TW202223914A - Lung nodule detection method on low-dose chest computer tomography images using deep learning and its computer program product - Google Patents

Abstract

A lung nodule detection system on low-dose chest computer tomography images using deep learning includes a lung nodule detection stage and a false positive elimination stage. In the lung nodule detection stage, an original data set is prepared and a cross-validation is performed; the CT images in the original data set are disassembled into multiple 2D slice images, and the CT images are performed with image enhancement, lung area detected and three-channel superimposition to obtain a first data set; a Mask R-CNN image segmentation model receives the first data set for training, and multiple Mask R-CNN models trained from cross-validation are provided to perform hard negative mining to determine hard negative examples; and the hard negative examples are integrated with multiple training sets generated by cross-validation to complete a second data set for training a Mask R-CNN model. In the false positive elimination stage, the positive cases in the original data set and the negative cases selected from the lung nodule detection stage are combined into a data set, and the CT images of the original data set are cut into cuboid 3D images by taking the center of mass coordinates of the false positive results predicted by the Mask R-CNN model trained in the lung nodule detection stage; and a 3D CNN model receives the cuboid 3D images for training.

Description

Deep learning lung nodule detection method and computer program product based on chest low-dose computed tomography images

The present invention relates to a pulmonary nodule detection method, in particular to a deep learning pulmonary nodule detection method based on chest low-dose computed tomography images.

In recent years, lung cancer has become a major health concern for Chinese people, and the detection of pulmonary nodules is an important key to early treatment of lung cancer. Therefore, a pulmonary nodule detection system has been developed to assist radiologists in their diagnostic work and improve their diagnostic quality and efficiency.

Although it is known that lung nodule detection has been developed based on deep learning and neural networks, there are still shortcomings such as complex model architecture, high training cost, and insufficient accuracy in practical applications. After the detection of pulmonary nodules is completed, the conventional pulmonary nodule detection system suffers from the problem of poor false positive rate when performing false positive exclusion.

In view of this, it is necessary to establish a deep learning method for detecting pulmonary nodules based on chest low-dose computed tomography images to effectively assist radiologists in their diagnostic work, thereby improving their diagnostic quality and efficiency.

The present invention proposes a deep learning pulmonary nodule detection method based on chest low-dose computed tomography images to overcome the aforementioned shortcomings of the prior art.

According to a feature of the present invention, a deep learning pulmonary nodule detection method based on chest low-dose computed tomography images is proposed, which includes: a pulmonary nodule detection stage, including the steps of: (A) preparing a CT image including a plurality of CT images and marking A source data set of lung nodules, and a cross-validation is performed on the source data set, wherein the cross-validation uses multiple training sets to train multiple Mask R-CNN models and obtain multiple test results; (B) The CT images in the source data set are disassembled into a plurality of 2D slice images, and image enhancement, lung region detection, and three-channel synthesis are performed on the CT images to obtain a first 2D image including multiple synthesized 2D images. A data set; (C) use a Mask R-CNN image segmentation model, accept the first data set for training, and use the multiple Mask R-CNN models trained by the cross-validation to mine negative hard examples (Hard Negative Mining) to screen out negative hard cases; and (D) integrating the negative hard cases with multiple training sets of the cross-validation to complete a second class with two categories of positive lung nodules and hard negative cases data set, and use the second data set to train a Mask R-CNN model; and a false positive exclusion stage, including the steps: (E) the lung nodules marked in the source data set are regarded as positive examples, and then from the lung nodules False positive cases were screened out as negative cases from the cross-validation test results in the detection phase, and combined into a third dataset with two categories of nodules and non-nodules, and then from the CT images of the source dataset, the lung The Mask R-CNN model trained in the nodule detection stage predicts the centroid coordinate of the false positive result as the center point, and cuts out a cuboid stereo image; and (F) a 3D CNN model accepts the cuboid stereo image for training.

