Colorectal cancer detection method and device based on conventional enhanced CT (Computed Tomography)

Abstract

The invention discloses a colorectal cancer detection method and device based on conventional enhanced CT, and the method comprises the steps: collecting a data set of an abdominal CT image, and carrying out the preprocessing of the data set; respectively labeling the colorectum and the colorectal cancer of the abdomen CT image as real labels of the detection model; a detection model is constructed, the detection model uses nnUNet as a backbone network of pixel-level colorectal and colorectal cancer segmentation, an auxiliary classification branch is combined to integrate multi-scale features from the nnUNet, and patch-level classification is realized; training the detection model, wherein the sum of segmentation loss and classification loss is taken as total loss during training; and performing colorectal cancer detection by using the trained detection model, and determining that the detection is successful when the detected tumor volume exceeds a set cutoff value. According to the method, the robustness and accuracy of colorectal cancer positioning and segmentation are improved, and the labor cost is greatly saved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a method and device for detecting colorectal cancer based on conventional enhanced CT. Background Art

[0002] Enhanced computed tomography (CT), a widely used clinical examination method, plays an important role in the diagnosis of various diseases and also provides a method for detecting colorectal cancer (CRC). Accurately identifying CRC through enhanced CT can help improve patient treatment outcomes and reduce deaths caused by the disease. However, traditional enhanced CT scans are often performed without proper bowel preparation, which may cause CRC to be obscured by intestinal material or collapsed areas. Coupled with the increasing workload in the medical environment, radiologists may focus more on confirming the primary indications for CT scans and fail to fully examine the colorectal region. In addition, the diagnosis of CRC is also affected by differences in experience among different radiologists.

[0003] Despite this, deep learning technology has demonstrated tremendous potential in the field of medical image analysis. However, most current research and development efforts using deep learning focus on segmenting and localizing colorectal cancer using CT, magnetic resonance imaging (MRI), and colonoscopy. Therefore, utilizing deep learning to detect colorectal cancer using conventional contrast-enhanced CT without bowel preparation is a research area in this field. Summary of the Invention

[0004] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the existing technology and provide a method and device for colorectal cancer detection based on conventional enhanced CT. By training the segmentation backbone and auxiliary classification branches to segment and classify CT image data, the robustness and accuracy of colorectal cancer positioning and segmentation are improved, and labor costs are greatly saved.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a method for detecting colorectal cancer based on conventional enhanced CT, comprising the following steps:

[0007] Collecting a data set of abdominal CT images and preprocessing the data set; the data set includes colorectal cancer patient data and healthy colorectal data;

[0008] The colorectal and colorectal cancer images in the abdominal CT images are labeled separately as the true labels of the detection model. The labels include pixel-level labels and image-level labels.

[0009] A detection model is constructed, which uses nnUNet as the backbone network for pixel-level colorectal and colorectal cancer segmentation, and is combined with an auxiliary classification branch to integrate multi-scale features from nnUNet to achieve patch-level classification. The nnUNet includes an encoding part and a decoding part, each of which includes multiple resolution stages, each resolution stage containing at least one block, and each block consists of a convolution, an instance normalization instanceNorm, and a leakyReLU nonlinear unit. A skip connection operation is set between the encoding part and the decoding part, and the skip connection operation transfers the features extracted by the encoding part to the decoding part to restore the feature map for accurate segmentation. In the auxiliary classification branch, features extracted from different resolution stages of nnUNet are used to distinguish between patches with colorectal cancer and patches without colorectal cancer.

[0010] The detection model is trained. During the training, on the one hand, a segmentation map of colorectal and colorectal cancer is predicted based on the pixel-level labels, and on the other hand, a classification prediction of whether the image contains colorectal cancer is performed based on the image-level labels; during the training, the sum of the segmentation loss and the classification loss is used as the total loss;

[0011] The trained detection model is used to detect colorectal cancer. The detection is successful when the detected tumor volume exceeds the set cutoff value.

[0012] As a preferred technical solution, the preprocessing includes excluding abnormal data, data enhancement processing and normalization processing;

[0013] The abnormal data includes:

[0014] Repeated CT image data of the same patient;

[0015] CT image data that does not cover the entire abdomen and pelvis;

[0016] CT images containing severe artifacts;

[0017] CT images of a patient presenting with severe ascites;

[0018] CT images of a patient with residual positive contrast agent in the colorectum;

[0019] The data enhancement processing includes rotation, scaling, Gaussian noise, Gaussian blur, brightness, contrast, low-resolution simulation, gamma and mirroring.

[0020] As a preferred technical solution, the colorectal and colorectal cancer images in the abdominal CT images are marked separately, specifically:

[0021] Use the automatic segmentation model to perform preliminary segmentation on the data set to obtain a rough segmentation map;

[0022] Correct and unify the rough segmentation map, and annotate the image-level labels corresponding to the colorectum and colorectal cancer.

