat the above steps until the loss function remains unchanged, and obtain the two-dimensional segmentation network.

[0065] In one embodiment, the three-dimensional segmentation network includes one or more of the following: 3D Vnet, 3DUNet, 3DFCN, AtlasNet, DenseVoxNet, nnUNet, 3D ResNet / ResNeXt.

[0066] In one embodiment, the training process of the 3D segmentation network is:

[0067] Obtain chest CT dataset and adrenal gland region labels;

[0068] Inputting the data set and labels into the three-dimensional segmentation network to be trained to obtain a region recognition result;

[0069] Compare the region recognition results with the adrenal gland region label to obtain the loss function;

[0070] The 3D segmentation network is optimized based on the loss function, and the above steps are repeated until the loss function remains unchanged to obtain the 3D segmentation network.

[0071] In one embodiment, a three-dimensional segmentation network performs region-of-interest segmentation using a 3D Vnet. The 3D Vnet includes an encoder and a decoder. The encoder downsamples chest CT data N times and outputs a feature map obtained by each downsampling, where N is a natural number greater than 1. The feature map obtained by the Nth downsampling is input to the decoder, which then upsamples the Nth-encoded feature map N times to obtain a region of interest. The downsampled feature map is then input to the decoder via a skip connection. The encoder and decoder utilize a skip connection to fuse the feature maps, and the feature dimensions of the encoding layer of the skip connection are the same as those of the decoding layer.

[0072] In one embodiment, S2 is replaced by S2': inputting the chest CT data into a segmentation network for segmentation to obtain a target region, performing region expansion based on the target region to generate a region bounding box, and performing segmentation based on the region bounding box to obtain a region of interest.

[0073] In a specific embodiment, the workflow of the deep learning algorithm of the present invention is as follows: Figure 5As shown, it includes two main stages: adrenal segmentation and adrenal classification. The first stage involves the segmentation of the adrenal glands, in which the 3D VNet deep learning network is utilized. The input data consists of a complete CT image (A) and the bilateral adrenal regions labeled by experts (B). The output of the segmentation model is the bilateral adrenal region prediction (C). The second stage is the classification of the adrenal glands, using the Med3D deep learning model. The adrenal region predicted from the segmentation model is used as a mask (D) to crop the CT image, producing the adrenal bounding box (E), which then becomes the input of the Med3D model. The classification label of the adrenal gland is marked as normal or abnormal (F). The cropped CT image and its label are used as input for the Med3D model, and the output is the probability that the adrenal gland is abnormal (G).

[0074] All training was performed on an NVIDIA Tesla P100 16G GPU. The algorithms were coded using Python 3.6, PyTorch 0.4.1, OpenCV 3.4.0.12, NumPy 1.16.2, and SimpleITK 1.2.0. Deep learning models were developed using the model development dataset, which was randomly partitioned into three subsets: training, validation, and test, with an approximate ratio of 8:1:1. To ensure evaluation integrity, data from the same patient were assigned to a separate subset. During training, the training set was used for gradient descent optimization, and the validation set was used for hyperparameter tuning and model selection. A test set was reserved for evaluating model performance after training.

[0075] In the first step, a three-dimensional deep neural network called 3D VNet was used. The network consists of an encoder and a decoder to accurately locate the location of the bilateral adrenal glands. The encoder performs four downsampling operations to construct feature maps, which are then fused in the decoder via skip connections. The decoder upsamples from the variables and concatenates them with the previous encoder feature maps. The input of the 3D VNet consists of the entire CT image, which is preprocessed before being fed into the network. To ensure the consistency of the data shape and voxel spacing, the image is resampled to a fixed resolution of 256×192×96 (x×y×z). The window width is set to 300 and the window center is set to 30. In addition, the voxel values ​​are normalized using z-score normalization. The training set data is augmented by applying random rotations (with a maximum angle of 10°), introducing random noise, and performing parallel translations within the range of [-0.1, 0.1] pixels on the horizontal and vertical axes. The loss function of the 3D VNet is the Dice loss, which is well suited to address the imbalanced classes often encountered in medical imaging. The Dice coefficient D is defined as:

[0076]

[0077] Where A represents the predicted segmentation and B represents the true value. The Dice loss LDice is defined as:

[0078]

[0079] where pi is the predicted probability of each voxel and gi is the true label.