According to another feature of the present invention, a computer program product is provided, stored in a non-transitory computer-readable medium, the computer program product having an instruction to enable a computer system to perform the aforementioned chest-based low-dose computed tomography Image-based deep learning pulmonary nodule detection method.

The above overview and the following detailed description are exemplary in nature, and are intended to further illustrate the scope of the present invention, and other objects and advantages of the present invention will be explained in the following descriptions and drawings.

The following description will provide various embodiments of the present invention. It is understood that these examples are not intended to be limiting. The features of the various embodiments of the invention can be modified, substituted, combined, separated and designed to apply to other embodiments.

The ordinal numbers such as "first", "second" and other terms used in this specification and claims are used to modify the requested elements, and do not imply and represent that there must be a smaller ordinal number before a larger ordinal number, nor do they represent a certain request The arrangement order of elements and another requested element, or the manufacturing order, the use of these ordinal numbers is only used to clearly distinguish a requested element with a certain name from another requested element with the same name.

In addition, the descriptions of "when..." or "when" in this document refer to aspects such as "now, before, or after", and are not limited to simultaneous occurrences, which are described in advance. Similar descriptions of "arranged on" and the like herein represent the corresponding positional relationship between the two elements. Unless otherwise specified, it does not limit whether there is contact between the two elements, which is described in advance here. Furthermore, if the word "or" is used between multiple functions (or elements) herein, it means that the functions (or elements) can exist independently, but it does not exclude the aspect that multiple functions (or elements) can exist simultaneously. .

In addition, the terms "connection", "electrical connection" or "coupling" in the present invention, if not particularly emphasized, include the aspects of direct connection and indirect connection. In addition, the terms "comprising", "including", "having", and "having" in this disclosure are all open-ended descriptions, and are described here first.

Furthermore, the following various embodiments of the present invention can be implemented by means of software programs or electronic circuits, and are not limited thereto.

The present invention is a deep learning pulmonary nodule detection method and computer program product based on chest low-dose computed tomography images. As shown in FIG. The lung nodule detection stage 11 and a false positive exclusion stage 12, wherein the lung nodule detection stage 11 is mainly responsible for finding out the possible position of the lung nodule from the 2D CT slice, and calculating the center of mass of the position. ) coordinates, and after the centroid coordinates are obtained, a stereoscopic image of possible positions of the lung nodule is extracted from the original 3D CT scan image according to the coordinates, and the stereoscopic image is input into the false positive elimination stage 12 . The false-positive exclusion stage 12 is responsible for classifying the stereoscopic images of possible locations of lung nodules into "positive" or "negative", so as to exclude the false-positive predictions generated by the model in the lung nodule detection stage and reduce the overall false-positive rate of the system.

As shown in FIG. 1 , in the lung nodule detection stage 11 of the deep learning lung nodule detection method, data preparation and processing 111 for lung nodule detection are first performed. Data preparation is used to prepare a source data set containing most CT images and marking a plurality of pulmonary nodules. The LIDC-IDRI (Lung Image Database Consortium image collection and Image Database Resource Initiative) dataset jointly organized by the Foundation for the National Institutes of Health (FNIH) and the U.S. Food and Drug Administration (FDA) serves as the original dataset. To obtain the source data set required by the present invention, the LIDC-IDRI original data set contains a total of 1018 CT scan images, each scan contains 0 to 6 pulmonary nodules, and each image is prepared by four experienced radiologists. Physicians perform two-stage image labeling. In the first stage, each radiologist independently labels and diagnoses the labeled area as one of the following three categories: (1) pulmonary nodules ≥ 3 mm; (2) < 3 Lung nodules of mm; (3) non-nodules; in the second stage, each radiologist referred to the diagnosis of the other three radiologists and gave his own final diagnosis.

In order to maintain the consistency of the resolution of the CT images included in the data set, the present invention only selects CT images with a slice thickness greater than or equal to 2.5 mm from the LIDC-IDRI data set, and those existing in these CT images and conforming to the nodule diameter ≧ Pulmonary nodule markers of 3 mm and those approved by 3 to 4 radiologists were used as training and testing data. The source datasets screened under these conditions included a total of 897 CT images and 1179 instances of positive pulmonary nodules.

