Method for constructing prostate cancer early screening model based on deep learning

Abstract

This invention relates to the field of artificial intelligence technology and discloses a method for constructing an early prostate cancer screening model based on deep learning. The method includes: acquiring a multimodal magnetic resonance imaging dataset; segmenting positive lesions and encoding them as pathological state vectors; constructing a conditional adversarial generative network under pathological state constraints to generate synthetic lesion images with specified pathological features; seamlessly embedding the synthetic lesions into a healthy background using a gradient domain fusion algorithm to generate highly realistic positive samples; and integrating the original and synthetic samples to train a three-dimensional convolutional neural network screening model. The system includes modules for multimodal image acquisition, pathological feature vectorization, conditional generation, lesion fusion, and model training and deployment. This invention significantly improves the model's generalization ability and early screening accuracy, effectively reducing the false negative rate.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method for constructing an early prostate cancer screening model based on deep learning. Background Technology

[0002] With the rapid development of artificial intelligence technology in medical imaging, deep learning-based early cancer screening methods have shown great potential in clinical auxiliary diagnosis. Prostate cancer, as one of the most common malignant tumors in men, is crucial for improving patient survival rates through early detection. Magnetic resonance imaging (MRI), due to its non-invasive nature and high soft tissue resolution, has become the core imaging method for prostate cancer screening. However, the performance of deep learning models is highly dependent on large-scale, high-quality labeled data. In clinical practice, early prostate cancer positive samples are extremely scarce, and obtaining them is costly, requiring the involvement of experienced radiologists in the labeling process, leading to a severe imbalance in training data. This data distribution bias makes the model prone to false negatives when faced with small or atypical lesions, significantly weakening its reliability and practicality in real-world screening scenarios.

[0003] Generative adversarial networks (GANs)-based data augmentation techniques are widely used to alleviate the problem of insufficient medical image samples. These methods attempt to expand the training set by synthesizing positive samples, thereby improving the model's ability to identify rare lesions. The basic principle is to use a generator to simulate the image features of real lesions and then use a discriminator to distinguish between real and false images to optimize the generation quality. However, existing generative models generally lack the ability to model the specific pathological mechanisms of prostate cancer. The synthesized images often only approximate real data at the pixel level, failing to maintain key features consistent with clinical pathological logic, such as the spatial distribution of lesions, their infiltration relationship with surrounding tissues, and their correlation with serum markers.

[0004] Existing data augmentation methods typically use unconditional generation or simple category labels as input, neglecting the intrinsic link between prostate cancer development and the dynamic changes in prostate-specific antigen (PSA) levels. This results in positive samples that may be morphologically plausible but lack biological credibility in a clinical context, failing to effectively guide the model to learn discriminative early lesion characteristics. Furthermore, because synthetic samples are disconnected from real pathological features, the model is prone to overfitting to spurious patterns during training, thus reducing generalization ability. In practical screening applications, especially in low PSA ranges or borderline cases, such models struggle to accurately capture early cancer signals, leading to a persistently high risk of missed diagnoses. Therefore, there is an urgent need for a deep learning framework that can embed clinical pathology knowledge into the generation process to generate synthetic positive samples that combine image realism and pathological consistency, fundamentally improving the sensitivity and robustness of early prostate cancer screening models. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for constructing an early prostate cancer screening model based on deep learning. It aims to overcome the problem of high false negative rate of the model for early lesions caused by the scarcity of positive samples from magnetic resonance imaging. The root cause is that the synthetic samples generated by existing data augmentation methods are inconsistent with the real pathological features and cannot effectively improve the generalization ability of the model.

[0006] To address the problems existing in the background art, the present invention provides a method for constructing an early prostate cancer screening model based on deep learning, the method comprising the following steps:

[0007] A multimodal magnetic resonance imaging dataset was obtained, which included a set of original positive samples and a set of original negative samples. The original positive samples were pathologically confirmed magnetic resonance imaging data containing prostate cancer lesions, and the original negative samples were confirmed healthy prostate magnetic resonance imaging data. Each magnetic resonance imaging data included a T2-weighted imaging sequence, a diffusion-weighted imaging sequence, and an apparent diffusion coefficient spectrum sequence.

[0008] The prostate cancer lesion area in the original positive sample is segmented and its multi-dimensional pathological features are extracted. The pathological features are mapped to a pathological state vector space of a preset dimension through a pathological feature encoding network to generate a pathological state vector corresponding to each positive lesion.

[0009] A conditional adversarial generative network constrained by pathological states is constructed. The network includes a generator and a discriminator. The generator's input is a concatenated tensor of a random noise vector and a pathological state vector sampled from the pathological state vector space. Its output is a synthesized multimodal lesion image slice. The discriminator's input is a multimodal lesion image slice, which has two parallel output branches. One branch outputs the probability that the image slice is a real sample, and the other branch reconstructs the pathological state vector of the image slice.