[0080] S3: Inputting the region of interest into a classifier for prediction to obtain a prediction result of whether the adrenal gland is normal or abnormal.

[0081] In one embodiment, the classifier makes predictions using a pre-trained deep learning network;

[0082] The pre-trained deep learning network includes one or more of the following: Med3D, MTCNN, DenseVNet, DGCNN, VoteNet, and ShellNet.

[0083] In one embodiment, the pre-trained deep learning network is a Med3D deep learning network; the decoding module of the Med3D deep learning network is replaced by L convolution modules and a fully connected layer, L is a natural number greater than 1, and the convolution module includes a maximum pooling layer and a convolution layer.

[0084] In one embodiment, the training process of the pre-trained deep learning network is:

[0085] Obtaining a set of regions of interest and labels of whether the adrenal glands are normal or abnormal;

[0086] The set of regions of interest and labels of normal or abnormal adrenal glands are input into a pre-trained deep learning network to be trained to obtain a pre-trained deep learning network. The pre-trained deep learning network is parameter-adjusted and weight-optimized based on the input data.

[0087] In one embodiment, the present invention proposes automatic segmentation of the entire adrenal gland without manual intervention, which, combined with automatic classification, can accurately identify adrenal abnormalities. This approach offers advantages over the labor-intensive process of interpreting LDCT scans and may help reduce the rate at which radiologists overlook adrenal abnormalities. Therefore, the purpose of this study was to investigate the feasibility of automatic adrenal gland segmentation and classification using a deep learning model to detect adrenal abnormalities in LDCT imaging.

[0088] In one specific embodiment, in the second step, the present invention employs a deep learning network called Med3D. This framework leverages the power of transfer learning to analyze 3D medical images. Med3D consists of a 3D convolutional neural network pre-trained on a wide range of medical image datasets. For the present application, the present invention fine-tunes this pre-trained model. Through transfer learning, features are extracted from the image using the encoder weights. The present invention retains the encoder component and replaces the decoder (deconvolution portion) of the network with convolutional and fully connected layers from the classification network architecture. The classification convolutional layer has the following structure: 1) a max pooling layer with a stride of 2; 2) a convolutional layer with a kernel size of 3; 3) another max pooling layer with a stride of 2; 4) a second convolutional layer with a kernel size of 3. The fully connected layer of the classification network consists of 128 neurons that integrate and classify the image features. The output is then calculated and the classification result is presented through a SoftMax function, which generates a probability distribution over the different classes. The input to the second stage is the bounding box derived from the adrenal gland mask extracted in the first step. These masks are used to crop bounding boxes from the entire CT image in preparation for the input to the second stage. Med3D then predicts that each adrenal gland is either normal or abnormal. The output of the second stage is a probability score indicating the likelihood of abnormality, with values ​​ranging from 0 to 1. The loss function used by Med3D is the binary cross entropy loss, which is defined as follows:

[0089] L BCE = -[ylog(p)+(1-y)log(1-p)]

[0090] Where y is the true label (0 or 1) and p is the predicted probability that the instance is in class 1.

[0091] The present invention uses the Adam optimizer with a learning rate of 0.01-0.0001 to update the parameters of the network. To prevent overfitting, early stopping is implemented and the final model is selected based on the criterion that the loss does not decrease in 50 consecutive iterations.