The pulmonary nodule detection stage 11 is based on the aforementioned source dataset (897 CT images, 1179 positive pulmonary nodule instances) for model training. The training at this stage is used to perform a cross-validation on the source data set, wherein the cross-validation uses multiple training sets to train multiple Mask Region-based Convolutional Neural Networks (Mask Region-based Convolutional Neural Networks). Neural Network, Mask R-CNN) model and obtain multiple test results. In this embodiment, the cross-validation is a 10-fold cross-validation (10-fold cross-validation). As shown in FIG. 2, the 10-fold cross-validation includes: Step S21: use the source data set as a unit of CT image , randomly divided into 10 equal parts such as sub-data set (0), sub-data set (1), ..., sub-data set (9) (each equal part contains 89~90 CT images); and, with index value i =0~9, carry out the following steps: Step S22: select the sub-data set (i) as the test set (i), and simultaneously combine the other 9 equal sub-data sets into the training set (i); and step S23: use the training set ( i) Train a Mask R-CNN model (i), and test the performance of the Mask R-CNN model (i) with a subset of datasets (i) after training; therefore, 10 different Mask R-CNN models are finally obtained. CNN (Mask R-CNN model (0) ~ Mask R-CNN model (9)) and the test results in their respective test sets (test set (0) ~ test set (9)), merge the 10 test results , which can be used as an evaluation index for the overall performance of the first-stage model.

， 其中， a和b是由選擇的窗寬(window width)和窗位(window level)決定，於本發明中，由於對所有CT 影像序列選擇的固定窗位為-400 HU、固定窗寬則為1500 HU，因此可得a=([2×(-400)-1500])⁄2=-1150 HU、b=([2×(-400)+1500])⁄2=350 HU。 3 shows that the data processing in the aforementioned lung nodule detection stage 11 mainly includes three processing steps: image enhancement step S31 , lung area detection step S32 , and three-channel synthesis step S33 . First, since the input of the model in the lung nodule detection stage 11 is a 2D image, the CT images in the source data set need to be disassembled into many 2D slice images to facilitate input into the model. To this end, the following formula 1 can be used to pair the source The Hounsfield Unit (HU) values in the 897 CT images in the dataset were calculated to achieve the effect of image enhancement.

, where a and b are determined by the selected window width and window level. In the present invention, since the fixed window level selected for all CT image sequences is -400 HU, the fixed window width is is 1500 HU, so a=([2×(-400)-1500])⁄2=-1150 HU, b=([2×(-400)+1500])⁄2=350 HU.

其中，X是二值化圖像，B是結構元素、Bx代表B在X上的平移； (6) 閉合(closing)運算步驟326：利用半徑為10的圓盤作為結構元素對二值化影像進行影像形態學中的閉合運算，以保留肺壁上的肺結節，閉合運算本質上是利用同一個結構元素，先對影像作擴張(dilation)運算，再作侵蝕運算(定義如公式2)，而擴張運算的定義則如公式3所示，且由侵蝕運算與擴張運算可組合出閉合運算，如公式4所示：

