Construction method of prostate cancer image recognition classification model based on deep learning

Abstract

The application relates to the technical field of medical image processing, in particular to a construction method of a prostate cancer image recognition classification model based on deep learning. The method comprises the following steps: acquiring a standardized multi-modal image set; constructing an inter-region characteristic ratio atlas according to the standardized multi-modal image set; constructing a region physical feature tensor according to the inter-region characteristic ratio atlas; dividing the standardized multi-modal image set into three regions of a peripheral zone, a transition zone and a central zone, introducing anatomical prior knowledge as a constraint condition, constructing a morphological physical mapping function set according to a nonlinear mapping function of the region physical feature tensor; generating a morphological physical joint feature map according to the region physical feature tensor and the morphological physical mapping function set; and constructing a physical characteristic progressive constraint set according to the morphological physical joint feature map. The classification method through multi-physical parameter fusion and anatomical constraint improves the accuracy, robustness and interpretability of prostate cancer image recognition classification.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for constructing a deep learning-based prostate cancer image recognition and classification model. Background Technology

[0002] Existing prostate cancer image recognition and classification methods mainly rely on single-modal analysis. The analysis results between different modalities are often independent, lacking an effective multi-parameter physical feature fusion mechanism. This results in an inability to comprehensively reflect the overall physical characteristics of prostate tissue, reducing classification accuracy. Existing segmentation methods typically use thresholding or region growing algorithms for "hard classification" when dealing with prostate region boundaries. This means that each pixel is assigned to a certain region in an either-or manner, ignoring the natural transition characteristics between prostate regions. This leads to unnatural jagged or stepped boundary segmentation results that do not match the actual physiological state. Existing technologies severely underutilize anatomical prior knowledge in prostate cancer image analysis. Most methods rely solely on surface features such as image grayscale and texture for classification, failing to organically integrate anatomical knowledge such as the positional relationships, morphological features, and physical characteristics of prostate regions into the classification process. This often results in classification results that violate basic physiological laws.

[0003] In summary, existing technologies suffer from problems such as insufficient multimodal data fusion, unnatural boundary classification, and inadequate utilization of anatomical prior knowledge, which urgently need to be addressed. Summary of the Invention

[0004] Therefore, it is necessary to provide a method for constructing a deep learning-based prostate cancer image recognition and classification model to solve at least one of the aforementioned technical problems.

[0005] To achieve the above objectives, a method for constructing a deep learning-based prostate cancer image recognition and classification model includes the following steps:

[0006] Step S1: Obtain a standardized multimodal image set; construct an inter-regional characteristic ratio map based on the standardized multimodal image set; construct a regional physical feature tensor based on the inter-regional characteristic ratio map;

[0007] Step S2: Divide the standardized multimodal image set into three regions: peripheral zone, transition zone, and central zone. Introduce anatomical prior knowledge as constraints. Construct a nonlinear mapping function based on the regional physical feature tensor to obtain a morphophysical mapping function set. Generate a morphophysical joint feature map based on the regional physical feature tensor and the morphophysical mapping function set.

[0008] Step S3: Construct a progressive constraint set of physical properties based on the morphological-physical joint feature map; perform peripheral zone priority segmentation on the morphological-physical joint feature map based on the progressive constraint set of physical properties to obtain a peripheral zone segmentation mask map; perform transition zone conditional segmentation based on the peripheral zone segmentation mask map to obtain a transition zone segmentation mask map; determine the central zone boundary and map the physical property distribution based on the transition zone segmentation mask map to obtain a regional physical property distribution map.

[0009] Step S4: Perform deep learning feature optimization on the regional physical characteristic distribution map to obtain a deep feature enhancement map; perform micro-medium-macro progressive verification and scoring on the deep feature enhancement map to obtain the regional classification credibility index.