[0092] In a specific embodiment, the performance of the model on the test set was evaluated from the model development dataset, the internal validation cohort, and the external validation cohort. For adrenal segmentation, the Dice score (DSC), volume similarity (VS), and average Hausdorff distance (HD) were used to evaluate the performance of the model. For adrenal classification, evaluation indicators such as the receiver operating characteristic (ROC) curve, precision-recall (PR) curve, and confusion matrix were used. The present invention also collected the radiology reports of these patients and extracted the description of the adrenal glands from the reports. The output of the model and the radiology reports were compared with the actual situation. The present invention uses integrated discrimination improvement (IDI) to evaluate the AI ​​model and improves the diagnostic accuracy of the radiology reports.

[0093] In one embodiment, statistical analysis was performed using R4.1.3 software. Normally distributed continuous variables were expressed as mean ± standard deviation, and non-normally distributed variables were expressed as median (interquartile range, IQR). Categorical variables were reported as frequency and percentage. The Kolmogorov-Smirnov (KS) test was used to assess the normality of the distribution, and the F test was used to assess the homogeneity of variance. The Mann-Whitney U test or the Kruskal-Wallis test was used to compare non-normally distributed data or groups with unequal variance. Multiple comparisons were corrected using the Bonferroni method. The DeLong test was used to compare the AUC. The Z statistic was used to test whether the IDI was significantly different from 0. A P value of less than 0.05 was considered statistically significant.

[0094] In one embodiment, Table 1 presents detailed demographic data and adrenal status.

[0095] In the model development dataset, the median age was 60.0 years (IQR: 49.0-69.0), and 2199 (58.2%) participants were male. Bilateral adrenal abnormalities were found in 2268 (60.0%) participants, the left adrenal gland was abnormal in 759 (20.1%), the right adrenal gland was abnormal in 217 (5.7%), bilateral adrenal abnormalities were found in 475 (12.6%), and adrenal scans showed insufficiency in 60 (1.6%) participants. In the internal validation cohort, the median age was 47.0 years (IQR: 37.0-60.0), and 1814 (49.3%) participants were male. Adrenal glands were normal in 3602 patients (97.9%), abnormal left adrenal glands in 46 patients (1.2%), abnormal right adrenal glands in 17 patients (0.5%), bilateral adrenal gland abnormalities in 2 patients (0.1%), and adrenal insufficiency in 14 patients (0.4%). The median age of the external validation cohort was 60.0 years (IQR: 48.0-69.0), and 503 patients (64.6%) were male. Adrenal glands were normal in 646 patients (82.9%), abnormal left adrenal glands in 105 patients (13.5%), abnormal right adrenal glands in 14 patients (1.8%), bilateral adrenal gland abnormalities in 8 patients (1.0%), and adrenal insufficiency in 6 patients (0.8%). Statistical analysis was performed on adrenal gland morphology, diameter, volume, and CT scan. All clinical parameters were statistically significant among the three groups (P < 0.05).

[0096] Table 1

[0097]

[0098]

[0099] In a specific embodiment, the DSC, VS and HD values ​​of each data set and the results of the statistical tests are detailed in Tables 2 and Figure 6 The DSC values ​​for the test dataset, internal validation set, and external validation set were 0.839 (IQR: 0.783-0.871), 0.870 (IQR: 0.819-0.902), and 0.799 (IQR: 0.729-0.849), respectively. The DSC values ​​among the three groups were statistically significant (P<0.05). The VS values ​​of the three groups were 0.947 (IQR: 0.893-0.973), 0.952 (IQR: 0.910-0.978), and 0.914 (IQR: 0.840-0.963), respectively. The VS values ​​among the three groups were statistically significant (P<0.05). The mean HD values ​​were 0.226 mm (IQR: 0.158-0.374), 0.163 mm (IQR: 0.0989-0.302), and 0.354 mm (IQR: 0.223-0.616), respectively. The differences in HD values ​​among the three groups were statistically significant (P<0.05). Focusing on the DSC value, the present invention further compared the performance of the AI ​​model in segmenting normal and abnormal adrenal glands. In the test set, internal validation cohort, and external cohort, the median DSC of abnormal adrenal glands was smaller than that of normal adrenal glands (all P<0.05).