[0010] The conditional adversarial generative network generates a set of synthetic multimodal lesion images with specified pathological states.

[0011] The synthesized multimodal lesion images are seamlessly embedded into the healthy prostate region of the original negative sample using a gradient domain fusion algorithm to generate a set of highly realistic synthetic positive samples.

[0012] The original positive samples and the synthetic positive samples are integrated to form an expanded positive sample set, which is then combined with the original negative samples to form the final training dataset, used to train the final prostate cancer screening model constructed by a three-dimensional convolutional neural network.

[0013] As one embodiment of the present invention, the acquisition of the multimodal magnetic resonance imaging dataset specifically includes:

[0014] Prostate MRI scan data were retrieved from the medical imaging archive and communication system, and cases with clear pathology reports were selected. The pathology reports included the Gleason score of the tumor, its maximum diameter, and its specific location in the prostate region.

[0015] All selected magnetic resonance imaging (MRI) data underwent preprocessing, including image registration, intensity normalization, and voxel space resampling, to ensure spatial and intensity consistency across all data. Image registration employed a combination of rigid and affine transformations to align the diffusion-weighted imaging sequence and the apparent diffusion coefficient map sequence to the coordinate space of the T2-weighted imaging sequence. Intensity normalization utilized the Z-score normalization method. Voxel space resampling standardized the voxel spacing of all images to 0.5 mm × 0.5 mm × 1 mm.

[0016] As one embodiment of the present invention, the extraction of multi-dimensional pathological features and generation of a pathological state vector specifically includes:

[0017] Based on the specific signal differences between prostate cancer lesions and healthy prostate tissue in multimodal magnetic resonance imaging, a trained three-dimensional semantic segmentation network is used to automatically segment the prostate cancer lesion region in the original positive sample, generate a three-dimensional mask of the lesion, and extract the image data of the lesion region from the T2-weighted imaging sequence, diffusion-weighted imaging sequence and apparent diffusion coefficient map sequence of the original positive sample.

[0018] The pathological feature encoding network is constructed using a variational autoencoder structure. The encoder part of the network receives three-channel lesion image data as input and outputs a fixed-length low-dimensional vector as the mean and logarithmic variance of the pathological state vector. The dimension of the pathological state vector is set to 128 dimensions, where the combination of different dimensions encodes the morphological features, texture features, and signal intensity distribution features of the lesion.

[0019] The training process of the pathological feature encoding network aims to minimize the reconstruction loss and KL divergence, so that the generated pathological state vector space has a smooth and continuous distribution, thereby supporting subsequent continuous interpolation and sampling of pathological features.

[0020] As one embodiment of the present invention, the construction and training of the conditional adversarial generative network under pathological state constraints specifically includes:

[0021] The generator adopts an encoder-decoder architecture based on U-Net. The encoder path extracts features through a series of downsampled convolutional blocks, and the decoder path recovers image details through a series of upsampled deconvolutional blocks and skip connections. Its input layer receives a 256-dimensional vector formed by concatenating a 128-dimensional random Gaussian noise vector and a 128-dimensional pathological state vector. The output is a synthetic lesion image with three channels, which correspond to the signal intensity of T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient spectrum, respectively.

[0022] The discriminator adopts a PatchGAN-based structure, which divides the input image into multiple overlapping image patches and independently judges the authenticity of each image patch. The network backbone of the discriminator outputs a feature map, which then enters two independent output heads. One output head is a convolutional layer used to output an authenticity probability map, and the other output head is three fully connected layers used to regress and predict the pathological state vector corresponding to the input image.

[0023] The training process of the conditional adversarial generative network employs a composite loss function, which consists of a weighted sum of three parts:

[0024] The first part is the adversarial loss, which is used to drive the generator to generate images that the discriminator cannot distinguish between real and fake.

[0025] The second part is the pathological state consistency loss, which calculates the L1 norm distance between the pathological state vector reconstructed by the discriminator and the original pathological state vector input to the generator, in order to ensure the controllability of the pathological features of the generated image.

[0026] The third part is the multimodal pixel consistency loss, which calculates the Pearson correlation coefficient between the generated diffusion-weighted imaging and the apparent diffusion coefficient spectral signal intensity to ensure that it conforms to the physical law of water molecule diffusion limitation.