； (7) 擴張運算步驟327：利用半徑為10的圓盤再次對二值化圖像進行擴張運算，以便更加完整地保留肺壁上的結節； (8) 填補空洞步驟328：將運算過程中於連通區域內產生的空洞填補起來，使其成為完整的區域； (9) 影像疊加步驟329：將二值化影像疊加到原始影像上，將對應到二值化影像中像素值為1的區域的原始影像區域留下，其餘區域刪除，而只留下肺部區域。 In the lung region detection step S32, lung region detection is performed on the sliced image after the aforementioned image enhancement with a series of image morphological functions, so as to remove noise in regions other than the lungs. More specifically, the aforementioned lung area detection step S32 includes nine steps as follows: (1) Image binarization (thresholding) step 321: the threshold used in this step is -400 HU, which can distinguish between the lung area and the extra-pulmonary area. For the effect of the body cavity, the image after binarization will become a binary image; (2) Boundary removal step S322: remove the 1 value near the image boundary in the binary image; ( 3) Connected region labeling (label) step 323: distinguish the binarized image into one or more connected regions; (4) Connected region screening step 324: use the label obtained in the previous step 223 to leave the image in the image. The pixel value of the top three connected areas in the area is 1, and the other parts are changed to 0 to ensure that all lung nodules are left; (5) erosion (erosion) operation step 325: use a disk with a radius of 2 ( disk) as a structuring element to perform the erosion operation in image morphology on the binarized image to separate the pulmonary nodule from the blood vessel. The definition of the erosion operation is shown in Equation 2:

Among them, X is the binarized image, B is the structuring element, and Bx represents the translation of B on X; (6) Closing operation step 326: use a disk with a radius of 10 as a structuring element to perform the binarization image Perform the closing operation in image morphology to preserve the lung nodules on the lung wall. The closing operation essentially uses the same structural element to perform a dilation operation on the image first, and then perform an erosion operation (defined as Equation 2), The definition of the expansion operation is shown in Equation 3, and the closure operation can be combined by the erosion operation and the expansion operation, as shown in Equation 4:

(7) Expansion operation step 327: use a disk with a radius of 10 to perform an expansion operation on the binarized image again, so as to preserve the nodules on the lung wall more completely; Fill up the holes generated in the connected area to make it a complete area; (9) Image superimposition step 329: superimpose the binarized image on the original image, and map the area corresponding to the pixel value of 1 in the binarized image The original image area is left, the rest is deleted, and only the lung area is left.

Fig. 4 shows the output image of each step S321～S329 of the aforementioned lung region detection, as can be seen from Fig. 4, after the slice image is processed by the lung region detection step 32, a clear lung image can be obtained.

In the three-channel synthesis step S33, for the above-mentioned 2D slice images detected in the lung region, three continuous slice images in depth are combined into a synthesized 2D image using the RGB channels of the picture for training, so as to increase the 2D image. 3D spatial information in . And for the same lung nodule (target lung nodule), three consecutive 2D images with different depths are selected as training data, so that the model can extract 3D space features from the 2D images. Among them, the selection method of the three 2D images is as follows: Assuming that the center of mass of the target lung nodule has a depth coordinate of z, the R channel of the first image will be composed of slices with a depth of (z+2), and the G channel is the depth (z+2). z+1), the B channel is the slice of depth z; the second image is the slice of depth (z+1), z, (z-1) in sequence; the third image is the depth in sequence Slices of z, (z-1), (z-2). This combination allowed the model to more accurately identify lung nodules in slices near the centroid depth, which in turn allowed more accurate nodule centroid coordinates to be derived from the model's predictions. However, because some nodules are not large enough to synthesize 3 images, only 1~2 2D images can be synthesized by all the slices containing the nodules in the same way, so the synthesized 2D images actually contain the lung nodules. There are only 3266 images, instead of 1179×3=3537 images, and as shown in Figure 1, the 3266 composite 2D images containing lung nodules are the lung nodule detection stage 11 generated images that only contain lung nodules. A first dataset 112 of a single category.

As shown in FIG. 1 again, the model of the lung nodule detection stage 11 is a Mask R-CNN image segmentation model 113 with ResNet50 as the backbone, and the Mask R-CNN image segmentation model 113 accepts the above-mentioned first data. The 2D image of set 112 is used as input data, and a mask is output to label the objects in the image.