[0023] Pixel-level labels for background, colorectal, and colorectal cancer.

[0024] As an optimal technical solution, the auxiliary classification branch includes a global average pooling layer, a connection node, a fully connected layer and a Softmax layer; multiple global average pooling layers are connected in parallel through the connection node, and then connected to the series fully connected layer, and the fully connected layer is connected to the Softmax layer.

[0025] As a preferred technical solution, in the detection model, a global average pooling operation is used to convert the feature maps obtained at different resolution stages into the same size, and then the feature maps are connected in series and fed into two fully connected layers and a softmax operation for classification.

[0026] As a preferred technical solution, the segmentation loss is the sum of the cross entropy loss and the Dice loss;

[0027] The segmentation cross entropy loss L CE Expressed as:

[0028]

[0029] The segmentation Dice loss L DC Expressed as:

[0030]

[0031] in, is the number of categories, HWD is the size of the original CT image; I(·) is the indicator function, when the true label of voxel t is category n, the function value is 1, otherwise it is 0; Y t is the true label of voxel t, Pred t,n is the segmentation probability that the t-th voxel belongs to category n;

[0032] The classification loss represents the image-level cross entropy loss, and the image-level cross entropy loss L cls Expressed as:

[0033]

[0034] Among them, y c , p c They correspond to the true category label and prediction result of category c respectively, and C is the total number of categories.

[0035] As a preferred technical solution, if the predicted tumor volume exceeds the cutoff value and the Dice index between the predicted tumor and the labeled tumor is greater than 0, the tumor detection is considered correct; when the predicted tumor volume is lower than the cutoff value, or the Dice index is equal to 0, the tumor detection is considered incorrect.

[0036] In a second aspect, the present invention provides a colorectal cancer detection system based on enhanced CT, which is applied to the colorectal cancer detection method based on conventional enhanced CT, and includes a preprocessing module, a labeling module, a model building module, a model training module, and a detection module;

[0037] The preprocessing module is used to collect a data set of abdominal CT images and preprocess the data set; the data set includes colorectal cancer patient data and healthy colorectal data;

[0038] The labeling module is used to label the colorectal and colorectal cancer in the abdominal CT image respectively as the true labels of the detection model, and the labels include pixel-level labels and image-level labels;

[0039] The model building module is used to build a detection model, which uses nnUNet as the backbone network for pixel-level colorectal and colorectal cancer segmentation, and combines an auxiliary classification branch to integrate multi-scale features from nnUNet to achieve patch-level classification; the nnUNet includes an encoding part and a decoding part, and the encoding part and the decoding part include multiple resolution stages, each resolution stage contains at least one block, and each block consists of a convolution, an instance normalization instanceNorm, and a leakyReLU nonlinear unit; a jump connection operation is set between the encoding part and the decoding part, and the jump connection operation transfers the features extracted by the encoding part to the decoding part to restore the feature map for accurate segmentation; in the auxiliary classification branch, the features extracted by different resolution stages of nnUNet are used to distinguish between blocks with colorectal cancer and blocks without colorectal cancer;

[0040] The model training module is used to train the detection model. During training, on the one hand, the segmentation map of colorectal and colorectal cancer is predicted based on the pixel-level labels, and on the other hand, the classification prediction of whether there is colorectal cancer is performed based on the image-level labels; during training, the sum of the segmentation loss and the classification loss is used as the total loss;

[0041] The detection module is used to detect colorectal cancer using the trained detection model, and the detection is successful when the detected tumor volume exceeds the set cutoff value.

[0042] In a third aspect, the present invention provides an electronic device, comprising:

[0043] at least one processor; and,

[0044] a memory communicatively connected to the at least one processor; wherein,

[0045] The memory stores computer program instructions that can be executed by the at least one processor, and the computer program instructions are executed by the at least one processor to enable the at least one processor to perform the colorectal cancer detection method based on conventional enhanced CT.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium storing a program, which, when executed by a processor, implements the method for detecting colorectal cancer based on conventional enhanced CT.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] (1) The present invention segments and classifies CT image data by training the segmentation backbone and auxiliary classification branches, and uses the trained detection model to detect colorectal cancer. When the detected tumor parameters exceed the set cutoff value, the detection is completed, thereby improving the robustness and accuracy of colorectal cancer detection and greatly saving labor costs.