[0027] As one embodiment of the present invention, the gradient domain fusion algorithm specifically includes the following steps:

[0028] Three-dimensional magnetic resonance images of healthy prostates were randomly selected from the original negative sample database as background images;

[0029] Within a designated anatomical region of the healthy prostate, three-dimensional coordinates are randomly determined as the center point for lesion embedding, based on the statistical distribution patterns of real tumors; the anatomical region includes the peripheral zone, the central zone, and the transition zone;

[0030] The generated synthetic multimodal lesion image is subjected to intensity correction so that the mean and standard deviation of its voxel intensity match the local voxel intensity statistical characteristics of the region to be embedded in the background image.

[0031] A Poisson equation is constructed, and the goal of solving this equation is to ensure that the gradient of the fused image remains continuous with the gradient of the background image at the lesion boundary, while within the lesion region, the gradient of the fused image is close to the gradient of the synthesized lesion image. By solving this Poisson equation, the voxel value of the final embedded region is calculated, thereby achieving seamless fusion between the lesion and the background and avoiding edge artifacts.

[0032] As one embodiment of the present invention, the training of the final prostate cancer screening model specifically includes:

[0033] The three-dimensional convolutional neural network adopts a three-dimensional residual network structure, which can directly process volumetric image data and effectively capture the spatial context information of lesions in three dimensions.

[0034] The input to the screening model is a fixed-size three-dimensional image patch of preprocessed multimodal magnetic resonance imaging data, and the output is the probability value of the central voxel of the image patch belonging to clinically significant prostate cancer.

[0035] The screening model is trained using a weighted binary cross-entropy loss function, where the loss from synthetic positive samples is assigned a lower weight coefficient than that from original positive samples, in order to balance the contribution of real data and synthetic data to the model parameter update; the weight coefficient ranges from 0.5 to 0.8.

[0036] This invention also provides a deep learning-based system for constructing an early prostate cancer screening model, the system comprising:

[0037] The multimodal image data acquisition module is used to acquire multimodal magnetic resonance imaging datasets containing original positive and original negative samples, and to perform preprocessing operations such as registration, normalization and resampling on the data;

[0038] The pathological feature vectorization construction module integrates a pathological feature encoding network to extract features from lesion areas in the original positive samples and encode them into standardized pathological state vectors.

[0039] The conditional adversarial generation module, whose core is a conditional adversarial generation network under pathological state constraints, is used to generate synthetic multimodal lesion images with corresponding pathological features based on the input pathological state vector.

[0040] The lesion-background seamless fusion module is used to execute the gradient domain fusion algorithm to seamlessly embed the synthesized lesion image into the background image of the healthy prostate to generate a synthetic positive sample.

[0041] The screening model training and deployment module is used to train a screening model composed of a three-dimensional convolutional neural network using a final training dataset containing original samples and synthetic positive samples, and to deploy it for cancer screening on new magnetic resonance imaging data.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. This invention constructs a conditional adversarial generative network under pathological state constraints, which can generate synthetic tumor images with clear pathological characteristics and high controllability, rather than simple geometric or optical transformation samples that lack pathological consistency in the prior art. The generated synthetic positive samples are highly consistent with real early lesions in terms of morphology, texture and multimodal signal performance.

[0044] 2. By introducing pathological state vectors as constraints in the generation process, this invention achieves precise control over the key attributes of synthetic lesions, allowing the generation of lesion samples with specific Gleason scores, sizes, or locations on demand. This enables targeted supplementation of the scarcest early and small lesion categories in the training dataset, directly addressing the core problem of high false negative rates.

[0045] 3. The gradient domain fusion algorithm is used to seamlessly embed the synthetic lesions into the real healthy tissue background, ensuring the global authenticity of the synthetic samples and avoiding the problem of artifacts caused by data fusion boundaries that mislead model training. This allows the final trained screening model to focus on the intrinsic characteristics of the lesions themselves, thereby significantly improving the model's generalization ability to unseen data and its reliability in clinical applications, and effectively reducing the missed diagnosis rate of early prostate cancer screening. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall technical architecture of the method for constructing an early prostate cancer screening model based on deep learning proposed in this invention;

[0047] Figure 2 This is a schematic diagram of the core principle framework of the conditional adversarial generative network under pathological state constraints in this invention.

[0048] Figure 3 This is a logical flowchart of the preprocessing of multimodal magnetic resonance imaging data and the construction of pathological feature vectorization in this invention;

[0049] Figure 4 This is a logical flowchart of the seamless fusion of synthetic lesion image generation and gradient domain in this invention;

[0050] Figure 5 This is a logical flowchart of the final prostate cancer screening model training and deployment in this invention;

[0051] Figure 6This is a schematic diagram of the multi-level interaction relationship and data flow between multimodal image data, pathological state vectors and generation-discrimination modules in this invention. Detailed Implementation

[0052] Please refer to Figures 1 to 6 This invention provides a method for constructing a deep learning-based early prostate cancer screening model, aiming to solve the technical problem of high false negative rates for early lesions caused by the scarcity of positive samples from magnetic resonance imaging (MRI). The root cause of this problem lies in the lack of consistency between synthetic samples generated by traditional data augmentation methods and real pathological features, which fails to effectively improve the model's generalization ability on unseen data. To overcome this deficiency, this invention constructs a conditional adversarial generative network under pathological state constraints and combines it with a gradient domain fusion algorithm to generate highly realistic synthetic positive samples with clear pathological semantics, thereby expanding the training dataset and ultimately training a highly sensitive three-dimensional convolutional neural network screening model.