[0010] This invention effectively integrates the physical characteristics of multimodal images, including tissue stiffness, water diffusion, and blood perfusion, by constructing an inter-regional characteristic ratio map and a regional physical feature tensor. It also captures the interactions between regions, providing a more comprehensive and discriminative feature representation for subsequent feature mapping and region segmentation, thereby improving the accuracy of prostate region identification. By introducing anatomical prior knowledge as constraints and combining it with nonlinear mapping using the regional physical feature tensor, a morphophysical joint feature map is constructed. This effectively links morphological features and physical characteristics, solving the problem of "hard classification" at boundaries in traditional methods. This results in more natural transitions between region boundaries, better conforming to the physiological structure of the prostate, thus improving the accuracy and robustness of segmentation. By constructing a progressive constraint set of physical characteristics and adopting a hierarchical segmentation strategy—first segmenting the peripheral zone, then the transition zone, and finally determining the central zone—the invention fully utilizes the differences and progressive relationships of the physical characteristics of different regions, avoiding mutual interference between regions, improving the accuracy and efficiency of segmentation, and ensuring the physiological rationality of the segmentation results. By optimizing features through deep learning, the consistency of regional boundaries and physical properties in space is enhanced. Combined with multi-scale verification and scoring mechanisms at the micro, meso, and macro levels, the segmentation results are comprehensively evaluated and optimized, effectively improving the reliability and credibility of the classification results. The resulting regional classification credibility index more accurately reflects the quality of the classification results. Therefore, this invention provides a method for constructing a prostate cancer image recognition and classification model based on deep learning. Through a technical route of "physical feature fusion - constraint mapping - hierarchical segmentation - multi-scale verification," a Regional Physical Feature Tensor (RPT) is constructed to organically integrate the three-dimensional feature spaces of tissue stiffness, water diffusion, and blood perfusion. A boundary gradient penalty term is introduced to ensure a natural transition of physical properties at morphological boundaries. Anatomical prior knowledge is encoded as constraints to guide the classification process, ensuring that the classification results conform to physiological laws. This classification method based on multi-physical parameter fusion and anatomical constraints effectively solves the accuracy and reliability problems in prostate cancer image recognition and classification, improving the accuracy, robustness, and interpretability of prostate cancer image recognition and classification. Attached Figure Description

[0011] Figure 1 A flowchart illustrating the steps involved in constructing a deep learning-based prostate cancer image recognition and classification model.

[0012] Figure 2 This is a detailed flowchart illustrating the implementation steps of step S2 in this invention.

[0013] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0015] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0016] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0017] In this embodiment of the invention, reference is made to Figure 1 The diagram shown illustrates the steps of constructing a deep learning-based prostate cancer image recognition and classification model according to the present invention. In this example, the method for constructing the deep learning-based prostate cancer image recognition and classification model includes the following steps:

[0018] Step S1: Obtain a standardized multimodal image set; construct an inter-regional characteristic ratio map based on the standardized multimodal image set; construct a regional physical feature tensor based on the inter-regional characteristic ratio map;

[0019] In this embodiment of the invention, three modalities of prostate imaging data—elastic imaging, DWI, and DCE-MRI—are acquired using MRI equipment and preprocessed (noise filtering, motion artifact correction, temporal alignment, and spatial registration) to generate a standardized multimodal image set. Then, physical characteristic parameters (tissue stiffness, ADC value, and blood perfusion parameters) for each voxel are extracted from the standardized image set to construct a regional physical parameter matrix. Next, the ratios of physical parameters (stiffness ratio, ADC ratio, and blood perfusion parameter ratio) between adjacent regions (peripheral zone, transition zone, and central zone) are calculated, and gradient weighting and anisotropic diffusion filtering are performed to construct an inter-regional characteristic ratio map. Finally, the regional physical parameter matrix and the inter-regional characteristic ratio map are fused and normalized using min-max normalization to construct a regional physical feature tensor (RPT). The RPT contains the original physical parameters of each voxel and information on the relative relationships between regions.