[0049] (2) The present invention performs block connection between the output features of the segmentation backbone and the auxiliary classification branch, introduces a fully connected layer and a Softmax operation, improves the stability and accuracy of the segmentation image of the segmentation backbone, and enhances the classification ability of the detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application, 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 application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0051] Figure 1 This is a flow chart of a method for detecting colorectal cancer based on conventional enhanced CT according to an embodiment of the present invention;

[0052] Figure 2 Schematic diagram of a segmentation model according to an embodiment of the present invention;

[0053] Figure 3 Schematic diagram of the structure of a colorectal cancer detection system based on enhanced CT according to an embodiment of the present invention;

[0054] Figure 4 2 is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described 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, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0056] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0057] See also Figure 1 The present embodiment provides a method for detecting colorectal cancer based on conventional enhanced CT, comprising the following steps:

[0058] S1. Collect a data set of colorectal CT images and preprocess the data set; the data set includes colorectal cancer patient data and healthy colorectal data.

[0059] CT image data of CRC patients from multiple centers and CT image data of healthy colorectal volunteers were collected as training sets and test sets for the model. The present invention collected data from four centers as training sets and randomly divided them into training sets, validation sets and internal test sets. In addition, data from two centers were collected as two independent external test sets. All CRC patients underwent abdominal enhanced CT scans within one month before surgery and were confirmed by postoperative pathology reports. Abdominal enhanced CT scans of healthy individuals were selected based on the statement that no obvious colorectal abnormalities were found in the original radiology report and further verified by 12 months of follow-up.

[0060] In particular, CT image data preprocessing methods include: excluding abnormal data, data enhancement, and normalization. Abnormal data include: repeated CT image data of the same patient, CT image data that does not cover the entire abdomen and pelvis, CT image data containing severe artifacts, CT image data with severe ascites, and CT image data with residual reagents from other colorectal tests, such as residual positive contrast agent in the colorectum.

[0061] Furthermore, the data enhancement processing includes rotation, scaling, Gaussian noise, Gaussian blur, brightness, contrast, low-resolution simulation, gamma and mirroring.

[0062] Furthermore, in this embodiment, all CT images are resampled to a fixed voxel spacing of 0.76×0.76×3 mm. 3 During model training, all images were randomly cropped to 192 × 192 × 96, and the training batch size was set to 16. The model was trained for 1000 epochs, with each epoch containing 250 mini-batches. Stochastic gradient descent was used with an initial learning rate of 0.01 and Nesterov momentum (0.99) to learn the network weights. The detection model was trained on four GPUs (NVIDIA GeForce RTX 3090 with 24GB RAM). Model development and training were implemented in Python 3.7.11 using PyTorch 1.8.0. For details on the construction and training of the detection model, see steps S3-S4.

[0063] S2. Label the colorectum and colorectal cancer in the colorectal CT image separately, and the labeling includes pixel-level labels and image-level labels.

[0064] In this example, an automated segmentation model for colorectal segmentation was employed to facilitate efficient annotation. The segmentation results were then simply corrected to provide a rough colorectal annotation. The colorectal cancer annotations were based on patient clinical and pathological information, including tumor location, size, and invasion depth. Both the colorectal correction and tumor annotation processes were performed in ITK-SNAP software.

[0065] Specifically, this example annotates colorectal and colorectal cancer images on 3D venous phase CT images of all cases to obtain true labels for detection model training. Pixel-level labels for model training are set to 0, 1, and 2, representing background, colorectal, and colorectal cancer, respectively. Image sets are labeled 0 and 1, representing the absence and presence of colorectal cancer, respectively.

[0066] S3. Construct a detection model, which includes a segmentation backbone and an auxiliary classification branch; the segmentation backbone includes an encoding part and a decoding part, and a jump connection is set between the encoding part and the decoding part; the auxiliary classification branch includes multi-scale features for classifying the presence or absence of colorectal cancer at the image level to assist the segmentation model in feature learning; after calculating the loss and back propagation, a trained detection model is obtained.

[0067] See also Figure 2 ,In this embodiment, nnUNet is used as the pixel-level image segmentation backbone, ,and an auxiliary classification branch is combined to integrate the multi-scale features ,from nnUNet to achieve patch-level classification.,The segmentation backbone and the auxiliary classification ,branch are connected through blocks.

[0068] The nnUNet consists of an encoder and a decoder, each consisting of five resolution stages, each with two blocks, including convolution, instance normalization (InstanceNorm), and leakyReLU nonlinear units. Skip connections transfer the extracted features from the encoder to the decoder to recover the feature map for accurate segmentation.

[0069] In the classification branch, features from the bottom three stages of nnUNet are used to distinguish blocks with CRC from blocks without CRC. First, the feature maps from different stages are converted to the same size using a global average pooling (GAP) operation. Second, the feature maps are concatenated and fed into two fully connected (FC) layers and a softmax operation for classification.

