Early liver cancer intelligent diagnosis system and method based on multi-temporal enhanced CT (Computed Tomography)

Abstract

The invention relates to the technical field of medical information systems, in particular to an early-stage liver cancer intelligent diagnosis system and method based on multi-temporal enhanced CT, and the system comprises a CT image module which is used for collecting a multi-temporal enhanced CT image of a patient; the image super-resolution reconstruction module is connected with the CT image module; the system comprises an image super-resolution reconstruction module, a radiomics feature module connected with the image super-resolution reconstruction module, and a depth feature module connected with the image super-resolution reconstruction module and used for constructing a depth convolutional neural network model; the knowledge distillation module is connected with the depth feature module; the early diagnosis module is connected with the radiomics feature module, the depth feature module and the knowledge distillation module and is used for integrating the high-throughput radiomics features and the depth features; based on an interpretable diagnosis strategy, an early liver cancer diagnosis result is generated, the prior knowledge of human experts is reserved, and the powerful feature learning ability of deep learning is fully utilized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information systems, and more specifically, to an intelligent diagnosis system and method for early liver cancer based on multi-phase enhanced CT. Background Art

[0002] With the increasing number of patients with chronic kidney disease, blood purification therapy is increasingly used in clinical practice. Liver cancer is one of the most common malignant tumors worldwide, and early diagnosis is crucial to improving patient survival. In recent years, with the continuous advancement of computed tomography (CT) technology, multi-phase enhanced CT has become an important tool for the diagnosis of liver cancer. However, traditional diagnostic methods mainly rely on the experience of radiologists, which have problems such as strong subjectivity and low efficiency.

[0003] In order to improve the accuracy and efficiency of diagnosis, researchers have begun to try to apply artificial intelligence technology to liver cancer diagnosis. Currently, the closest existing technologies are mainly concentrated in the following aspects:

[0004] First, some researchers have proposed methods based on radiomics. These methods assist diagnosis by extracting quantitative features from CT images, such as shape and texture. Although radiomics methods have improved the objectivity of diagnosis to a certain extent, there are still problems such as difficulty in feature selection and high requirements for image quality.

[0005] Secondly, with the development of deep learning technology, some researchers have tried to use models such as convolutional neural networks to directly learn features from CT images. Although such methods have achieved good results on some data sets, they face challenges such as poor interpretability and limited generalization ability.

[0006] In addition, some researchers have tried to combine radiomics and deep learning methods. However, these methods are often simple feature splicing or model fusion, which fails to give full play to the advantages of both methods.

[0007] Although existing technologies have made certain progress, the following problems still exist in the diagnosis of early liver cancer:

[0008] 1. The temporal information of multi-phase CT images is not fully utilized, and the dynamic enhancement characteristics of the tumor cannot be fully captured.

[0009] 2. Image quality limits the accuracy of diagnosis, especially for the detection of early small lesions.

[0010] 3. The model is not interpretable enough, making it difficult to gain the trust of clinicians and be widely used.

[0011] 4. Lack of effective integration of expert experience and failure to fully utilize existing medical knowledge.

[0012] 5. The diagnostic results lack foresight and cannot provide sufficient support for clinical decision-making.

[0013] In view of the above problems, there is an urgent need for an intelligent diagnostic system that can fully utilize multi-phase enhanced CT information, improve image quality, enhance model interpretability, effectively integrate expert experience, and have predictive capabilities. Summary of the invention

[0014] The present invention aims to solve the above technical problems and provide an intelligent diagnosis system and method for early liver cancer based on multi-phase enhanced CT. The system realizes the intelligence of the whole process from image acquisition, preprocessing, feature extraction to final diagnosis through the collaborative work of multiple modules.

[0015] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0016] The intelligent diagnosis system for early liver cancer based on multi-phase enhanced CT includes:

[0017] CT imaging module for:

[0018] Acquire multi-phase enhanced CT images of patients;

[0019] Preprocessing the multi-phase enhanced CT image;

[0020] An image super-resolution reconstruction module is connected to the CT image module and is used to:

[0021] Receiving the preprocessed multi-phase enhanced CT image sent by the CT image module;

[0022] Based on the preprocessed multi-phase enhanced CT image, super-resolution reconstruction is performed to obtain a reconstructed high-resolution CT image;

[0023] The radiomics feature module is connected to the image super-resolution reconstruction module and is used to:

[0024] extracting high-throughput radiomics features from the reconstructed high-resolution CT images;

[0025] A deep feature module, connected to the image super-resolution reconstruction module, is used to:

[0026] Build a deep convolutional neural network model;

[0027] Extracting deep features from the reconstructed high-resolution CT image using the deep convolutional neural network model;

[0028] The knowledge distillation module is connected to the deep feature module and is used to:

[0029] Form a training set based on expert annotations;

[0030] Training the deep feature module to form an interpretable diagnostic strategy;

[0031] An early diagnosis module, connected to the radiomics feature module, the deep feature module and the knowledge distillation module, is used to:

[0032] integrating the high-throughput radiomics features and the deep features;

[0033] Based on the explainable diagnostic strategy, early liver cancer diagnosis results are generated.