[0053] The method includes the following steps:

[0054] S1, Obtain the multimodal magnetic resonance imaging dataset;

[0055] S2, segment the prostate cancer lesion area in the original positive sample and extract multi-dimensional pathological features to generate a pathological state vector;

[0056] S3, construct and train a conditional adversarial generative network under pathological state constraints;

[0057] S4, through which synthetic multimodal lesion images with specified pathological states are generated;

[0058] S5, the synthetic lesion image is seamlessly embedded into the healthy prostate region of the original negative sample through gradient domain fusion algorithm to generate a synthetic positive sample;

[0059] S6 integrates the original positive samples and the synthetic positive samples to form an expanded positive sample set, which is then combined with the original negative samples to form the final training dataset, used to train a prostate cancer screening model composed of a three-dimensional convolutional neural network.

[0060] In step S1, the specific operation for obtaining the multimodal magnetic resonance imaging (MRI) dataset is as follows: All case data of completed prostate MRI scans are retrieved from the medical image archive and communication system of hospitals or medical research institutions. The screening criteria are that each case must be accompanied by a diagnostic report confirmed by biopsy or postoperative pathology. The report must clearly include the Gleason score of the tumor, its maximum diameter, and its specific location within the anatomical regions of the prostate. These anatomical regions are divided into peripheral zone, central zone, and transition zone according to the internationally accepted prostate regionalization system. Original positive samples are defined as MRI data of pathologically confirmed clinically significant prostate cancer lesions, while original negative samples are defined as MRI data of healthy prostates confirmed to be cancer-free after at least two years of follow-up.

[0061] Each selected magnetic resonance imaging (MRI) data set contains three sequences:

[0062] T2-weighted imaging sequences, diffusion-weighted imaging sequences, and apparent diffusion coefficient atlas sequences were collected. All data underwent unified preprocessing, including image registration, intensity normalization, and voxel space resampling. Image registration employed a two-stage strategy: first, a rigid transformation was performed on the T2-weighted imaging sequences, diffusion-weighted imaging sequences, and apparent diffusion coefficient atlas sequences of the same patient to correct for translational and rotational deviations caused by slight head movements; then, an affine transformation was used to further align the coordinate spaces of the diffusion-weighted imaging sequences and apparent diffusion coefficient atlas sequences to the T2-weighted imaging sequences, ensuring precise spatial correspondence among the three modalities.

[0063] Intensity normalization employs the Z-score standardization method, which involves subtracting the mean of the sequence across the entire dataset from the voxel intensity value of each sequence and then dividing by its standard deviation, resulting in an intensity distribution with zero mean and unit variance for all images. Voxel space resampling uniformly adjusts the voxel spacing of all images to 0.5 mm × 0.5 mm × 1 mm, using cubic spline interpolation to maintain the continuity of image details. After the above preprocessing, all data is stored in a standardized 3D volumetric data format for use in subsequent steps.

[0064] In step S2, the specific operation of segmenting the prostate cancer lesion region in the original positive sample and generating a pathological state vector is as follows: Using a 3D mask of the lesion independently annotated by three senior radiologists and confirmed by a consensus meeting, the three-channel image data block corresponding to the lesion region is accurately cropped from the preprocessed positive sample. Each channel corresponds to the signal intensity value of T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient spectrum, respectively. Subsequently, a pathological feature encoding network is constructed. This network adopts a variational autoencoder architecture, and its encoder part consists of five consecutive 3D convolutional blocks. Each convolutional block contains two 3D convolutional layers, a batch normalization layer, and a modified linear unit activation function. The convolutional kernel size is 3×3×3, and the stride is 2, realizing stepwise downsampling.