[0070] S4. The training of the detection model is mainly divided into two parts. One part is based on the pixel-level label Y seg Perform colorectal and colorectal cancer segmentation map P seg The other part is based on the image-level label Y cls Classify P cls The specific process is as follows:

[0071] S41. In the detection network f, a predicted segmentation map P is obtained for each input image X. seg And the classification result P cls :

[0072] P seg ,P cls =f(X,Y seg ,Y cls )

[0073] S42, the loss of the two parts (segmentation loss L seg , classification loss L cls ) are combined as a joint loss to optimize the performance of the entire model, where the semantic segmentation loss L seg is the cross entropy loss L CE and Dice loss L DC Comprehensive:

[0074]

[0075] in, is the number of pixel-level categories, HWD is the dimension of the input image, I(·) is an indicator function, when Y t =n when I(Y t =n) = 1, otherwise I(Y t =n)=0.

[0076] The loss for classification loss is the image-level cross entropy loss, which can be expressed as:

[0077]

[0078] Where M is the number of image-level categories.

[0079] Therefore, the total loss function L for detection model training is:

[0080] L=L seg +αL cls

[0081] Back propagate and repeat steps (1)-(2) until the set stopping condition is reached.

[0082] Finally, we collected colorectal cancer CT data and healthy colorectal CT data from multiple centers and annotated the tumor area for verification. Each CT image was input into the segmentation network to calculate the prediction result:

[0083] P pred =f(X)

[0084] In the validation phase, the trained network model only needs to take the CT image X as input and simply retain the segmentation prediction P pred as the final output.

[0085] S5. Use the trained detection model to detect colorectal cancer. When the detected tumor parameters exceed the set cutoff value, the detection is successful.

[0086] For detected tumors, the relevant detection rules should be followed as follows:

[0087] A cutoff value is set based on the predicted tumor volume: if the predicted tumor volume exceeds the cutoff value and the Dice index between the predicted tumor and the labeled tumor is greater than 0, the tumor detection is considered correct; when the predicted tumor volume is lower than the cutoff value, or the Dice index is equal to 0, the tumor detection is considered incorrect.

[0088] Furthermore, the following model validation steps are included:

[0089] S6. Steps to validate model performance with multi-center data:

[0090] S61. Collect colorectal cancer CT data and healthy colorectal CT data from multiple centers and mark the tumor area for verification.

[0091] S62. Input each CT image and calculate the prediction result in the segmentation network:

[0092] P pre d=f(X)

[0093] S63. In the verification phase, the trained network model only needs to take the CT image X as input and simply retain the segmentation prediction P pred as the final output.

[0094] S64. Model detection rules: A cutoff value is set based on the predicted tumor volume: if the predicted tumor volume exceeds the cutoff value and the Dice index between the predicted tumor and the labeled tumor is greater than 0, the tumor is considered to be detected correctly; when the predicted tumor volume is lower than the cutoff value, or the Dice index is equal to 0, the tumor is considered to be detected incorrectly.

[0095] S65. Model detection performance evaluation indicators: receiver operating characteristic (ROC) curve, area under the receiver operating characteristic curve (AUC), detection sensitivity, specificity, accuracy, positive predictive value and negative predictive value.

[0096] Furthermore, the following auxiliary diagnosis steps are also included:

[0097] S7, human-machine detection performance comparison and detection model-assisted diagnosis;

[0098] S71. Collect enhanced CT image data of patients with colorectal cancer and healthy colorectum, and recruit multiple radiologists with different diagnostic experience to diagnose each CT image. The diagnostic goal is to determine whether the CT image contains colorectal cancer. If so, roughly mark the layer with the largest area of the suspected tumor.

[0099] S72. Participating radiology suspects first perform a diagnosis on the CT images without the aid of a detection model. All suspects are not involved in the aforementioned tumor annotation process and are unaware of the distribution and proportion of cancer patients and healthy individuals. During the diagnostic process, radiologists are provided only with the CT images and are unaware of any clinical, radiological, or pathological diagnostic information.

[0100] S73. After an interval of at least 30 days, all radiologists in step S72 re-diagnose the CT images with the assistance of the detection model. The radiologists annotate the tumor regions based on the reference CT images and the tumors detected by the detection model. The performance of the detection model and the independently interpreted detection performance are not disclosed to the radiologists.

[0101] S74. Doctor detection rule: For the same CT image, the consistent diagnosis results of more than half of the radiologists are regarded as the comprehensive diagnosis results of all doctors for the case.

[0102] S75. McNemar's test was used to compare the detection performance between the detection model and radiologists. The significance level was set at P < 0.5.

[0103] In a specific embodiment, data collected by Guangdong Provincial People's Hospital are collected to further illustrate the solution of the present invention.