[0100] In addition, the present invention also compared the performance of the AI ​​model in segmenting the left and right adrenal glands. In the test set, the median DSC value of the right adrenal gland was greater than that of the left adrenal gland. In the internal validation cohort, the median DSC value of the left adrenal gland was greater than that of the right adrenal gland (P < 0.001). In the external validation cohort, the difference in median DSC values ​​between the left and right adrenal glands was not statistically significant (P = 0.511).

[0101] Figure 7 Confusion matrices, receiver operating characteristic (ROC) curves, and predictive value (PR) curves for adrenal gland classification using the AI ​​model and radiology reports in the test, internal, and external validation sets are presented. Table 3 details the AUC values, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, and negative likelihood ratio for classification in the three datasets. The DeLong test was used to compare the AUC values ​​of the AI ​​model and radiology reports for each dataset; the results are shown in Table 4.

[0102] Table 2

[0103]

[0104]

[0105] P 1: Statistical test results of the difference in DSC values ​​between normal and abnormal adrenal glands or between left and right adrenal glands. 2 : Statistical test results of the difference in VS values ​​between the left and right adrenal glands or between normal and abnormal adrenal glands. P 3 : Statistical test results of the difference in HD values ​​between normal and abnormal adrenal glands or between left and right adrenal glands. P 4 : Statistical test results of the differences in DSCs, VSs, and HDs among the test set, internal validation set, and external validation set.

[0106] Table 3

[0107]

[0108]

[0109] Table 4

[0110]

[0111] The AUC values ​​of the ROC curves of the AI ​​model in the test set, internal validation set, and external validation set were 0.956 (95% CI: 0.941-0.972), 0.942 (95% CI: 0.906-0.978), and 0.977 (95% CI: 0.956-0.997), respectively. The AUC values ​​of the radiology reports in the three data sets were 0.889 (95% CI: 0.861-0.916), 0.705 (95% CI: 0.650-0.760), and 0.551 (95% CI: 0.525-0.576), respectively. In all three data sets (all P < 0.05), the AUC values ​​of the AI ​​model were significantly greater than those of the radiology reports.

[0112] Given the significant imbalance in the number of normal and abnormal adrenal glands in the internal and external validation cohorts, the classification ability of the AI ​​model and radiology reports was further evaluated using PR curves. The PR AUC values ​​of the AI ​​model and radiology reports were 0.913 and 0.809 in the test set, 0.753 and 0.708 in the internal validation cohort, and 0.927 and 0.591 in the external validation cohort, respectively, all of which were statistically significant (P < 0.05).

[0113] The IDI was used to investigate the AI ​​model's enhancement of radiology reports. Using the radiology report as the reference, the net reclassification index (NRI) of the AI ​​model relative to the radiology report in the test set, internal validation cohort, and external validation cohort was -0.172 (95% CI: -0.261, -0.084), -0.821 (95% CI: -1.039 to -0.602), and 0.685 (95% CI: 0.610-0.761), respectively (all P < 0.001). The IDI of the AI ​​model relative to the radiology report was -0.130 (95% CI: -0.197, -0.063) in the test set, 0.221 (95% CI: 0.066-0.376) in the internal validation cohort, and 0.590 (95% CI: 0.523-0.657) in the external validation cohort (all P < 0.05).

[0114] Studies have reported that adrenal abnormalities are common in radiology reports but are often overlooked, which highlights the need for increased awareness in radiology. The AI ​​model proposed in this study can automatically detect abnormalities in the adrenal region from LDCT images with an AUC of 0.942-0.977, which exceeds the detection capability of radiology reports (AUC: 0.551-0.889, P < 0.05). This shows that the AI ​​model of the present invention can assist the work of radiologists. The envisioned application scenario is that the AI ​​model acts as the first reader to assist radiologists in automatically evaluating scans after the LDCT is completed. If the AI ​​model identifies an abnormality, it can remind radiologists to pay attention to the adrenal glands during image interpretation and reporting.