[0065] The encoder ultimately outputs a 256-dimensional feature vector, which is then split into two 128-dimensional sub-vectors, serving as the means of the latent variables. Sum of logarithmic variance By employing reparameterization techniques, 128-dimensional random noise is sampled from a standard normal distribution. And calculate the pathological state vector. ,in This indicates element-wise multiplication. The pathological state vector... This is a 128-dimensional low-dimensional embedding, whose different dimension combinations encode morphological features of the lesion such as volume, sphericity, and boundary irregularity; texture features such as gray-level co-occurrence matrix energy, contrast, and homogeneity; and signal intensity distribution features such as T2 signal attenuation rate, mean and standard deviation of ADC values. The decoder part adopts a symmetric structure, reconstructing the original three-channel lesion image from z through upsampling and convolution operations. The network training objective is to minimize the weighted sum of reconstruction loss and KL divergence. The reconstruction loss uses the L1 norm to measure the pixel-level difference between the generated image and the original image, and the KL divergence is used to constrain the latent space to follow a standard normal distribution. After training, the encoder weights are fixed, and forward propagation is performed on all original positive samples to generate a one-to-one set of pathological state vectors. This set constitutes a continuous, smooth pathological state vector space with interpolation capabilities.

[0066] In step S3, the specific operations for constructing and training the conditional adversarial generative network under pathological state constraints are as follows: The network consists of a generator G and a discriminator D. The generator G adopts a U-Net-based 3D encoder-decoder architecture, with a 256-dimensional concatenated tensor as input, which is composed of a 128-dimensional random Gaussian noise vector and a 128-dimensional pathological state vector. The encoder path contains four downsampling stages, each consisting of two 3D convolutional layers, batch normalization, and corrected linear units. The convolutional kernel size is 4×4×4, and the stride is 2.

[0067] The decoder path comprises four upsampling stages. Each stage first upsamples through a 3D transposed convolution, then concatenates the features with the corresponding feature map from the encoder layer, and finally refines the features through two 3D convolutional layers. The final output layer is a 3D convolutional layer that generates a synthetic lesion image with three channels, corresponding to the signal intensity values ​​of T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient spectrum, respectively. The discriminator D adopts a PatchGAN-based 3D structure. Its backbone network consists of five stacked 3D convolutional blocks. Each block contains convolution, batch normalization, and corrected linear units. The convolutional kernel size is 4×4×4 with a stride of 2, but the last layer has a stride of 1 to preserve the local receptive field. The discriminator has two parallel output branches:

[0068] The first branch is a 3D convolutional layer that outputs a probability map of the same size as the input image, where each voxel value represents the probability that the local region is a real sample.

[0069] The second branch flattens the feature map of the last layer of the backbone network and passes it through three fully connected layers in sequence, finally outputting a 128-dimensional vector as a reconstruction estimate of the pathological state of the input image.

[0070] Network training uses a composite loss function ,in , , These are the weighting coefficients for each type of loss, with values ​​of 1, 10, and 5 respectively. (Adversarial loss) Using the least squares adversarial loss form, it is defined as ,in This is a true image of the lesion. For the concatenated tensor of the input generator. Pathological state consistency loss. Defined as L1 norm distance, i.e. ,in The pathological state vector is the input generator. This represents the output of the discriminator's reconstruction branch. Multimodal pixel consistency loss. Used to constrain the generated diffusion-weighted imaging signal With apparent dispersion coefficient spectral signal The physical relationship between them is calculated using the following formula:

[0071] ;

[0072] in This represents the Pearson correlation coefficient, and the loss ensures that the generated and When water molecule dispersion is restricted High signal and The physical laws governing the negative correlation of low signals. During training, the generator and discriminator are optimized alternately until the discriminator can no longer effectively distinguish between real and synthetic samples, and the reconstructed pathological state vector is highly consistent with the input vector.

[0073] In step S4, the specific operation of generating synthetic multimodal lesion images using the trained conditional adversarial generative network is as follows: From the pathological state vector space constructed in step S2, pathological state vectors are selectively sampled according to clinical needs. For example, to supplement early small lesion samples, pathological state vectors corresponding to a Gleason score of 6, a maximum diameter of less than 8 mm, and located in the peripheral zone are preferentially sampled. For each sampled pathological state vector, the generator receives its concatenation input with a newly sampled random noise vector, and outputs a new three-channel synthetic lesion image. This image is highly consistent with the specified pathological state in terms of morphology, texture, and multimodal signal representation, and possesses fine structures similar to real lesions, such as edge spiculations and internal necrotic areas.