[0034] Preferably, the CT image module comprises:

[0035] An image acquisition unit, used for acquiring multi-phase enhanced CT images of patients;

[0036] An image preprocessing unit, used for performing normalization and standardization processing on the multi-phase enhanced CT image;

[0037] The image alignment unit is used to register the multi-phase enhanced CT image to eliminate the problem of uneven distribution of CT values.

[0038] Preferably, the image super-resolution reconstruction module comprises:

[0039] The encoding-decoding module is built on the U-Net architecture;

[0040] Multi-level residual structure, using a combination of global residual and local residual;

[0041] Atrous convolution unit, used to reduce the number of parameters;

[0042] The encoding-decoding module includes alternating convolutional layers and residual convolutional layers.

[0043] Preferably, the radiomics feature module is used for:

[0044] Extract the dynamic enhancement curve characteristics of lesions;

[0045] Extract sequence features;

[0046] Extract texture features;

[0047] Extract shape features;

[0048] Among them, the radiomics feature module uses a multi-channel convolutional neural network layer to extract high-throughput image features, and uses a deep convolutional neural network layer to fuse high-throughput image features with high-throughput image features.

[0049] Preferably, the deep feature module comprises:

[0050] Temporal attention model, used to model dynamic temporal attention;

[0051] Convolutional neural network model, used for feature extraction and fusion;

[0052] Among them, the deep feature module adopts void convolution and attention mechanism to build a deep convolutional network model, and uses image pyramid and residual structure to optimize the network model.

[0053] Preferably, the temporal attention model comprises:

[0054] Encoding module, including convolutional layer, fully connected layer and pooling layer;

[0055] Decoding module, including dense attention layer;

[0056] The encoding module is used to extract CT image features and lesion timing patterns, and the decoding module is used to extract correlations between features.

[0057] Preferably, the knowledge distillation module is used to:

[0058] Use expert experience as new samples;

[0059] Converting the expert experience into a feature vector through a feature extraction network;

[0060] Matching the feature vector with the prediction result of the deep feature module;

[0061] Optimizing the prediction results of the deep feature module.

[0062] Preferably, the early diagnosis module comprises:

[0063] A multi-feature fusion unit, used for integrating the high-throughput radiomics features and the deep features;

[0064] A prediction model building unit, used to build an early liver cancer recognition prediction model based on the explainable diagnostic strategy;

[0065] The diagnosis result generating unit is used to generate an early liver cancer diagnosis result by using the early liver cancer recognition prediction model.

[0066] As a preference, it also includes:

[0067] The lesion growth prediction module is connected to the early diagnosis module and is used to:

[0068] Predicting the growth of lesions based on the early liver cancer diagnosis results and the patient's historical medical data;

[0069] Generate early diagnosis recommendations.

[0070] The intelligent diagnosis method for early liver cancer based on multi-phase enhanced CT includes the following steps:

[0071] S1: acquiring a multi-phase enhanced CT image of a patient, and preprocessing the multi-phase enhanced CT image;

[0072] S2: performing super-resolution reconstruction on the preprocessed multi-phase enhanced CT image to obtain a reconstructed high-resolution CT image;

[0073] S3: extracting high-throughput radiomics features from the reconstructed high-resolution CT image;

[0074] S4: extracting deep features from the reconstructed high-resolution CT image using a deep convolutional neural network model;

[0075] S5: forming a training set based on expert annotations, training the deep convolutional neural network model, and forming an interpretable diagnostic strategy;

[0076] S6: Integrate the high-throughput radiomics features and the deep features to generate early liver cancer diagnosis results based on the interpretable diagnostic strategy;

[0077] S7: Based on the early liver cancer diagnosis results and the patient's historical medical data, predict the growth of the lesion and generate early diagnosis recommendations.

[0078] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0079] First, the present invention constructs a complete diagnostic framework, and each module works closely together to form a closed-loop diagnostic process. The CT image module provides high-quality raw data for subsequent analysis; the image super-resolution reconstruction module significantly improves the image quality and creates conditions for the detection of early small lesions; the radiomics feature module and the deep feature module extract image features from different angles, respectively, which not only retains the prior knowledge of human experts, but also makes full use of the powerful feature learning ability of deep learning; the knowledge distillation module cleverly integrates expert experience into the model, improving the interpretability and credibility of the system; the early diagnosis module integrates multi-source information to give the final diagnosis result.

[0080] Secondly, the present invention effectively solves several contradictions in the prior art. For example, by introducing super-resolution reconstruction technology, the image quality is improved without increasing the radiation dose; by combining radiomics and deep learning features, the interpretability of the features is guaranteed and the performance of the model is improved; and by using knowledge distillation technology, a balance is achieved between model complexity and interpretability.