[0115] In recent years, advances in deep learning technology have led to remarkable progress in accurately delineating abdominal organs in medical imaging. However, automatic adrenal gland segmentation remains challenging due to their small size, unclear boundaries with surrounding tissues, and significant variability in shape and position. Our research focuses on adrenal gland segmentation in chest LDCT images, which typically have lower image quality and thus increase the difficulty of the task. Despite this, our study achieved competitive DSC values ​​for 9,606 adrenal glands in a diverse validation set, ranging from 0.799 to 0.870, consistent with previous studies.

[0116] The present study differs from the above studies in three main aspects. First, the present invention focuses on images from chest LDCT scans. Second, a larger data set was used during training, more extensive validation was performed, and the results of the AI ​​model of the present invention were compared with radiology reports, demonstrating the potential clinical benefits of using the AI ​​model. The validation cohort of the present invention included an internal validation cohort (7523 adrenal glands) and an external validation cohort (1545 adrenal glands). The model showed robust classification performance in both cohorts, with a sensitivity of 0.821-0.964, a specificity of 0.927-0.994, an ROC AUC of 0.942-0.977, and a PR AUC of 0.753-0.927. In both internal and external validation cohorts, the AI ​​model was significantly better than radiology reports in detecting adrenal abnormalities. Third, the present study divided the adrenal glands into two categories: normal and abnormal, rather than the normal vs. mass-containing adrenal glands or normal vs. hyperplasia classifications used in previous studies. The definition of abnormality in the present invention includes various morphological and density abnormalities of the adrenal gland, which is more in line with clinical practice.

[0117] In our study, the incidence of adrenal abnormalities was 1.8% in the internal validation cohort (65 of 3681 patients) and 15.4% in the external validation cohort (127 of 779 patients with follow-up). This significant difference in incidence is likely attributable to demographic differences between the cohorts, with the external cohort having a higher number of elderly patients (median age 60.0 years, IQR: 48.0-69.0) than the internal cohort (median age 47.0 years, IQR: 37.0-60.0), a difference that was statistically significant (P < 0.001).

[0118] The AI ​​model proposed in this paper is designed to accurately detect these adrenal abnormalities. Its application in future large-scale studies could enhance understanding of the prevalence of adrenal abnormalities and potentially serve as a preliminary screening tool for community health assessments, making a valuable contribution to the field.

[0119] The disclosed embodiments of the present invention further provide a computer program product or system, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for predicting adrenal lesions based on chest CT.

[0120] Figure 2 The schematic diagram of the system for predicting adrenal lesions based on chest CT provided by an embodiment of the present invention specifically includes:

[0121] Acquisition unit: acquire chest CT data;

[0122] Segmentation unit: inputting the chest CT data into the segmentation network for segmentation to obtain a region of interest;

[0123] Prediction unit: inputs the region of interest into the classifier for prediction to obtain a prediction result of whether the adrenal gland is normal or abnormal.

[0124] Figure 3 A schematic diagram of a device for predicting adrenal lesions based on chest CT provided by an embodiment of the present invention specifically includes:

[0125] A memory and a processor; the memory is used to store program instructions; the processor is used to call program instructions, and when the program instructions are executed, any one of the above-mentioned methods for predicting adrenal lesions based on chest CT is performed.

[0126] The disclosed embodiments of the present invention further provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it performs any of the above-mentioned methods for predicting adrenal lesions based on chest CT.

[0127] The verification results of this verification example show that assigning inherent weights to indications can improve the performance of the present method relative to the default settings. Those skilled in the art will clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways.