[0020] Step S2: Divide the standardized multimodal image set into three regions: peripheral zone, transition zone, and central zone. Introduce anatomical prior knowledge as constraints. Construct a nonlinear mapping function based on the regional physical feature tensor to obtain a morphophysical mapping function set. Generate a morphophysical joint feature map based on the regional physical feature tensor and the morphophysical mapping function set.

[0021] In this embodiment of the invention, based on prostate anatomy knowledge (or atlas registration), the standardized multimodal image set is initially divided into three regions: the peripheral zone, the transition zone, and the central zone. Then, morphological features (shape factor, boundary curvature, relative position) of each region are extracted from the Regional Physical Feature Tensor (RPT) and quantified into a morphological feature description set. Simultaneously, anatomical prior knowledge (regional shape range, relative positional relationships, boundary curvature patterns) is encoded into constraint rules, forming an anatomical prior constraint library. Next, using a Support Vector Machine (SVM) and a Radial Basis Function (RBF) kernel, a nonlinear mapping function is constructed. The physical features from the RPT, the morphological feature description set, and the constraint rules from the anatomical prior constraint library (introduced through penalty terms) are integrated into the SVM training to obtain a set of morphophysical mapping functions. Finally, the physical features of each voxel in the RPT and the morphological features of its region are input into the morphophysical mapping function set to calculate the probability of it belonging to each region and fuse it with the original physical features to generate a morphophysical joint feature map (MP). The MP contains the original physical features, morphological features and anatomical prior knowledge.

[0022] Step S3: Construct a progressive constraint set of physical properties based on the morphological-physical joint feature map; perform peripheral zone priority segmentation on the morphological-physical joint feature map based on the progressive constraint set of physical properties to obtain a peripheral zone segmentation mask map; perform transition zone conditional segmentation based on the peripheral zone segmentation mask map to obtain a transition zone segmentation mask map; determine the central zone boundary and map the physical property distribution based on the transition zone segmentation mask map to obtain a regional physical property distribution map.

[0023] In this embodiment of the invention, statistical analysis is performed on the physical property parameters (tissue stiffness, ADC value, blood perfusion parameters) of each region (based on preliminary segmentation) in the morphophysical joint feature map (MP) to construct a three-dimensional physical property parameter space. The difference matrix between regions is calculated to determine the region segmentation priority. Then, based on the region parameter space model and segmentation priority, progressive constraint rules (parameter ranges for each region) are designed to obtain a progressive constraint set of physical properties. Next, the peripheral zone parameter range in the progressive constraint set of physical properties is used to enhance the MP, obtaining a peripheral zone feature response map. An adaptive threshold is determined using the Otsu method for preliminary segmentation and candidate region extraction. Then, physical parameter constraint screening and morphological optimization (opening, closing, Gaussian filtering) are performed to obtain a peripheral zone segmentation mask map (PZM). Finally, the PZM is excluded from the MP to obtain the remaining region map, and a transition zone feature scoring map is constructed. Seed points are selected, and a conditional constraint region growing algorithm is used to obtain a transition zone segmentation mask map (TZM). Finally, the central zone boundary was determined by excluding PZM and TZM regions, and morphological optimization was performed to obtain the Region Segmentation Label Map (RSL). The RSL was fused with the MP to obtain the Region Physical Property Distribution Map (RPD), which contains the segmentation results of each region and the corresponding distribution of physical property parameters.

[0024] Step S4: Perform deep learning feature optimization on the regional physical characteristic distribution map to obtain a deep feature enhancement map; perform micro-medium-macro progressive verification and scoring on the deep feature enhancement map to obtain the regional classification credibility index.