[0081] Furthermore, the effects of the various modules of the present invention are highly complementary and synergistic. For example, super-resolution reconstruction not only improves image quality, but also provides richer information for subsequent feature extraction; the combination of radiomics features and deep learning features enables the model to capture both human-interpretable features and potential complex patterns; the introduction of knowledge distillation not only improves the interpretability of the model, but also makes up for the problem of insufficient training data to a certain extent.

[0082] Finally, the overall design of the present invention takes into account the actual needs of clinical applications. By providing interpretable diagnostic results and lesion growth predictions, it provides comprehensive support for clinicians' decision-making. This design not only improves the accuracy of diagnosis, but also enhances the usability and acceptability of the system in actual clinical settings.

[0083] In summary, the present invention achieves high accuracy, high efficiency and high interpretability in the diagnosis of early liver cancer through the organic combination of multiple innovative technologies, providing a powerful tool for improving the detection rate of early liver cancer and improving patient prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a logic block diagram of the overall structure of the system of the present invention.

[0085] Figure 2 It is a logic block diagram of the CT image module of the present invention.

[0086] Figure 3 It is a logic block diagram of the image super-resolution reconstruction module of the present invention.

[0087] Figure 4 It is a logical block diagram of the deep feature module of the present invention.

[0088] Figure 5 It is a logic block diagram of the early diagnosis module of the present invention. DETAILED DESCRIPTION

[0089] like Figure 1-5 As shown, the present invention provides an intelligent diagnosis system and method for early liver cancer based on multi-phase enhanced CT. The system achieves high-precision diagnosis of early liver cancer through multimodal data fusion and deep learning technology. The specific implementation of the present invention will be described in detail below.

[0090] First, the intelligent diagnosis system for early liver cancer based on multi-phase enhanced CT of the present invention includes the following modules: CT image module 1, image super-resolution reconstruction module 2, radiomics feature module 3, deep feature module 4, knowledge distillation module 5 and early diagnosis module 6. These modules work together to complete the intelligent diagnosis task of early liver cancer.

[0091] The CT image module 1 is used to collect multi-phase enhanced CT images of the patient and pre-process these images. In practical applications, the CT image module 1 is usually directly connected to the CT equipment of the hospital. Preferably, the present invention uses a 3.0T superconducting magnetic resonance scanner (such as Toshiba Aquilion ONE320 row) for image acquisition. The collected multi-phase enhanced CT images usually include images of the arterial phase, portal phase and delayed phase.

[0092] The preprocessing steps in CT image module 1 include image normalization and standardization. Normalization can scale the pixel values ​​to the range [0,1], while standardization can make the mean of the image 0 and the standard deviation 1. These preprocessing steps help improve the stability and accuracy of subsequent processing.

[0093] The image super-resolution reconstruction module 2 is connected to the CT image module 1, and is used to receive the pre-processed multi-phase enhanced CT image and perform super-resolution reconstruction to obtain a reconstructed high-resolution CT image. The present invention uses a deep learning model based on the U-Net architecture for super-resolution reconstruction. The mathematical expression of the model is as follows:

[0094] Y=F(X)=F n (F n-1 (...F1(X)))

[0095] Among them, X is the input low-resolution image, Y is the reconstructed high-resolution image, and F i represents the i-th layer transformation of the network.

[0096] Preferably, the super-resolution reconstruction model of the present invention adopts a multi-level residual structure combining global residual and local residual. The global residual can be expressed as:

[0097] Y=F(X)+X

[0098] The local residual is applied inside each convolutional block. This structure helps alleviate the gradient vanishing problem and improves the convergence speed and performance of the model.

[0099] In addition, the present invention also introduces dilated convolution to expand the receptive field and reduce the number of parameters. The mathematical expression of dilated convolution is:

[0100]

[0101] Among them, l is the expansion rate, F is the input feature map, and k is the convolution kernel.

[0102] The radiomics feature module 3 is connected to the image super-resolution reconstruction module 2, and is used to extract high-throughput radiomics features from the reconstructed high-resolution CT image. These features include, but are not limited to, shape features, texture features, intensity features, and wavelet features. For example, shape features may include volume, surface area, sphericity, etc.; texture features may include features derived from gray level co-occurrence matrix (GLCM), run length matrix (RLGL), etc.; intensity features may include statistics such as mean, variance, skewness, and kurtosis.

[0103] A notable feature of the present invention is the introduction of dynamic enhancement features. Specifically, the radiomics feature module 3 extracts the dynamic enhancement curve features of the lesion. This can be achieved by the following steps:

[0104] 1. Register multi-temporal images.

[0105] 2. Segment the region of interest (ROI) at each time phase.

[0106] 3. Calculate the average CT value within the ROI and obtain the time-density curve (TDC).