[0128] For example, the device embodiments described above are merely illustrative. For example, the division of units described is merely a logical functional division. In actual implementation, other divisions may be employed. For example, multiple units or components may be combined or integrated into another system, or some features may be omitted or not implemented. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through interfaces, or indirect coupling or communication connection between devices or units, which may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in a single location or distributed across multiple network units. Some or all of these units may be selected to achieve the objectives of the present embodiments as needed. Furthermore, the functional units in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. These integrated units may be implemented in either hardware or software functional units. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0129] Those skilled in the art will understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. The above-mentioned medium storage can be a read-only memory, a disk or an optical disk, etc.

[0130] The above is a detailed introduction to a computer device provided by the present invention. For those skilled in the art, according to the concept of the embodiments of the present invention, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting adrenal lesions based on chest CT, characterized in that: include: S1. Obtain low-dose chest CT data to determine whether there are density and morphological abnormalities within the adrenal glands. Abnormal adrenal density includes calcification, hemorrhage, cysts, and fatty infiltration; morphological abnormalities include diffuse adrenal enlargement, multiple nodules, or masses. S2. Inputting the low-dose chest CT data into a segmentation network for segmentation to obtain a target region, performing region expansion based on the target region to generate a region bounding box, and performing segmentation based on the region bounding box to obtain a region of interest; A 3D VNet three-dimensional segmentation network, consisting of an encoder and a decoder, is used to accurately locate the bilateral adrenal glands. The encoder performs N downsampling operations and outputs each downsampling to construct a feature map, where N is a natural number greater than 1. The output is then fused in the decoder via skip connections. The decoder upsamples N times from the variables and concatenates them with the previous encoder feature maps; The 3D segmentation network training process is as follows: obtaining a chest CT dataset and adrenal gland region labels; inputting the dataset and labels into the 3D segmentation network to be trained to obtain region recognition results; comparing the region recognition results with the adrenal gland region labels to obtain a loss function; and optimizing the 3D segmentation network based on the loss function until the loss function remains unchanged, thereby obtaining the 3D segmentation network. S3, inputting the region of interest into a classifier for prediction to obtain a prediction result of whether the adrenal gland is normal or abnormal; The classifier makes predictions using a pre-trained deep learning network, the Med3D deep learning network. The network extracts features from images using the weights of the encoder through transfer learning, retains the encoder components, and replaces the decoder deconvolution portion of the network with convolutional and fully connected layers in the classification network architecture. The classification convolutional layer consists of the following: a maximum pooling layer with a stride of 2; a convolutional layer with a kernel size of 3; a maximum pooling layer with a stride of 2; a second convolutional layer with a kernel size of 3; and a fully connected layer of 128 neurons in the classification network. The adrenal region predicted from the 3D segmentation network is used as a mask to crop the CT image to generate an adrenal bounding box, which then becomes the input to the Med3D model.

2. The method for predicting adrenal lesions based on chest CT according to claim 1, characterized in that: The training process of the pre-trained deep learning network is: Obtaining a set of regions of interest and labels of whether the adrenal glands are normal or abnormal; The set of regions of interest and labels of normal or abnormal adrenal glands are input into a pre-trained deep learning network to be trained to obtain a pre-trained deep learning network. The pre-trained deep learning network is parameter-adjusted and weight-optimized based on the input data.

3. The method for predicting adrenal lesions based on chest CT according to claim 1, characterized in that: The method further includes data preprocessing, wherein the chest CT data is preprocessed to obtain processed data, and the processed data is input into a three-dimensional segmentation network.

4. The method for predicting adrenal lesions based on chest CT according to claim 3, characterized in that: The data preprocessing includes: data enhancement, data resampling, and data standardization.

5. A computer product comprising a computer program or instructions, characterized in that The computer program or instructions are executed by a processor to implement the method for predicting adrenal lesions based on chest CT according to any one of claims 1 to 4.

6. A computer device comprising a memory, a processor, and a computer program or instruction stored in the memory, wherein: The computer program or instructions are executed by a processor to implement the method for predicting adrenal lesions based on chest CT according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: The computer program or instructions are executed by a processor to implement the method for predicting adrenal lesions based on chest CT according to any one of claims 1 to 4.

Images