FIG. 5 shows the model training process of the lung nodule detection stage 11. First, in step S51, the first data set 112 that only contains a single type of lung nodules is directly used for training. Referring to FIG. 1 , step S52 is performed to perform Hard Negative Mining 114, which is 10 models (Mask R-CNN model (0) ~ Mask R-CNN model ( 9)) Predict the respective training sets (training set(0) ~ training set(9)). The results returned by the model include the segmentation contour in the image of the pulmonary nodule it detected, and the predicted score (score) representing the probability that the segment is a pulmonary nodule. Next, according to the preset conditions: (1) the predicted score is higher than 0.5, and (2) the intersection rate with any nodule (Intersection over Union, IoU) is less than 0.01, screen out the predictions from the model that meet the preset conditions. Predict the result, and define this prediction result as a Hard Negative Example.

Finally, as shown in Figure 1, the aforementioned negative examples are integrated with the original training set (training set (0) ~ training set (9)) to complete a two-category report with positive pulmonary nodules and negative examples. The second dataset 115 is used to train a new Mask R-CNN model 116 to reduce false positives in the model prediction results. The two training steps of the aforementioned training process are both completed after 180,000 iterations of training at a learning rate of 0.002.

Referring to FIG. 1 , in the false-positive elimination stage 12 of the deep learning lung nodule detection method, data preparation and processing 121 for false-positive elimination are first performed, wherein the input accepted by the model in the false-positive elimination stage 12 is 3D Therefore, the marked nodules in the source data set are regarded as positive cases, and the false positive cases are screened out from the cross-validation test results in stage 11 of lung nodule detection as negative cases, and combined into a nodule with nodules and non-positive cases. The third data set of two categories such as nodules, among which, the screening conditions for negative cases are: (1) prediction probability > 0.2 (prediction confidence threshold); (2) the centroid coordinates of the prediction area and the straight line distance of any nodule are not smaller than the nodule radius. After screening out negative cases (false positives), based on the third dataset with two categories of nodules and non-nodules, false positive results can be predicted from the CT images of the source dataset using the model of lung nodule detection stage 11 The coordinate of the center of mass is the center point, and a cuboid three-dimensional image with a length of 40, a width of 40, and a height of 20 is cut out. The cuboid stereoscopic image is then image enhanced by the function described in Equation 1, and the fixed window width and window level used in the image enhancement are also the same as those used in the lung nodule detection stage 11 .

As shown in Figure 6, the model of the false positive exclusion stage 12 is composed of 1 layer of 3D convolution layer (convolution), 3 layers of 3D mixed depthwise convolution layer (mixed depthwise convolution, MixConv), and 3 layers of fully connected layers (fully connected layers). The three-dimensional convolutional neural network (3D Convolutional Neural Network, 3D CNN) model 122. The present invention uses batch normalization and dropout in the model to avoid the problem of overfitting, and inserts a Rectified Linear Unit (ReLU) between layers as an activation function (activation function) function). The cuboid stereoscopic image input to this 3D CNN model 122 will first pass through 4 layers of convolution layers (1 layer of 3D convolution layers and 3 layers of mixed depth convolution layers) to be converted into feature maps, and then flattened and input to fully connected Floor. In order to make full use of the features of different scales in CT images and reduce the amount of calculation, the present invention uses the MixConv technology, in each layer of the 3-layer 3D mixed depth convolution layer, five different sizes of convolution kernels (kernel) are used to input tensors. Perform operations, as shown in Figure 7 for a detailed description of the MixConv technique, where the input tensor of a MixConv layer is X, that is, X is a tensor of size (c, h, w, d), where c is The number of tensor channels (channel), h is height, w is width, d is depth, MixConv technology will divide X from c dimension into five partial tensors of size (c⁄5,h,w,d) G1, G2,..., G5, after that, these five partial tensors G1, G2,..., G5 will each perform a convolution operation with a convolution kernel Wi (i=1~5), and set the W1 to W5 here. The sizes are (3,3,3), (5,5,3), (7,7,5), (9,9,7), (11,11,9), because considering the depth of CT images The resolution in the direction is usually low, and the convolution kernels W2 to W5 are relatively short on the depth axis. Therefore, when performing convolution operations, use appropriate zero-padding to maintain the height, width and depth of each tensor. Unchanged, after this convolution operation, a part of the output tensor Yi (i=1~5) will be produced for each part of the tensor, and then Y1 to Y5 will be concatenated in the c dimension to produce the output tensor Y .