[0107] 4. Extract features from TDC, such as peak enhancement, peak time, wash-out rate, etc.

[0108] These dynamic features are important for distinguishing early liver cancer from benign lesions.

[0109] The deep feature module 4 is also connected to the image super-resolution reconstruction module 2 to construct a deep convolutional neural network model and extract deep features from the reconstructed high-resolution CT image. The present invention adopts an architecture based on a residual network (ResNet) and introduces an attention mechanism. The mathematical expression of the residual block is as follows:

[0110] y=F(x,{W i})+x

[0111] Among them, x and y are the input and output of the residual block respectively, F(x,{W i}) represents the residual mapping.

[0112] The attention mechanism can be expressed as follows:

[0113]

[0114] Among them, Q, K, and V represent query, key, and value respectively. k is the dimension of the key.

[0115] This structural design enables the model to automatically learn more effective feature representations, which is especially advantageous for detecting tiny lesions such as early liver cancer.

[0116] The knowledge distillation module 5 is connected to the deep feature module 4 to form a training set based on expert annotations and train the deep feature module to form an interpretable diagnostic strategy. The core idea of ​​knowledge distillation is to let a complex "teacher model" guide the learning of a relatively simple "student model". In the present invention, the diagnostic experience of radiology experts is regarded as the "teacher model" and the deep learning model is used as the "student model".

[0117] The loss function of knowledge distillation can be expressed as:

[0118] L=αL CE (y,σ(z s / T))+(1-α)L KL (σ(z t / T),σ(z s / T))

[0119] Among them, L CE is the cross entropy loss, L KL is the KL divergence, z t and z s are the output logits of the teacher model and the student model respectively, T is the temperature parameter, and α is the balance coefficient.

[0120] In this way, the diagnostic experience of experts can be effectively integrated into the deep learning model, improving the interpretability and credibility of the model.

[0121] The early diagnosis module 6 is connected with the radiomics feature module 3, the deep feature module 4 and the knowledge distillation module 5 to integrate high-throughput radiomics features and deep features and generate early liver cancer diagnosis results based on an interpretable diagnostic strategy. This module adopts the idea of ​​ensemble learning to fuse features from different sources to obtain more robust and accurate diagnostic results.

[0122] Specifically, the early diagnosis module 6 first normalizes the radiomics features and deep features, and then performs feature fusion by weighted summation:

[0123] F fused =w1F radiomics +w2F deep

[0124] Among them, F radiomics and F deep , represent radiomic features and deep features respectively, and w1 and w2 are the corresponding weights.

[0125] The fused features are input into a final classifier (such as a support vector machine or random forest) to obtain the diagnosis result. In order to improve the interpretability of the model, SHAP (SHapley Additive exPlanations) values ​​can also be used to explain the contribution of each feature to the final decision.

[0126] The method of the present invention achieves high-precision diagnosis of early liver cancer through multimodal data fusion and deep learning technology. Compared with traditional methods, the present invention has the following advantages:

[0127] 1. The use of multi-phase CT images fully captures the dynamic enhancement characteristics of the tumor and improves the sensitivity to early lesions.

[0128] 2. Through super-resolution reconstruction technology, the image quality is improved without increasing the radiation dose, which is conducive to the detection of tiny lesions.

[0129] 3. Combining manually designed radiomics features and automatically learned deep features, it not only retains the prior knowledge of human experts, but also makes full use of the powerful feature learning ability of deep learning.

[0130] 4. The knowledge distillation technology is introduced to incorporate expert experience into the model, which improves the interpretability and credibility of the model.

[0131] 5. Adopting integrated learning strategy, integrating multiple features and diagnostic strategies, it improves the stability and accuracy of diagnosis.

[0132] In summary, the present invention provides a comprehensive, efficient and accurate intelligent diagnosis system and method for early liver cancer, which is expected to play an important role in clinical practice and improve the detection rate and diagnostic accuracy of early liver cancer.

[0133] In one embodiment of the present invention, the structure of the CT image module 1 is further clarified. The module includes an image acquisition unit 11, an image preprocessing unit 12 and an image alignment unit 13. These three units work together to ensure the quality of the CT image input to the subsequent modules.

[0134] The image acquisition unit 11 is responsible for acquiring multi-phase enhanced CT images of the patient. Preferably, the present invention adopts dynamic enhanced scanning technology, and scans at different time points after intravenous injection of iodine contrast agent to obtain CT images of the arterial phase, portal phase and delayed phase. This multi-phase acquisition method can fully capture the blood supply characteristics of the tumor, which is of great significance for the detection of early liver cancer.

[0135] The image preprocessing unit 12 performs normalization and standardization on the acquired CT images. Normalization usually uses the minimum-maximum normalization method to scale the pixel values ​​to the range of [0,1], and standardization uses the Z-score standardization method to make the image mean 0 and the standard deviation 1. These preprocessing steps help eliminate the differences caused by different devices and scanning parameters and improve the stability of subsequent processing.