[0025] In this embodiment of the invention, a lightweight convolutional neural network (CNN) (encoder-decoder structure, including an adaptive domain attention module and skip connections) is constructed using a Regional Physical Characteristics Distribution Map (RPD) and a Standardized Multimodal Image Set (SMI) as input. The network is then trained under physical constraints (the loss function includes region segmentation loss, boundary loss, and physical consistency loss) to obtain a physically constrained optimized model. This model is used for boundary-region co-enhancement, and combined with physiological validation and fine-tuning, to obtain a Deep Feature Enhancement Map (DFE). Next, microscale feature analysis (pixel-level feature consistency evaluation) is performed on the DFE to obtain a Microscale Feature Consistency Map (MFC). Then, mesoscale structural evaluation (sub-region feature analysis and scoring) is performed to obtain a mesoscale structural evaluation matrix (MSE). Finally, macroscale overall consistency validation (boundary sharpness, smoothness, and anatomical conformity evaluation) is performed on the MFC, MSE, and MGI to obtain a macroscale consistency index set (MGI). Finally, feature space standardization, information entropy weight calculation, and fusion of deep features and multiscale features are performed on the MFC, MSE, and MGI to construct a consistency scoring function and generate and optimize a multiscale consistency scoring map (MSC). The Region Classification Confidence Index (RCI) for each voxel is calculated based on MSC and DFE, and the RCI map is obtained.

[0026] Preferably, step S1 includes the following steps:

[0027] Step S11: Acquire and preprocess prostate multimodal medical images to obtain a standardized multimodal image set, which includes elastography, DWI sequence and DCE-MRI sequence;

[0028] Step S12: Calculate the regional feature parameters of the standardized multimodal image set to obtain the regional physical parameter matrix;

[0029] Step S13: Construct a region characteristic ratio map based on the region physical parameter matrix;

[0030] Step S14: Perform multi-parameter feature fusion and tensor construction on the regional physical parameter matrix and the inter-regional characteristic ratio map to obtain the regional physical feature tensor.

[0031] In this embodiment of the invention, multimodal imaging data of the patient's prostate are acquired using a hospital's magnetic resonance imaging (MRI) equipment. Specifically, a 3.0 Tesla (T) MRI scanner equipped with a body phased array coil is used. Three modalities of images are acquired: (1) Elastography: Shear wave elastography (SWE) technology is used, the mechanical excitation frequency is set to 50Hz, 5 repeated scans are acquired, and the scan time is 10 seconds each, to acquire elastography data. (2) DWI sequence: Single excitation spin echo planar imaging (SE-EPI) sequence is used, with b values ​​of 0, 500, 1000, and 1500 s / mm², repetition count (TR) of 4000ms, echo time (TE) of 70ms, slice thickness of 3mm, field of view (FOV) of 200mm×200mm, and matrix size of 128×128, to acquire DWI sequence data. (3) DCE-MRI sequence: A three-dimensional phase-scrambled gradient echo (3D-SPGR) sequence was used with gadopentetate dimeglumine (Gd-DTPA) contrast agent at a dose of 0.1 mmol / kg body weight and an injection rate of 2 ml / s. A set of baseline images was acquired before contrast agent injection, followed by continuous dynamic scanning for a total of 60 phases. Each phase lasted 5 seconds, with a TR of 4 ms, a TE of 2 ms, a flip angle of 12°, a slice thickness of 3 mm, a FOV of 240 mm × 240 mm, and a matrix size of 192 × 192. DCE-MRI sequence data were then acquired. The acquired raw image data were then preprocessed. For elastography data, a 3 × 3 median filter was used to remove salt-and-pepper noise from the images. For DWI sequences, a rigid registration algorithm based on mutual information was used to register images with different b values ​​onto images with b = 0 s / mm² to correct motion artifacts. For DCE-MRI sequences, the first phase image is selected as the reference image. Then, a non-rigid registration algorithm based on mutual information is used to register subsequent phase images with the reference image, correcting image distortion caused by respiratory motion. Finally, the image data of the three modalities are spatially registered, and a cubic B-spline interpolation algorithm is used to resample the elastography and DCE-MRI images to the same resolution and slice thickness as the DWI images, ensuring that all voxels correspond one-to-one, generating a standardized multimodal image set. The region physical parameter matrix is ​​a three-dimensional matrix, represented as RPM[x, y, z], where x, y, and z represent the coordinates of the voxels in the three spatial dimensions, respectively.