[0104] (1) Data collection;

[0105] The present invention included 984 colorectal cancer patients between January 2008 and December 2018 and 1073 individuals without colorectal abnormalities between January 2018 and June 2018 at Guangdong Provincial People's Hospital. These data were randomly divided into a training set (788 colorectal cancer patients and 919 normal individuals), a validation set (98 colorectal cancer patients and 115 normal individuals), and an internal test set (98 colorectal cancer patients and 115 normal individuals). The present invention also recruited 310 patients with T1 colorectal cancer from four hospitals to enrich the training set, including 38 from the Shanghai Institute of Oncology, Fudan University, from May 2016 to May 2021; 43 from the Affiliated Hospital of Qingdao University, from October 2016 to November 2022; 42 from Guangdong Provincial People's Hospital, from March 2011 to March 2021; and 187 from the Sixth Affiliated Hospital of Sun Yat-sen University, from December 2012 to September 2022. To verify the performance of the detection model, two independent data sets were also included for external testing. External test set 1 included 202 colorectal cancer patients and 265 normal individuals from the First Affiliated Hospital of Chongqing Medical University between January 2018 and December 2018. The external test set 2 included 252 colorectal cancer patients and 481 normal individuals from Shaanxi Peking University Hospital between January 2016 and September 2022. To further explore the detection performance of the detection model in the real world, the present invention collected a group of consecutive patients who underwent enhanced CT scans at Guangdong Provincial People's Hospital between February 22 and March 28, 2023. After the inclusion and exclusion process, the real-world cohort included 53 colorectal cancer patients confirmed by postoperative or biopsy pathology reports, and 1524 normal individuals without radiologically suspected colorectal abnormalities. The baseline information of all data is shown in Table 1.

[0106] Table 1 Baseline information of all datasets

[0107]

[0108] (2) Data annotation;

[0109] The present invention annotates colorectal and colorectal cancer images on 3D venous phase CT images of all cases to obtain true labels for detection model training. Pixel-level labels for model training are set to 0, 1, and 2, representing background, colorectal, and colorectal cancer, respectively. Image set labels are set to 0 and 1, representing the absence and presence of colorectal cancer, respectively.

[0110] (3) Detection model training;

[0111] During model training, all CT images were resampled to a fixed voxel spacing of 0.76 × 0.76 × 3 mm³. The intensities of all resampled images were normalized by subtracting the global mean and dividing by the global standard deviation. Data augmentation techniques including rotation, scaling, Gaussian noise, Gaussian blur, brightness, contrast, low-resolution simulation, gamma, and mirroring were used during model training. All images were randomly cropped to 192 × 192 × 96, and the training batch size was set to 16. The model was trained for 1000 epochs, with each epoch containing 250 mini-batches. Stochastic gradient descent was used with an initial learning rate of 0.01 and Nesterov momentum (0.99) to learn the network weights. The detection model was trained on four GPUs (NVIDIA GeForce RTX 3090 with 24GB RAM). Model development and training were implemented in Python 3.7.11 using PyTorch 1.8.0.

[0112] (4) Detection model center test;

[0113] The model obtained by the present invention has an AUC of 0.988 in the validation set, 0.994 on the internal test set, 0.965 on the Chongqing test set 1, 0.957 on the Shanxi test set, and 0.968 on the real-world test set. In the validation set, the detection model had a sensitivity of 95.9%, a specificity of 97.4%, and an accuracy of 96.7% (95% CI 93.4-98.7) in detecting CRC in enhanced CT images without colon preparation. The detection model achieved a sensitivity of 96.9%, a specificity of 97.4%, and an accuracy of 97.2% in the internal test set. In the Chongqing test set, the detection model had a sensitivity of 93.6%, a specificity of 95.9%, and an accuracy of 94.9%. In the Shanxi test set, the detection model had a sensitivity of 91.7%, a specificity of 95.6%, and an accuracy of 94.3%. In the real-world test set, the detection model achieved a sensitivity of 94.3%, a specificity of 97.4%, and an accuracy of 97.3% to distinguish patients with CRC from normal individuals (see Table 2 for details).

[0114] Table 2 Detection performance evaluation of the detection model on the validation set and test set

[0115]

[0116] (5) Comparison of the detection performance of the detection model and that of radiologists;

[0117] Nine radiologists of varying experience performed CT image interpretation on the internal test set and the Chongqing test set. The detection performance is shown in Table 3. Compared with the overall radiologists, the detection model had significantly higher sensitivity in the internal test set (96.9% vs. 82.7%, P < .0001) and the Chongqing test set (93.6% vs. 81.7%, P < .0001). In the internal test set, the specificity of the detection model was not significantly different from that of the radiologists (97.4% vs. 99.1%, P = .625). In the Chongqing test set, the specificity of the detection model was lower than that of the overall radiologists (95.9% vs. 98.9%, P = .039). The accuracy of the detection model significantly exceeded that of the radiologists in the internal test set (97.2% vs. 91.6%, P < .0001) and the Chongqing test set (94.9% vs. 91.4%, P < .0001).