[0136] The image alignment unit 13 is responsible for registering the multi-phase CT images to eliminate the problem of uneven distribution of CT values. The present invention preferably adopts a non-rigid registration algorithm, such as B-spline registration or optical flow registration. Taking B-spline registration as an example, its deformation field can be expressed as:

[0137]

[0138] Where x is the image coordinate, c i is the control point, β 3 is the cubic B-spline basis function, and h is the control point spacing.

[0139] By minimizing the mutual information or mean square error between images, the optimal deformation field parameters can be obtained to achieve image alignment.

[0140] In one embodiment of the present invention, the structure of the image super-resolution reconstruction module 2 is further refined. The module includes an encoding-decoding module 21, a multi-level residual structure 22 and a dilated convolution unit 23. This design fully considers the characteristics of medical images and can effectively improve image resolution without introducing artifacts.

[0141] The encoding-decoding module 21 is built on the U-Net architecture. An important feature of U-Net is the skip connection, which can directly pass the low-level features of the encoder to the corresponding layer of the decoder, helping to restore the detailed information of the image. The skip connection can be expressed as:

[0142] y i =H i (x i )+F i (x i )

[0143] Among them, H i represents the identity mapping, F i represents the residual function.

[0144] The multi-level residual structure 22 uses a combination of global residual and local residual. The global residual helps to learn the residual mapping between input and output, while the local residual is applied inside each convolutional block to help the network learn more fine-grained features. This structure can be expressed as:

[0145]

[0146] Among them, F(X) represents the global residual map, R i (X)) represents the local residual map.

[0147] The dilated convolution unit 23 is used to reduce the number of parameters and expand the receptive field. An important parameter of the dilated convolution is the dilation rate, which determines the spacing between the convolution kernel elements. Preferably, the present invention adopts multi-scale dilated convolution, that is, dilated convolutions with different dilation rates are used in the same layer, and then the results are spliced. This can be expressed as:

[0148]

[0149] in, The expansion rate is d i The dilated convolution.

[0150] In one embodiment of the present invention, the function of the radiomics feature module 3 is further clarified. The module is responsible for extracting various types of radiomics features, including lesion dynamic enhancement curve features, sequence features, texture features and shape features.

[0151] The dynamic enhancement curve characteristics of the lesion reflect the blood supply characteristics of the tumor and are important indicators for distinguishing benign and malignant lesions. The present invention extracts such characteristics through the following steps:

[0152] 1. Segment the region of interest (ROI) on each phase image.

[0153] 2. Calculate the average CT value within the ROI and obtain the time-density curve (TDC).

[0154] 3. Extract features from TDC, such as peak enhancement (PE), time to peak (TTP), wash-out rate (WOR), etc.

[0155] These features can be expressed as:

[0156] PE=max(HU t )-HU0

[0157] TTP=argmax t (HU t )

[0158]

[0159] Among them, HU t represents the CT value at time t, and HU0 is the baseline CT value.

[0160] Sequence features reflect the change patterns of lesions in different phases. The present invention uses a method based on sequence pattern mining to extract such features. Specifically, the change pattern of CT value is encoded into a sequence, and then a sequence pattern mining algorithm (such as PrefixSpan) is used to discover frequent patterns.

[0161] Texture features reflect the heterogeneity inside the lesion. The present invention uses methods such as gray level co-occurrence matrix (GLCM) and run length matrix (RLGL) to extract texture features. Taking GLCM as an example, the following features can be calculated:

[0162] Texture features reflect the heterogeneity inside the lesion. The present invention uses methods such as gray level co-occurrence matrix (GLCM) and run length matrix (RLGL) to extract texture features. Taking GLCM as an example, the following features can be calculated:

[0163] Energy: ∑ i,j p(i,j) 2

[0164] Contrast ratio:∑ i,j |ij| 2 p(i,j)

[0165] Relevance:

[0166] where p(i,j) is an element in GLCM.

[0167] The shape feature reflects the geometric characteristics of the lesion. The shape features extracted by the present invention include but are not limited to:

[0168] Volume: V = ∑ x,y,z I(x,y,z)

[0169] Surface Area:

[0170] Sphericity:

[0171] Among them, I(x,y,z) is the binary image after segmentation.

[0172] In one embodiment of the present invention, the structure of the deep feature module 4 is further refined. The module includes a temporal attention model 41 and a convolutional neural network model 42, which work together to capture complex features in CT images.

[0173] The temporal attention model 41 is used to model dynamic temporal attention. The present invention adopts a temporal modeling method based on a self-attention mechanism. Specifically, for the input sequence X = [x1, x2, ..., x T ], first calculate the self-attention:

[0174]

[0175] Among them, Q, K, and V are query, key, and value matrices, respectively, obtained by linear transformation of X. Then, a gating mechanism is used to fuse the original features and the attention features:

[0176] g t =σ(W g [x t ,a t ]+b g )

[0177] y t =g t ⊙x t +(1-g t )⊙a t

[0178] Among them, a t The attention feature, g t is the gate value, and ⊙ represents element-wise multiplication.