[0032] Based on the standardized multimodal image set obtained in step S11, the physical characteristic parameters of each voxel are calculated. For elastography data, the shear wave velocity (SWS) value of each voxel is directly read from the SWE image, and then Young's modulus E is calculated as the tissue stiffness index based on the shear wave velocity value. For DWI sequences, the signal attenuation curve of each voxel is fitted using a double exponential model, and the apparent diffusion coefficient (ADC) value of each voxel is obtained by fitting the model using the Levenberg-Marquardt nonlinear least squares algorithm. For DCE-MRI sequences, the time-signal intensity curve of each voxel is fitted using an extended Tofts model. The AIF is obtained by automatically selecting arteries in the reference area, and then the extended Tofts model is fitted using the Levenberg-Marquardt algorithm to obtain the Ktrans, Kep, and Ve (extracellular space volume fraction) values ​​of each voxel. The obtained tissue stiffness index, ADC value, Ktrans, Kep and Ve values ​​are stored in the corresponding positions of the corresponding region physical parameter matrix RPM, RPM[x, y, z, p], where p=1, 2, 3, 4, 5 correspond to these five parameters respectively.

[0033] Based on the Regional Physical Parameter Matrix (RPM) obtained in step S12, the physical property ratios between adjacent regions are calculated. First, three main regions within the prostate are defined: the peripheral zone (PZ), the transition zone (TZ), and the central zone (CZ). Initial outlines of these three regions are manually or automatically delineated based on prior anatomical knowledge. Then, the physical property ratios between each region and its neighboring regions are calculated. For example, for the boundary region between the peripheral zone (PZ) and the transition zone (TZ), the stiffness ratio, ADC ratio, and Ktrans ratio of each PZ voxel are calculated between it and its nearest neighbor TZ voxel. To avoid division by zero errors, a small constant is added to the denominator when calculating the ratios. Similar ratio calculations are performed for other region boundaries (PZ-CZ, TZ-CZ). The calculated ratios are stored in the corresponding inter-regional property ratio map (IRR).

[0034] Based on the regional physical parameter matrix RPM obtained in step S12 and the inter-regional characteristic ratio map IRR obtained in step S13, they are fused to construct the regional physical feature tensor RPT. RPT is a four-dimensional tensor, RPT[x, y, z, c], where x, y, and z represent the spatial coordinates of voxels, and c represents the feature channel. The five parameters in RPM (tissue stiffness index, ADC, Ktrans, Kep, Ve) are used as the first five channels of RPT, i.e., RPT[x, y, z, 1:5] = RPM[x, y, z, 1:5]. The three ratios in IRR (stiffness ratio, ADC ratio, Ktrans ratio) are used as the last three channels of RPT, i.e., RPT[x, y, z, 6:8] = IRR[x, y, z, 1:3]. To balance the contributions of different features, the feature values ​​of each channel are normalized. The min-max normalization method is used to scale the feature values ​​of each channel to the interval [0, 1]. Specifically, for each channel c, the maximum value max(c) and minimum value min(c) of all voxels in that channel are calculated. Then, the eigenvalues ​​of each voxel are transformed as follows: RPT[x, y, z, c] = (RPT[x, y, z, c] - min(c)) / (max(c) - min(c)). In this way, the resulting RPT tensor contains 8 feature channels, with each channel's eigenvalues ​​falling within the range [0, 1], while preserving the spatial structure and inter-regional relationships of the original data.

[0035] Preferably, step S13 includes the following steps:

[0036] Step S131: Define the boundary transition zone based on the regional physical parameter matrix to obtain the boundary transition zone location map;

[0037] Step S132: Calculate the multi-parameter directional gradient based on the boundary transition zone location map and the regional physical parameter matrix to obtain the directional parameter gradient set;

[0038] Step S133: Calculate the local weighted ratio based on the regional physical parameter matrix and the gradient set of directional parameters to obtain the local parameter ratio set;

[0039] Step S134: Perform regional feature enhancement processing on the local ratio set of parameters to obtain an enhanced ratio map set;

[0040] Step S135: Perform multi-parameter ratio fusion on the enhanced ratio map to obtain the inter-regional characteristic ratio map.