[0118] Table 3 Comparison of detection performance between the detection model and radiologists on the internal test set and the Chongqing test set

[0119]

[0120] (6) Detection models to assist radiologists in their diagnosis;

[0121] In both the internal and Chongqing test sets, the detection model improved the radiologists' sensitivity (95.9% vs. 82.7%, P = .0002; 94.1% vs. 81.7%, P < .0001) without reducing their specificity (99.1% vs. 99.1%, P > .99; 98.1% vs. 98.9%, P = .688). Thus, the accuracy of the radiologists' detection assistance was significantly higher in the internal test set (97.7% vs. 91.6%, P = .001) and also higher in the Chongqing test set (96.4% vs. 91.4%, P < .0001). The detection performance is shown in Table 4.

[0122] Table 4 Comparison of detection performance before and after the detection model assisted radiologists

[0123]

[0124] In summary, automated CRC detection was achieved using conventional enhanced CT images, which have greater potential for application, validating the potential of artificial intelligence models for computer-assisted colorectal cancer detection. The invented detection model achieved an AUC of 0.957-0.994 and an accuracy of 0.943-0.973 in multicenter data, with significantly higher sensitivity than that achieved by multiple radiologists with varying levels of experience. Furthermore, the detection model significantly improved the diagnostic accuracy of radiologists.

[0125] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0126] Based on the same concept as the conventional enhanced CT-based colorectal cancer detection method in the above-mentioned embodiment, the present invention also provides a colorectal cancer detection system based on enhanced CT, which can be used to perform the above-mentioned conventional enhanced CT-based colorectal cancer detection method. For ease of explanation, the structural diagram of the embodiment of the colorectal cancer detection system based on enhanced CT only shows the parts related to the embodiment of the present invention. Those skilled in the art will understand that the illustrated structure does not constitute a limitation of the device, and it can include more or fewer components than shown, or combine certain components, or arrange the components differently.

[0127] See also Figure 3 In another embodiment of the present application, a colorectal cancer detection system 10 based on enhanced CT is provided, which includes a preprocessing module 11, a labeling module 12, a model building module 13, a model training module 14 and a detection module 15;

[0128] The preprocessing module 11 is used to collect a data set of abdominal CT images and preprocess the data set; the data set includes colorectal cancer patient data and healthy colorectal data;

[0129] The labeling module 12 is used to label the colorectal and colorectal cancer in the abdominal CT image respectively as the true label of the detection model, and the label includes pixel-level label and image-level label;

[0130] The model construction module 13 is used to construct a detection model, which uses nnUNet as the backbone network for pixel-level colorectal and colorectal cancer segmentation, and is combined with an auxiliary classification branch to integrate multi-scale features from nnUNet to achieve patch-level classification; the nnUNet includes an encoding part and a decoding part, and the encoding part and the decoding part include multiple resolution stages, each resolution stage includes at least one block, and each block is composed of a convolution, an instance normalization instanceNorm, and a leakyReLU nonlinear unit; a jump connection operation is set between the encoding part and the decoding part, and the jump connection operation transfers the features extracted by the encoding part to the decoding part to restore the feature map for accurate segmentation; in the auxiliary classification branch, the features extracted by different resolution stages of nnUNet are used to distinguish between blocks with colorectal cancer and blocks without colorectal cancer;

[0131] The model training module 14 is used to train the detection model. During training, the colorectal and colorectal cancer segmentation maps are predicted based on pixel-level labels, and the classification prediction of whether the image contains colorectal cancer is performed based on image-level labels. During training, the sum of the segmentation loss and the classification loss is used as the total loss.

[0132] The detection module 15 is used to detect colorectal cancer using the trained detection model, and the detection is successful when the detected tumor volume exceeds the set cutoff value.

[0133] It should be noted that the enhanced CT-based colorectal cancer detection system of the present invention corresponds one-to-one to the conventional enhanced CT-based colorectal cancer detection method of the present invention. The technical features and beneficial effects described in the above-mentioned embodiment of the conventional enhanced CT-based colorectal cancer detection method are applicable to the embodiment of the conventional enhanced CT-based colorectal cancer detection method. For specific contents, please refer to the description in the embodiment of the method of the present invention. No further details will be given here. This is hereby declared.

[0134] In addition, in the implementation of the enhanced CT-based colorectal cancer detection system in the above-mentioned embodiment, the logical division of each program module is only an example. In actual application, the above-mentioned functions can be assigned to different program modules as needed, for example, for the configuration requirements of the corresponding hardware or the convenience of software implementation. That is, the internal structure of the enhanced CT-based colorectal cancer detection system is divided into different program modules to complete all or part of the functions described above.

[0135] See also Figure 4In one embodiment, an electronic device for implementing a colorectal cancer detection method based on conventional enhanced CT is provided. The electronic device 20 may include a first processor 21, a first memory 22 and a bus, and may also include a computer program stored in the first memory 22 and executable on the first processor 21, such as a colorectal cancer detection program 23 based on enhanced CT.