[0179] The convolutional neural network model 42 is used to realize feature extraction and fusion. The present invention adopts a residual network-based architecture and introduces a dilated convolution and attention mechanism. The mathematical expression of the residual block is as follows:

[0180] y=F(x,{W i})+x

[0181] Among them, F(x,{W i}) represents the residual mapping.

[0182] The use of dilated convolution can expand the receptive field without increasing the number of parameters. The present invention uses multi-scale dilated convolution, that is, dilated convolutions with different dilation rates are used in the same layer, and then the results are spliced.

[0183] In addition, the present invention also uses spatial attention and channel attention mechanisms. Spatial attention can help the model focus on important areas in the image, while channel attention can highlight important feature channels. These two attention mechanisms can be expressed as:

[0184]

[0185] M c =σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0186] Among them, F is the input feature map, represents 7X7 convolution, and MLP represents multi-layer perceptron.

[0187] The structure of the temporal attention model 41 is further clarified. The model includes an encoding module 411 and a decoding module 412, which together extract and model the temporal features of the CT image.

[0188] The encoding module 411 includes a convolutional layer, a fully connected layer, and a pooling layer. The convolutional layer is used to extract local features and can be expressed as:

[0189]

[0190] Among them, * represents the convolution operation, and f is the activation function (such as ReLU).

[0191] The fully connected layer is used to integrate global information and can be expressed as:

[0192] y=f(Wx+b)

[0193] The pooling layer is used to reduce the feature dimension and increase the translation invariance of the model. The maximum pooling can be expressed as:

[0194]

[0195] Among them, R i,j represents the pooling area centered at (i, j).

[0196] The decoding module 412 includes a dense attention layer. The core idea of ​​the dense attention layer is to calculate attention at different scales and then fuse the results. This can be expressed as:

[0197] A k =softmax(f k (Q)f k (K) T )f k (V)

[0198] Y=Concat[A1,A2,...,A K ]W

[0199] Among them, f k represents the feature transformation of the kth scale, and W is the weight matrix used for feature fusion.

[0200] Through this design, the temporal attention model 41 can effectively capture the changing characteristics of CT images in the time dimension, providing an important basis for the diagnosis of early liver cancer.

[0201] In general, the present invention achieves a comprehensive analysis of multi-phase enhanced CT images by refining the structure and function of each module. From image acquisition, preprocessing, super-resolution reconstruction, to radiomics feature extraction, deep feature learning, and time series modeling, each step has been carefully designed to maximize the accuracy of early liver cancer diagnosis. This multimodal, multi-scale analysis method is expected to significantly improve the detection rate of early liver cancer and provide important support for the timely treatment of patients.

[0202] In one embodiment of the present invention, the function of the knowledge distillation module 5 is further clarified. The module is intended to effectively integrate expert experience into the deep learning model to improve the interpretability and credibility of the model. Specifically, the knowledge distillation module 5 performs the following steps:

[0203] First, expert experience is used as a new sample. The expert experience here may include the diagnosis results of CT images by radiologists, the interpretation of specific imaging features, etc. Preferably, the present invention records expert experience in a structured manner, such as using a predefined feature dictionary and scoring criteria.

[0204] Secondly, the expert experience is converted into a feature vector through the feature extraction network. The purpose of this step is to convert qualitative expert experience into quantifiable feature representation. The feature extraction network can use a pre-trained natural language processing model, such as BERT. For the input expert experience text T, the feature extraction process can be expressed as:

[0205] v=BERT(T)

[0206] Among them, v is the extracted feature vector.

[0207] Then, the feature vector is matched with the prediction result of the deep feature module. The core of this step is to calculate the similarity between the expert experience feature vector and the model prediction result. The present invention uses cosine similarity as the matching metric:

[0208]

[0209] Here, p is the prediction vector of the model.

[0210] Finally, the prediction results of the deep feature module are optimized. This step is achieved by introducing knowledge distillation loss to "distill" expert knowledge into the model. The knowledge distillation loss can be expressed as:

[0211]

[0212] Among them, z t and z sare the output logits of the teacher model (expert) and the student model (deep learning model), T is the temperature parameter, KL is the Kullback-Leibler divergence, CE is the cross entropy loss, and α is the balance coefficient.

[0213] In this way, the present invention achieves effective integration of expert knowledge and improves the interpretability and clinical acceptability of the model.

[0214] In claim 8, the structure of the early diagnosis module 6 is further refined. The module comprises a multi-feature fusion unit 61, a prediction model building unit 62 and a diagnosis result generating unit 63, which work together to generate a final diagnosis result.