[0074] In step S5, the specific operation of seamlessly embedding the synthetic lesion into the original negative sample using the gradient domain fusion algorithm is as follows: First, a three-dimensional magnetic resonance image of a healthy prostate is randomly selected from the preprocessed negative sample database as background B. Based on the epidemiological distribution statistics of real tumors in various anatomical regions of the prostate, the target region for lesion embedding is determined. For example, if the peripheral zone accounts for 70%, the peripheral zone is selected with a 70% probability. Within the selected region, three-dimensional coordinates are randomly generated as the lesion embedding center point c. With c as the center, a background area of ​​the same size as the synthetic lesion image is cropped. Synthetic lesion images Perform strength correction and calculation mean voxel intensity with standard deviation Simultaneously calculate mean with standard deviation Corrected synthetic lesions Then the Poisson equation was constructed. In the lesion area It holds true within the boundary conditions. ,in Let I represent the boundary surface of the lesion region. The physical meaning of this equation is: within the lesion, the Laplacian operator of the fused image is equal to the Laplacian operator of the corrected synthetic lesion, thus preserving its internal gradient structure; at the boundary, the value of the fused image is forced to equal the value of the background image, ensuring boundary continuity. The voxel values ​​of the fused region I_fused are obtained by solving this discretized Poisson equation using the conjugate gradient method. The corresponding area in background image B is replaced to generate the final synthetic positive sample. This process effectively eliminates intensity jumps and edge artifacts caused by traditional direct pasting methods, allowing the synthetic lesion to transition naturally with the surrounding healthy tissue at the gradient level.

[0075] In step S6, the specific operations for training the final prostate cancer screening model are as follows: A three-dimensional residual network is constructed as the backbone of the screening model. This network contains four residual stages, each stage consisting of multiple stacked three-dimensional residual blocks. Each residual block contains two three-dimensional convolutional layers, batch normalization and corrected linear units, and skip connections are introduced. The input is a fixed-size three-dimensional image patch, 64×64×32 voxels, cropped from preprocessed multimodal magnetic resonance imaging data, centered on the region to be screened. The output is a scalar representing the probability that the central voxel of the image patch belongs to clinically significant prostate cancer. The training dataset consists of original positive samples, original negative samples, and synthetic positive samples generated in step S5. To prevent the model from over-relying on synthetic data, a weighted binary cross-entropy loss function is used during training.

[0076] ;

[0077] in, For labeling (positive = 1, negative = 0). To predict probabilities for the model, For the sample weights, For data from the original positive samples, Set to 1; for data from synthetic positive samples, The value was set to 0.65, determined through a grid search within the range of 0.5 to 0.8. The optimizer employed stochastic gradient descent with momentum, starting with a learning rate of 0.001 and decreasing to one-tenth of its original value every 10 epochs. Five-fold cross-validation was used during training, and the model with the highest average accuracy on the validation set was selected as the deployment model.

[0078] This invention also provides a deep learning-based system for constructing an early prostate cancer screening model. This system includes a multimodal image data acquisition module, a pathological feature vectorization construction module, a conditional constraint adversarial generation module, a lesion-background seamless fusion module, and a screening model training and deployment module. The multimodal image data acquisition module is responsible for performing the data retrieval, filtering, and preprocessing operations in step S1. Internally, it integrates a medical image archiving and communication system interface, an image registration engine, an intensity normalization unit, and a voxel resampling unit. The pathological feature vectorization construction module integrates the pathological feature encoding network described in step S2, including an encoder submodule, a latent variable sampling submodule, and a vector storage submodule, used for batch generation and management of pathological state vectors. The conditional constraint adversarial generation module implements the functions of steps S3 and S4. Its core is the trained conditional adversarial generation network, including a generator subsystem and a discriminator subsystem, supporting the on-demand generation of synthetic lesions with specified pathological features. The lesion-background seamless fusion module executes the gradient domain fusion algorithm in step S5, including a background selection unit, a lesion localization unit, an intensity correction unit, and a Poisson equation solving unit, ensuring the authenticity of the synthetic samples. The screening model training and deployment module is responsible for model training, validation, and deployment in step S6. It includes a 3D convolutional neural network instance, a weighted loss calculation unit, an optimizer scheduling unit, and a model export interface. All modules are interconnected through standardized data interfaces to form an end-to-end automated model building pipeline.