[0136] The first memory 22 includes at least one type of readable storage medium, including a flash memory, a mobile hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the first memory 22 may be an internal storage unit of the electronic device 20, such as a mobile hard disk of the electronic device 20. In other embodiments, the first memory 22 may also be an external storage device of the electronic device 20, such as a plug-in mobile hard disk, a smart memory card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 20. Furthermore, the first memory 22 may include both an internal storage unit of the electronic device 20 and an external storage device. The first memory 22 can be used not only to store application software and various types of data installed in the electronic device 20, such as the code of the enhanced CT-based colorectal cancer detection program 23, but also to temporarily store data that has been output or is about to be output.

[0137] In some embodiments, the first processor 21 may be composed of an integrated circuit, such as a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and a combination of various control chips. The first processor 21 is the control core (Control Unit) of the electronic device, connecting the various components of the entire electronic device using various interfaces and lines, and executing or executing programs or modules stored in the first memory 22, as well as calling data stored in the first memory 22, to perform various functions of the electronic device 20 and process data.

[0138] Figure 4 Only the electronic device with components is shown, and it can be understood by those skilled in the art that Figure 4 The structure shown does not constitute a limitation on the electronic device 20 , and the electronic device 20 may include fewer or more components than shown in the figure, or combine certain components, or arrange the components differently.

[0139] The colorectal cancer detection program 23 based on enhanced CT stored in the first memory 22 of the electronic device 20 is a combination of multiple instructions. When running in the first processor 21, it can achieve the following:

[0140] Collecting a data set of abdominal CT images and preprocessing the data set; the data set includes colorectal cancer patient data and healthy colorectal data;

[0141] The colorectal and colorectal cancer images in the abdominal CT images are labeled separately as the true labels of the detection model. The labels include pixel-level labels and image-level labels.

[0142] A detection model is constructed, which uses nnUNet as the backbone network for pixel-level colorectal and colorectal cancer segmentation, and is combined with an auxiliary classification branch to integrate multi-scale features from nnUNet to achieve patch-level classification. The nnUNet includes an encoding part and a decoding part, each of which includes multiple resolution stages, each resolution stage containing at least one block, and each block consists of a convolution, an instance normalization instanceNorm, and a leakyReLU nonlinear unit. A skip connection operation is set between the encoding part and the decoding part, and the skip connection operation transfers the features extracted by the encoding part to the decoding part to restore the feature map for accurate segmentation. In the auxiliary classification branch, features extracted from different resolution stages of nnUNet are used to distinguish between patches with colorectal cancer and patches without colorectal cancer.

[0143] The detection model is trained. During the training, on the one hand, a segmentation map of colorectal and colorectal cancer is predicted based on the pixel-level labels, and on the other hand, a classification prediction of whether the image contains colorectal cancer is performed based on the image-level labels; during the training, the sum of the segmentation loss and the classification loss is used as the total loss;

[0144] The trained detection model is used to detect colorectal cancer. The detection is successful when the detected tumor volume exceeds the set cutoff value.

[0145] Furthermore, if the modules / units integrated in the electronic device 20 are implemented as software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium. The computer-readable medium may include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0146] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0147] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0148] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for detecting colorectal cancer based on conventional enhanced CT, characterized in that: The steps include: Collecting a data set of abdominal CT images and preprocessing the data set; the data set includes colorectal cancer patient data and healthy colorectal data; The colorectal and colorectal cancer images of the abdominal CT images are labeled separately as the true labels of the detection model. The labels include pixel-level labels and image-level labels. A detection model is constructed, which uses nnUNet as the backbone network for pixel-level colorectal and colorectal cancer segmentation, and combines an auxiliary classification branch to integrate multi-scale features from nnUNet to achieve patch-level classification. The nnUNet includes an encoding part and a decoding part, each of which includes multiple resolution stages, each resolution stage contains at least one block, and each block consists of a convolution, an instance normalization instanceNorm, and a leakyReLU nonlinear unit. A skip connection operation is set between the encoding part and the decoding part, and the skip connection operation transmits the features extracted by the encoding part to the decoding part to recover the feature map for accurate segmentation. In the auxiliary classification branch, features extracted from different resolution stages of nnUNet are used to distinguish between patches with colorectal cancer and patches without colorectal cancer; The detection model is trained to predict colorectal and colorectal cancer segmentation maps based on pixel-level labels and to predict whether a patient has colorectal cancer based on image-level labels; During training, the sum of segmentation loss and classification loss is the total loss; The trained detection model is used to detect colorectal cancer. The detection is successful when the detected tumor volume exceeds the set cutoff value.