[0215] The multi-feature fusion unit 61 is responsible for integrating high-throughput radiomics features and deep features. The present invention adopts a feature fusion method based on the attention mechanism. Specifically, for the radiomics feature F r and deep features F d , the fusion process can be expressed as:

[0216] w r =softmax(W r F r ),w d =softmax(W d F d )

[0217] F fused =w r F r +w d F d

[0218] Among them, W r and W d is a learnable weight matrix.

[0219] The prediction model building unit 62 builds an early liver cancer recognition prediction model based on an interpretable diagnostic strategy. The present invention adopts the idea of ​​ensemble learning and combines multiple basic classifiers to improve the generalization ability of the model. Preferably, the method uses random forest as an ensemble model. For the input feature x, the prediction of random forest can be expressed as:

[0220]

[0221] Where M is the number of decision trees, f m Represents the mth decision tree.

[0222] The diagnosis result generating unit 63 generates an early liver cancer diagnosis result using the early liver cancer recognition prediction model. In order to improve the interpretability of the result, the present invention also introduces the SHAP value to explain the contribution of each feature to the final decision. The SHAP value can be calculated by the following formula:

[0223]

[0224] Among them, F is the set of all features, S is the subset that does not contain feature i, and f S are the model predictions using only the feature subset S.

[0225] In one embodiment of the present invention, the present invention introduces a lesion growth prediction module 7, which further enhances the clinical value of the system. This module is connected to the early diagnosis module 6 and is used to predict the growth of the lesion and generate early diagnosis suggestions.

[0226] The lesion growth prediction module 7 first constructs a time series model based on the early liver cancer diagnosis results and the patient's historical medical data. The present invention preferably uses a long short-term memory network (LSTM) to capture the long-term dependency of lesion growth.

[0227] Based on the prediction results of LSTM, the lesion growth prediction module 7 also generates personalized early diagnosis suggestions. These suggestions take into account factors such as the growth rate of the lesion, the patient's age, and liver function status, providing comprehensive support for clinical decision-making.

[0228] Finally, the invention provides a complete intelligent diagnosis method for early liver cancer based on multi-phase enhanced CT. The method includes the following steps:

[0229] S1: Collect multi-phase enhanced CT images of patients and perform preprocessing. This step is usually completed in the radiology department of the hospital, using high-end CT equipment (such as 64-row and above spiral CT) for scanning. Preprocessing includes image denoising, standardization and other operations.

[0230] S2: Perform super-resolution reconstruction on the pre-processed multi-phase enhanced CT image to obtain a reconstructed high-resolution CT image. This step is completed using the aforementioned image super-resolution reconstruction module 2, which can significantly improve the image detail information.

[0231] S3: Extract high-throughput radiomic features from the reconstructed high-resolution CT images. This step is completed using the radiomic feature module 3, and the extracted features include multiple dimensions such as shape, texture, and intensity.

[0232] S4: Extract deep features from the reconstructed high-resolution CT images using a deep convolutional neural network model. This step is completed using the deep feature module 4, which can automatically learn complex image features.

[0233] S5: Based on the expert annotations, a training set is formed and the deep convolutional neural network model is trained to form an interpretable diagnostic strategy. This step is completed using the knowledge distillation module 5 to effectively integrate expert experience into the model.

[0234] S6: Integrate high-throughput radiomics features and deep features to generate early liver cancer diagnosis results based on an interpretable diagnostic strategy. This step is completed using the early diagnosis module 6, which comprehensively considers multiple features to improve the accuracy of diagnosis.

[0235] S7: Based on the early liver cancer diagnosis results and the patient's historical medical data, predict the growth of the lesion and generate early diagnosis suggestions. This step is completed using the lesion growth prediction module 7 to provide support for subsequent clinical decision-making.

[0236] Through this series of steps, the method of the present invention realizes the intelligence of the whole process from image acquisition to diagnosis suggestion generation, providing a powerful tool for accurate diagnosis of early liver cancer. Compared with traditional methods, this method has the following advantages:

[0237] 1. Multimodal data fusion: Combining radiomics features and deep learning features to comprehensively capture the imaging characteristics of tumors.

[0238] 2. Utilization of temporal information: Through multi-phase CT images and temporal attention models, the dynamic enhancement characteristics of the tumor are fully utilized.

[0239] 3. Knowledge integration: Through knowledge distillation technology, expert experience is effectively integrated into the model, improving interpretability.

[0240] 4. Predictive analysis: It not only provides current diagnostic results, but also predicts the future growth of lesions, providing more comprehensive information for clinical decision-making.

[0241] In summary, the intelligent diagnosis system and method for early liver cancer based on multi-phase enhanced CT provided by the present invention achieves high accuracy, high efficiency and high interpretability in the diagnosis of early liver cancer through the organic combination of multiple innovative technologies. This method is expected to play an important role in clinical practice, improve the detection rate of early liver cancer, and contribute to the timely treatment and improved prognosis of patients.