The training method of the model in stage 12 of false positive exclusion is to complete the training after 90 epochs at the learning rate of 0.001, and in order to achieve the effect of data augmentation, the image will be randomly rotated during the training process (rotation) , and in addition, it can be further proposed by Smith (Smith, L.N., A disciplined approach to neural network hyper-parameters: Part 1--learning rate, batch size, momentum, and weight decay. arXiv preprint arXiv:1803.09820, 2018.) The one-cycle training strategy adjusts the learning rate to achieve better training results.

From the above description, it can be seen that the present invention can improve the false positive rate of the model and reduce the false positive rate from the perspective of the training process by adopting the negative difficult mining (HNM) technology in the training process. Model burden and training cost. And by selecting the mask RCNN image segmentation model in the lung nodule detection stage, the position of the lung nodule in the image can be more accurately pointed out. In addition, by using the 2D plane image training model with less computing resources in the lung nodule detection stage, and using the three-channel superimposition image processing technology to increase the three-dimensional spatial information in the 2D image, it is possible to To strike a balance between training cost and model performance,

Although the present invention has been illustrated by the foregoing embodiments, it will be understood that many modifications and variations are possible in accordance with the spirit of the invention and the scope of the claimed invention.

11: Pulmonary Nodule Detection Phase 12: False positive exclusion stage 111, 121: Data preparation and processing 112: The first data set 113: Mask R-CNN Image Segmentation Model 114: Negative Hard Case Mining 115: Second data set 116: Mask R-CNN Model 122: 3D CNN Models S21~S23: Steps S31~S33: Steps S321~S329: Steps S51~S52: Steps

1 is a schematic diagram of a deep learning pulmonary nodule detection method based on chest low-dose computed tomography images of the present invention; Figure 2 is a flow chart of 10-aliquot cross-validation according to the method of the present invention; 3 is a data processing flow chart of the lung nodule detection phase of the method according to the present invention; Fig. 4 shows the output image of lung region detection according to the method of the present invention; 5 is a model training process in a lung nodule detection stage according to the method of the present invention; Figure 6 schematically shows the 3D CNN model of the false positive exclusion stage of the method according to the invention; and Figure 7 schematically shows the use of the MixConv technique in a 3D CNN model according to the method of the present invention.

11: Pulmonary Nodule Detection Phase

12: False positive exclusion stage

111, 121: Data preparation and processing

112: The first data set

113: Mask R-CNN Image Segmentation Model

114: Negative Hard Case Mining

115: Second data set

116: Mask R-CNN Model

122: 3D CNN Models

Claims (14)