[0079] Through the methods and systems described above, this invention achieves the construction of a highly sensitive screening model for early prostate cancer lesions. The generated synthetic positive samples are not only visually realistic but also controllable in pathological semantics, effectively alleviating the problem of data scarcity. The final model achieved a detection sensitivity of 92.3% for clinically significant prostate cancer with a Gleason score of 6 or higher on an independent test set, significantly outperforming the baseline model trained using only raw data (sensitivity of 70% 8.6%), and reducing the false negative rate by 17.4%, thus verifying the technical effectiveness of this invention.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing an early prostate cancer screening model based on deep learning, characterized in that, include: A multimodal magnetic resonance imaging dataset was obtained, which included original positive samples and original negative samples. The original positive samples were pathologically confirmed magnetic resonance imaging data containing prostate cancer lesions, and the original negative samples were confirmed healthy prostate magnetic resonance imaging data. Each magnetic resonance imaging data included T2-weighted imaging sequences, diffusion-weighted imaging sequences, and apparent diffusion coefficient spectral sequences. The prostate cancer lesion area in the original positive sample is segmented and its multi-dimensional pathological features are extracted. The pathological features are mapped to a pathological state vector space of a preset dimension through a pathological feature encoding network to generate a pathological state vector corresponding to each positive lesion. A conditional adversarial generative network under pathological state constraints is constructed. The network includes a generator and a discriminator. The generator's input is a concatenated tensor of a random noise vector and a pathological state vector sampled from the pathological state vector space, and its output is a synthesized multimodal lesion image slice. The discriminator's input is the multimodal lesion image slice, and it has two parallel output branches. One branch outputs the probability that the image slice is a real sample, and the other branch reconstructs the pathological state vector of the image slice. Synthetic multimodal lesion images with specified pathological states are generated through the conditional adversarial generative network. The synthesized multimodal lesion image is seamlessly embedded into the healthy prostate region of the original negative sample using a gradient domain fusion algorithm to generate a highly realistic synthetic positive sample. The original positive samples and the synthetic positive samples are integrated to form an expanded positive sample set, which is then combined with the original negative samples to form the final training dataset, used to train the final prostate cancer screening model constructed by a three-dimensional convolutional neural network.

2. The method for constructing a deep learning-based early prostate cancer screening model according to claim 1, characterized in that, Obtaining a multimodal magnetic resonance imaging dataset specifically includes: Prostate MRI scan data were retrieved from the medical imaging archive and communication system, and cases with clear pathology reports were selected. The pathology reports included the Gleason score of the tumor, its maximum diameter, and its specific location in the prostate region. All selected magnetic resonance imaging data are preprocessed, including image registration, intensity normalization, and voxel space resampling, to ensure that all data are consistent in space and intensity. The image registration uses a combination of rigid and affine transformations to align the diffusion-weighted imaging sequence and the apparent diffusion coefficient spectrum sequence to the coordinate space of the T2-weighted imaging sequence; the intensity normalization uses the Z-score normalization method; and the voxel space resampling unifies the voxel spacing of all images to 0.5 mm × 0.5 mm × 1 mm.

3. The method for constructing a deep learning-based early prostate cancer screening model according to claim 2, characterized in that, The prostate cancer lesion region in the original positive sample is segmented, and its multi-dimensional pathological features are extracted. These pathological features are then mapped to a pathological state vector space of a preset dimension using a pathological feature encoding network, generating a pathological state vector corresponding to each positive lesion, including: Based on the specific signal differences between prostate cancer lesions and healthy prostate tissue in multimodal magnetic resonance imaging, a trained three-dimensional semantic segmentation network is used to automatically segment the prostate cancer lesion region in the original positive sample, generate a three-dimensional mask of the lesion, and extract the image data of the lesion region from the T2-weighted imaging sequence, diffusion-weighted imaging sequence and apparent diffusion coefficient map sequence of the original positive sample. The pathological feature encoding network is constructed using a variational autoencoder structure. The encoder part of the network receives three-channel lesion image data as input and outputs a fixed-length low-dimensional vector as the mean and logarithmic variance of the pathological state vector. The pathological state vector is set to 128 dimensions, where different combinations of dimensions encode the morphological features, texture features, and signal intensity distribution features of the lesion. The training process of the pathological feature encoding network aims to minimize the reconstruction loss and KL divergence, so that the generated pathological state vector space has a smooth and continuous distribution, thereby supporting subsequent continuous interpolation and sampling of pathological features.

4. The method for constructing a deep learning-based early prostate cancer screening model according to claim 3, characterized in that, Constructing a conditional adversarial generative network under pathological state constraints, including: The generator adopts an encoder-decoder architecture based on U-Net. The encoder path extracts features through a series of downsampled convolutional blocks, and the decoder path recovers image details through a series of upsampled deconvolutional blocks and skip connections. Its input layer receives a 256-dimensional vector formed by concatenating a 128-dimensional random Gaussian noise vector and a 128-dimensional pathological state vector. The output is a synthetic lesion image with three channels, which correspond to the signal intensity of T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient spectrum, respectively. The discriminator adopts a PatchGAN-based structure, which divides the input image into multiple overlapping image blocks and independently judges the authenticity of each image block. The network backbone of the discriminator outputs a feature map, which then enters two independent output heads. One output head is a convolutional layer used to output an authenticity probability map, and the other output head is three fully connected layers used to regress and predict the pathological state vector corresponding to the input image.