[0242] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto; any person familiar with the art who, within the scope disclosed by the present invention, makes substitutions or changes based on the scheme and improved concepts of the present invention shall be covered within the protection scope of the present invention.

Claims

1. An intelligent diagnosis system for early liver cancer based on multi-phase enhanced CT, characterized by: include: CT imaging module for: Acquire multi-phase enhanced CT images of patients; Preprocessing the multi-phase enhanced CT image; An image super-resolution reconstruction module is connected to the CT image module and is used to: Receiving the preprocessed multi-phase enhanced CT image sent by the CT image module; Based on the preprocessed multi-phase enhanced CT image, super-resolution reconstruction is performed to obtain a reconstructed high-resolution CT image; The radiomics feature module is connected to the image super-resolution reconstruction module and is used to: extracting high-throughput radiomics features from the reconstructed high-resolution CT images; A deep feature module, connected to the image super-resolution reconstruction module, is used to: Build a deep convolutional neural network model; Extracting deep features from the reconstructed high-resolution CT image using the deep convolutional neural network model; The knowledge distillation module is connected to the deep feature module and is used to: Form a training set based on expert annotations; Training the deep feature module to form an interpretable diagnostic strategy; An early diagnosis module, connected to the radiomics feature module, the deep feature module and the knowledge distillation module, is used to: integrating the high-throughput radiomics features and the deep features; Based on the explainable diagnostic strategy, early liver cancer diagnosis results are generated.

2. The system according to claim 1, characterized in that The CT image module includes: An image acquisition unit, used for acquiring multi-phase enhanced CT images of patients; An image preprocessing unit, used for performing normalization and standardization processing on the multi-phase enhanced CT image; The image alignment unit is used to register the multi-phase enhanced CT image to eliminate the problem of uneven distribution of CT values.

3. The system according to claim 1, characterized in that The image super-resolution reconstruction module comprises: The encoding-decoding module is built on the U-Net architecture; Multi-level residual structure, using a combination of global residual and local residual; Atrous convolution unit, used to reduce the number of parameters; The encoding-decoding module includes alternating convolutional layers and residual convolutional layers.

4. The system according to claim 1, characterized in that The radiomics feature module is used to: Extract the dynamic enhancement curve characteristics of lesions; Extract sequence features; Extract texture features; Extract shape features; Among them, the radiomics feature module uses a multi-channel convolutional neural network layer to extract high-throughput image features, and uses a deep convolutional neural network layer to fuse high-throughput image features with high-throughput image features.

5. The system according to claim 1, characterized in that The deep feature module includes: Temporal attention model, used to model dynamic temporal attention; Convolutional neural network model, used for feature extraction and fusion; Among them, the deep feature module adopts void convolution and attention mechanism to build a deep convolutional network model, and uses image pyramid and residual structure to optimize the network model.

6. The system according to claim 5, characterized in that The temporal attention model includes: Encoding module, including convolutional layer, fully connected layer and pooling layer; Decoding module, including dense attention layer; The encoding module is used to extract CT image features and lesion timing patterns, and the decoding module is used to extract correlations between features.

7. The system according to claim 1, characterized in that The knowledge distillation module is used to: Use expert experience as new samples; Converting the expert experience into a feature vector through a feature extraction network; Matching the feature vector with the prediction result of the deep feature module; Optimizing the prediction results of the deep feature module.

8. The system according to claim 1, characterized in that The early diagnosis module comprises: A multi-feature fusion unit, used for integrating the high-throughput radiomics features and the deep features; A prediction model building unit, used to build an early liver cancer recognition prediction model based on the explainable diagnostic strategy; The diagnosis result generating unit is used to generate an early liver cancer diagnosis result by using the early liver cancer recognition prediction model.

9. The system according to claim 1, characterized in that Also includes: The lesion growth prediction module is connected to the early diagnosis module and is used to: Predicting the growth of lesions based on the early liver cancer diagnosis results and the patient's historical medical data; Generate early diagnosis recommendations.

10. An intelligent diagnosis method for early liver cancer based on multi-phase enhanced CT, characterized in that: The following steps are involved: S1: acquiring a multi-phase enhanced CT image of a patient, and preprocessing the multi-phase enhanced CT image; S2: performing super-resolution reconstruction on the preprocessed multi-phase enhanced CT image to obtain a reconstructed high-resolution CT image; S3: extracting high-throughput radiomics features from the reconstructed high-resolution CT image; S4: extracting deep features from the reconstructed high-resolution CT image using a deep convolutional neural network model; S5: forming a training set based on expert annotations, training the deep convolutional neural network model, and forming an interpretable diagnostic strategy; S6: Integrate the high-throughput radiomics features and the deep features to generate early liver cancer diagnosis results based on the interpretable diagnostic strategy; S7: Based on the early liver cancer diagnosis results and the patient's historical medical data, predict the growth of the lesion and generate early diagnosis recommendations.

Images