A deep learning pulmonary nodule detection method based on chest low-dose computed tomography images, comprising: A lung nodule detection stage, including steps: (A) Prepare a source dataset containing most CT images and multiple pulmonary nodules, and perform a cross-validation on the source dataset, wherein the cross-validation is based on multiple training sets to train multiple Mask R- CNN model and obtain multiple test results; (B) Disassemble the CT images in the source data set into most 2D slice images, and perform image enhancement, lung region detection, and three-channel synthesis on the CT images to obtain a 2D image containing multiple composites The first dataset of images; (C) Use a Mask R-CNN image segmentation model to accept the first data set for training, and use the multiple Mask R-CNN models trained by the cross-validation to perform negative hard example mining to filter out negative case; and (D) Integrate the negative and difficult examples with multiple training sets of the cross-validation to complete a second dataset with two categories of positive pulmonary nodules and hard negative examples, and use the second dataset to train a Mask R - CNN models; and A false positive exclusion stage, including steps: (E) Take the marked lung nodules in the source data set as positive cases, and then screen out false positive cases as negative cases from the cross-validation test results in the lung nodule detection stage, and combine them into a nodule with nodules and non-nodules Wait for the third data set of the second category, and then use the CT image of the source data set to use the Mask R-CNN model trained in the lung nodule detection stage to predict the centroid coordinates of the false positive results as the center point, and cut out the cuboid stereoscopic images; and (F) A 3D CNN model is trained on the cuboid stereo image. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images as claimed in claim 1, wherein, in step (A), the cross-validation is a 10-section cross-validation, which includes the steps: (A1) The source dataset is randomly divided into 10 equal sub-data sets, including sub-data sets (0) ~ sub-data sets (9), with CT images as the unit; and (A2) with the index value i=0~9, carry out the following steps (A3)~(A5); (A3) Select the sub-data set (i) as the test set (i), and at the same time combine the other 9 equal sub-data sets into the training set (i); (A4) train a Mask R-CNN model (i) with the training set (i); and (A5) Test the performance of the Mask R-CNN model (i) with the sub-dataset (i) to obtain the test results of the Mask R-CNN model (i) in its test set (i). The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images according to claim 1, wherein, in step (B), the HU value in the CT image of the source data set is based on the following image enhancement formula Perform operations to achieve the effect of image enhancement:

, where a and b are determined by the window width and window level selected for the CT image sequence. The deep learning lung nodule detection method based on chest low-dose computed tomography images as claimed in claim 1, wherein, in step (B), the lung region detection is performed on the image-enhanced slice images by a series of The Image Morphology function performs lung region detection to remove noise outside the lung. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images according to claim 4, wherein the lung area detection comprises: image binarization step; boundary removal step; connected area marking step; connected area A screening step; an erosion operation step; a closure operation step; an expansion operation step; a void filling step; and an image overlay step. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images as claimed in claim 1, wherein, in step (B), the three-channel synthesis is to use the RGB of the image for three consecutive slice images in depth. The channels are combined into a synthetic 2D image for training to increase the three-dimensional spatial information, and for the same lung nodule, three consecutive 2D images with different depths are selected as training data, so that the features of the 3D space can be extracted from the 2D images. . The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images according to claim 1, wherein the Mask R-CNN image segmentation model is a Mask R-CNN image segmentation model with ResNet50 as the backbone. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images as claimed in claim 1, wherein the mining of negative difficult cases is based on 10 Mask R-CNN models trained by the 10 equal parts cross-validation. The respective training sets are predicted, and then a prediction result that meets a preset condition is selected from the prediction, so as to define the prediction result as a negative difficult case. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images as claimed in claim 3, wherein, in step (E), image enhancement is further performed on the cut-out cuboid three-dimensional image using the image enhancement formula. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images as claimed in claim 1, wherein, in step (F), the 3D CNN model is composed of 1 layer of 3D convolutional layers, 3 layers of 3D mixed depth volume It consists of built-up layers and 3 fully connected layers. The deep learning lung nodule detection method based on chest low-dose computed tomography image as claimed in claim 10, wherein, in the 3D CNN model, batch normalization and discarding method are used to avoid overfitting, and the layer and A linear rectification function is inserted between the layers as the excitation function. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images according to claim 10, wherein the cuboid stereoscopic image input into the 3D CNN model first passes through four convolution layers to convert into feature maps and then is compressed Flatten and enter the fully connected layer. The deep learning pulmonary nodule detection method based on chest low-dose computed tomography images as claimed in claim 10, wherein, in the 3D CNN model, each layer of the 3-layer 3D mixed depth convolutional layer is composed of five kinds of Convolution kernels of different sizes operate on the input tensors to make full use of features at different scales in CT images and reduce computation. A computer program product, stored in a non-transitory computer-readable medium, the computer program product having an instruction to cause a computer system to perform the chest-based low-dose computed tomography of any one of claims 1 to 13 Image-based deep learning pulmonary nodule detection method.

Images