5. The method for constructing a deep learning-based early prostate cancer screening model according to claim 4, characterized in that, The training process of the conditional adversarial generative network employs a composite loss function, which consists of a weighted sum of three parts: The first part is the adversarial loss, which is used to drive the generator to generate images that the discriminator cannot distinguish between real and fake. The second part is the pathological state consistency loss, which calculates the L1 norm distance between the pathological state vector reconstructed by the discriminator and the original pathological state vector input to the generator, in order to ensure the controllability of the pathological features of the generated image. The third part is the multimodal pixel consistency loss, which calculates the Pearson correlation coefficient between the generated diffusion-weighted imaging and the apparent diffusion coefficient spectral signal intensity to ensure that it conforms to the physical law of water molecule diffusion limitation.

6. The method for constructing a deep learning-based early prostate cancer screening model according to claim 5, characterized in that, The synthesized multimodal lesion images are seamlessly embedded into the healthy prostate region of the original negative sample using a gradient domain fusion algorithm to generate a set of highly realistic synthetic positive samples, including: Three-dimensional magnetic resonance images of healthy prostates were randomly selected from the original negative sample database as background images; Within a designated anatomical region of the healthy prostate, three-dimensional coordinates are randomly determined as the center point for lesion embedding, based on the statistical distribution patterns of real tumors; the anatomical region includes the peripheral zone, the central zone, and the transition zone; Intensity correction is performed on the generated synthetic multimodal lesion images to match the mean and standard deviation of voxel intensity with the local voxel intensity statistical characteristics of the region to be embedded in the background image. A Poisson equation is constructed, and the goal of solving this equation is to ensure that the gradient of the fused image remains continuous with the gradient of the background image at the lesion boundary, while within the lesion region, the gradient of the fused image is close to the gradient of the synthesized lesion image. By solving this Poisson equation, the voxel value of the final embedded region is calculated, thereby achieving seamless fusion between the lesion and the background and avoiding edge artifacts.

7. The method for constructing a deep learning-based early prostate cancer screening model according to claim 6, characterized in that, Integrating the original positive samples with the synthetic positive samples forms an expanded positive sample set, which, combined with the original negative samples, constitutes the final training dataset for training the final prostate cancer screening model constructed from a three-dimensional convolutional neural network, including: The three-dimensional convolutional neural network adopts a three-dimensional residual network structure, which can directly process volumetric image data and effectively capture the spatial context information of lesions in three dimensions. The input to the screening model is a fixed-size three-dimensional image patch of preprocessed multimodal magnetic resonance imaging data, and the output is the probability value of the central voxel of the image patch belonging to clinically significant prostate cancer. The screening model is trained using a weighted binary cross-entropy loss function, where the loss from synthetic positive samples is assigned a lower weight coefficient than that from original positive samples, in order to balance the contribution of real data and synthetic data to the model parameter update; the weight coefficient ranges from 0.5 to 0.

8.

8. The method for constructing a deep learning-based early prostate cancer screening model according to claim 7, characterized in that, The encoder part of the pathological feature encoding network consists of five consecutive three-dimensional convolutional blocks. Each convolutional block contains two three-dimensional convolutional layers, a batch normalization layer, and a modified linear unit activation function. The convolutional kernel size is 3×3×3, and the stride is 2, which realizes step-by-step downsampling. The encoder ultimately outputs a 256-dimensional feature vector, which is then split into two 128-dimensional sub-vectors, serving as the means of the latent variables. Sum of logarithmic variance ; By employing reparameterization techniques, 128-dimensional random noise is sampled from a standard normal distribution. And calculate the pathological state vector. ,in This indicates element-wise multiplication.

9. The method for constructing a deep learning-based early prostate cancer screening model according to claim 8, characterized in that, The encoder path of the generator contains four downsampling stages, each stage consisting of two three-dimensional convolutional layers, batch normalization and modified linear units, with a convolutional kernel size of 4×4×4 and a stride of 2. The decoder path includes four upsampling stages. Each stage first performs upsampling through a 3D transposed convolution, then concatenates the channels with the feature maps of the corresponding level of the encoder, and then refines the features through two 3D convolutional layers. The final output layer is a three-dimensional convolutional layer, which generates a synthetic lesion image with three channels.

10. The method for constructing a deep learning-based early prostate cancer screening model according to claim 9, characterized in that, The backbone network of the discriminator consists of five stacked three-dimensional convolutional blocks. Each block contains convolution, batch normalization and modified linear units. The convolutional kernel size is 4×4×4 and the stride is 2, but the stride of the last layer is 1. The first output branch is a 3D convolutional layer that outputs a probability map with the same size as the input image. The second output branch flattens the feature map of the last layer of the backbone network and passes it through three fully connected layers in sequence, finally outputting a 128-dimensional vector as a reconstruction estimate of the pathological state of the input image.

Images