[0041] In this embodiment of the invention, based on the Regional Physical Parameter Matrix (RPM) obtained in step S12, it is first necessary to preliminarily determine the boundary positions of different regions within the prostate. Using the ADC value in the RPM as the primary basis, the boundaries are initially identified by calculating its three-dimensional spatial gradient magnitude map. A gradient threshold is set, and voxels with gradient magnitudes greater than the preset threshold are marked as boundary candidate voxels. Then, morphological dilation is performed on the boundary candidate voxels to form a transition zone containing the boundary region and its adjacent tissues, generating a boundary transition zone location map BT.

[0042] Based on the boundary transition zone location map BT obtained in step S131 and the region physical parameter matrix RPM obtained in step S12, the directional gradients of multiple physical parameters for each voxel within the boundary transition zone are calculated. For each voxel within the boundary transition zone, the gradients of its three key parameters (tissue stiffness index, ADC value, and Ktrans value) in the RPM are calculated in three orthogonal directions (x, y, z). These gradient values ​​are combined into a directional parameter gradient set DPG.

[0043] Based on the region physical parameter matrix RPM obtained in step S12 and the directional parameter gradient set DPG obtained in step S132, the local weighted ratio of each voxel within the boundary transition zone is calculated. For each voxel within the boundary transition zone, the principal gradient direction of that voxel is determined, and then the parameter ratio of that voxel to its neighboring voxels along the principal direction is calculated. These ratios are then weighted and averaged according to the gradient magnitude to obtain the local weighted ratio. A similar operation is performed for the ADC value and the Ktrans value. These three weighted ratios are combined into a local parameter ratio set PLR.

[0044] Based on the local ratio set PLR obtained in step S133, an anisotropic diffusion filtering method is used for region feature enhancement. For each weighted ratio in PLR, a diffusion tensor is constructed based on its gradient information, and then anisotropic diffusion filtering is performed iteratively. After anisotropic diffusion filtering, the ratio difference at the region boundary is enhanced, while the ratio fluctuation within the region is suppressed. The three filtered weighted ratios are combined to form the enhanced ratio map set ERV.

[0045] Based on the enhancement ratio atlas ERV obtained in step S134, the multiple enhancement ratios (tissue stiffness ratio, ADC ratio, and Ktrans ratio) contained therein are fused into a single inter-regional characteristic ratio atlas. The specific operation is as follows: A weighted average method is used for fusion. For each voxel in the ERV, the weighted average of its three enhancement ratios is calculated. The weights are determined based on the ability of each parameter to distinguish different prostate regions. Specifically, the coefficient of variation (CV) of each parameter between different regions is first calculated. For each parameter p (p=1, 2, 3 corresponding to tissue stiffness, ADC, and Ktrans respectively), its mean μp, PZ, μp, TZ, μp, CZ and standard deviation σp, PZ, σp, TZ, σp, CZ are calculated in the three regions PZ, TZ, and CZ. Then, calculate the coefficient of variation for this parameter in each region: CVp,PZ=σp,PZ / μp,PZ,CVp,TZ=σp,TZ / μp,TZ,CVp,CZ=σp,CZ / μp,CZ. Then, calculate the average coefficient of variation for this parameter across the three regions: CVp=(CVp,PZ+CVp,TZ+CVp,CZ) / 3. A larger coefficient of variation indicates a stronger ability of the parameter to distinguish different regions. Therefore, the normalized value of the coefficient of variation is used as the weight. Normalize the coefficients of variation for the three parameters: Wp=CVp / (CV1+CV2+CV3). Then, for each voxel in the ERV, calculate the weighted average of its three enhancement ratios: IRR[x,y,z]=W1×ERV[x,y,z,1]+W2×ERV[x,y,z,2]+W3×ERV[x,y,z,3]. Here, IRR[x, y, z] represents the value of the corresponding voxel in the final inter-region characteristic ratio map. In this way, multiple enhancement ratios are merged into a single ratio map, which comprehensively reflects the differences of different physical parameters at the region boundaries, providing more discriminative features for subsequent feature mapping and region segmentation.