2. The method for detecting colorectal cancer based on conventional enhanced CT according to claim 1, characterized in that: The preprocessing includes excluding abnormal data, data enhancement processing and normalization processing; The abnormal data includes: Repeated CT image data of the same patient; CT image data that does not cover the entire abdomen and pelvis; CT images containing severe artifacts; CT images of a patient presenting with severe ascites; CT images of a patient with residual positive contrast agent in the colorectum; The data enhancement processing includes rotation, scaling, Gaussian noise, Gaussian blur, brightness, contrast, low-resolution simulation, gamma and mirroring.

3. The method for detecting colorectal cancer based on conventional enhanced CT according to claim 1, characterized in that: The colorectum and colorectal cancer of the abdominal CT image are respectively marked as follows: Use the automatic segmentation model to perform preliminary segmentation on the data set to obtain a rough segmentation map; Correct and unify the rough segmentation map, and annotate the image-level labels corresponding to the colorectum and colorectal cancer. Pixel-level labels for background, colorectal, and colorectal cancer.

4. The method for detecting colorectal cancer based on conventional enhanced CT according to claim 1, characterized in that: The auxiliary classification branch includes a global average pooling layer, a connection node, a fully connected layer and a Softmax layer; multiple global average pooling layers are connected in parallel through the connection node, and then connected to the series fully connected layer, and the fully connected layer is connected to the Softmax layer.

5. The method for detecting colorectal cancer based on conventional enhanced CT according to claim 4, characterized in that: In the detection model, a global average pooling operation is used to convert feature maps obtained at different resolution stages to the same size, and then the feature maps are concatenated and fed into two fully connected layers and a softmax operation for classification.

6. The method for detecting colorectal cancer based on conventional enhanced CT according to claim 1, characterized in that: The segmentation loss is the sum of the cross entropy loss and the Dice loss; The segmentation cross entropy loss L CE Expressed as: The segmentation Dice loss L DC Expressed as: in, is the number of categories, HWD is the size of the original CT image; I(·) is the indicator function, when the true label of voxel t is category n, the function value is 1, otherwise it is 0; Y t is the true label of voxel t, Pred t,n is the segmentation probability that the t-th voxel belongs to category n; The classification loss represents the image-level cross entropy loss, and the image-level cross entropy loss L cls Expressed as: Among them, y c , p c They correspond to the true category label and prediction result of category c respectively, and C is the total number of categories.

7. The method for detecting colorectal cancer based on conventional enhanced CT according to claim 1, characterized in that: If the predicted tumor volume exceeds the cutoff value and the Dice index between the predicted tumor and the labeled tumor is greater than 0, the tumor detection is considered correct; when the predicted tumor volume is lower than the cutoff value, or the Dice index is equal to 0, the tumor detection is considered incorrect.

8. A colorectal cancer detection system based on enhanced CT, characterized in that: A method for detecting colorectal cancer based on conventional enhanced CT, as applied to any one of claims 1 to 7, comprising a preprocessing module, a labeling module, a model building module, a model training module, and a detection module; The preprocessing module is used to collect a data set of abdominal CT images and preprocess the data set; the data set includes colorectal cancer patient data and healthy colorectal data; The labeling module is used to label the colorectal and colorectal cancer in the abdominal CT image respectively as the true labels of the detection model, and the labels include pixel-level labels and image-level labels; The model building module is used to build a detection model, which uses nnUNet as the backbone network for pixel-level colorectal and colorectal cancer segmentation, and combines an auxiliary classification branch to integrate multi-scale features from nnUNet to achieve patch-level classification; the nnUNet includes an encoding part and a decoding part, and the encoding part and the decoding part include multiple resolution stages, each resolution stage contains at least one block, and each block consists of a convolution, an instance normalization instanceNorm, and a leakyReLU nonlinear unit; a jump connection operation is set between the encoding part and the decoding part, and the jump connection operation transfers the features extracted by the encoding part to the decoding part to restore the feature map for accurate segmentation; in the auxiliary classification branch, the features extracted by different resolution stages of nnUNet are used to distinguish between blocks with colorectal cancer and blocks without colorectal cancer; The model training module is used to train the detection model. During training, on the one hand, the colorectal and colorectal cancer segmentation maps are predicted based on pixel-level labels, and on the other hand, the classification prediction of whether there is colorectal cancer is based on image-level labels; During training, the sum of segmentation loss and classification loss is the total loss; The detection module is used to detect colorectal cancer using the trained detection model, and the detection is successful when the detected tumor volume exceeds the set cutoff value.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores computer program instructions that can be executed by the at least one processor, and the computer program instructions are executed by the at least one processor so that the at least one processor can execute the colorectal cancer detection method based on conventional enhanced CT as described in any one of claims 1-7.

10. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, the method for detecting colorectal cancer based on conventional enhanced CT according to any one of claims 1 to 7 is implemented.

Images