[0046] As an example of the present invention, reference is made to... Figure 2 As shown, step S2 in this example includes:

[0047] Step S21: Based on anatomical knowledge, the standardized multimodal image set is divided into three regions: peripheral zone, transition zone, and central zone. Morphological features are extracted and quantified based on the regional physical feature tensor to obtain a morphological feature description set.

[0048] In this embodiment of the invention, based on anatomical knowledge of the prostate, a preliminary regional division is performed on the standardized multimodal image set. After obtaining the preliminary regional division, morphological features of each region are extracted based on the Regional Physical Feature Tensor (RPT) obtained in step S14. Specifically, for each region (PZ, TZ, CZ), its shape factor, boundary curvature, and relative position are calculated. These morphological features (shape factor, mean curvature, standard deviation of curvature, centroid distance, centroid direction) are combined into a morphological feature description set (MFD).

[0049] Step S22: Encode the anatomical prior knowledge into the morphological feature description set to obtain the anatomical prior constraint library;

[0050] In this embodiment of the invention, the morphological feature description set (MFD) obtained in step S21 is encoded into anatomical prior knowledge to form a constraint library. The specific operation is as follows: Based on prostate anatomy knowledge, the normal range of morphological features for each region is determined. For example, the peripheral zone is usually located posterolaterally to the prostate, with a relatively flat shape and a large shape factor; the transition zone is located in the middle of the prostate, surrounding the urethra, with a relatively rounded shape and a small shape factor; the central zone is located at the base of the prostate and is cone-shaped. Based on this knowledge, a reasonable range is set for each morphological feature in each region. For example, the shape factor range for PZ is set to [0.6, 0.8], the shape factor range for TZ is set to [0.4, 0.6], and the shape factor range for CZ is set to [0.5, 0.7]. The relative positional relationship between regions is set. For example, PZ should be located lateral to TZ and CZ, TZ should be located between PZ and CZ, and CZ should be located above PZ and TZ. The relative positional relationship between regions can be represented by vectors. For example, the direction of the vector (PZ centroid - TZ centroid) should generally point posterolaterally. Define typical curvature variations in the region boundaries. For example, the boundary between PZ and TZ is usually smoother with less curvature variation; the boundary between TZ and CZ is more curved with greater curvature variation. This anatomical prior knowledge can be encoded into a constraint library APK. APK can be a collection of multiple rules. Each rule takes the form: IF(condition)THEN(constraint). For example: IF(region=PZ)THEN(0.6≤shape factor≤0.8). IF(region=TZ)THEN(direction of vector (PZ centroid - TZ centroid) ∈ [135°, 225°]). APK contains multiple such rules, covering prior knowledge of the morphological characteristics and inter-regional relationships of various regions of the prostate.

[0051] Step S23: Construct a nonlinear mapping function based on the region physical feature tensor, morphological feature description set, and anatomical prior constraint library to obtain the morphophysical mapping function set;

[0052] In this embodiment of the invention, a nonlinear mapping function is constructed based on the region physical feature tensor (RPT) obtained in step S14, the morphological feature description set (MFD) obtained in step S21, and the anatomical prior constraint library (APK) obtained in step S22. Support Vector Machine (SVM) is used as the basis for the nonlinear mapping function. First, a training dataset is constructed. Then, an SVM with radial basis function (RBF) kernels is used to train the training dataset. The optimal SVM parameters are determined through cross-validation. During training, rules from the anatomical prior constraint library (APK) are introduced as constraints; specifically, a penalty term is added to the loss function for training samples that violate the rules. In this way, anatomical prior knowledge is integrated into the SVM training process. After training, a nonlinear mapping function set (MPM) is obtained.

[0053] Step S24: Generate a morphophysical joint feature map based on the region physical feature tensor and the morphophysical mapping function set.

[0054] In this embodiment of the invention, a morphophysical joint feature map MP is generated based on the Region Physical Feature Tensor (RPT) obtained in step S14 and the Morphophysical Mapping Function Set (MPM) obtained in step S23. The specific operation is as follows: For each voxel in the RPT, the values ​​of its eight physical feature channels and the six morphological features of the region to which the voxel belongs, calculated in step S21 (if the voxel does not belong to any defined region, the morphological features are set to 0), are input into the morphophysical mapping function set (MPM) obtained in step S23. Each SVM model in the MPM outputs a probability value representing the probability that the voxel belongs to the corresponding region (PZ, TZ, CZ). These three probability values ​​are used as three new channels of MP. That is, MP[x, y, z, 1] represents the probability that the voxel belongs to PZ, MP[x, y, z, 2] represents the probability that the voxel belongs to TZ, and MP[x, y, z, 3] represents the probability that the voxel belongs to CZ. These three probability values ​​are then fused with the eight physical feature channels in the RPT using a weighted average method. The weights can be determined based on the discriminative power of each feature or set empirically. For example, the weight of the probability value can be set to 0.6, and the weight of the physical feature to 0.4. Then, each channel is normalized to scale the feature values ​​to the [0, 1] interval. The resulting MP is an 11-channel feature map (8 original physical features + 3 region probabilities). The MP not only contains the original physical feature information but also integrates morphological features and anatomical prior knowledge, providing a more discriminative and robust feature representation for subsequent region segmentation.

[0055] Preferably, step S23 includes the following steps:

[0056] Step S231: Extract key physical property parameters from the region physical feature tensor, including tissue stiffness, water diffusion, and blood perfusion; extract morphological features from the morphological feature description set, including boundary curvature, shape factor, and relative position.

[0057] Step S232: Perform feature correlation on key physical property parameters and morphological features to obtain the feature correlation matrix;

[0058] Step S233: Extract the boundary curvature distribution from the morphological feature description set, construct the curvature-related weight function based on the feature correlation matrix, and obtain the curvature weight function set;

[0059] Step S234: Extract anatomical constraint rules for the prostate region from the anatomical prior constraint library, construct anatomical constraint penalty terms based on the curvature weight function set, and obtain the anatomical penalty function;

[0060] Step S235: Design the boundary gradient penalty term based on the anatomical penalty function and the morphological feature description set to obtain the boundary gradient penalty function;

[0061] Step S236: Integrate nonlinear mapping functions based on the feature correlation matrix, curvature weight function set, anatomical penalty function, and boundary gradient penalty function to obtain the morphophysical mapping function set.

[0062] In this embodiment of the invention, from the Regional Physical Feature Tensor (RPT) obtained in step S14, tissue stiffness (RPT[x, y, z, 1]), ADC value (RPT[x, y, z, 2]), and Ktrans value (RPT[x, y, z, 3]) are selected as key physical characteristic parameters. These three parameters represent different physical characteristics of prostate tissue and have good complementarity in distinguishing different regions. From the Morphological Feature Description Set (MFD) obtained in step S21, the average value of boundary curvature (MFD[r, 2]), shape factor (MFD[r, 1]), and distance from the region centroid to the prostate centroid (MFD[r, 4]) are selected as key morphological features. These three features describe the morphological characteristics of the prostate region from different perspectives.

[0063] Based on the key physical property parameters and morphological features extracted in step S231, the correlation between them is calculated to construct a feature correlation matrix. First, the key physical property parameters and morphological features are combined into a feature vector. Then, the Pearson correlation coefficients between each pair of these six features are calculated, resulting in a 6×6 symmetric matrix, which is the feature correlation matrix CM. Specifically, for each pair of features (i,j), the Pearson correlation coefficient is calculated as follows: First, N sample points are randomly selected from all voxels (usually N>1000 to ensure statistical reliability), and the feature values ​​of each sample point are recorded to form feature vectors Xi and Xj. Then, the mean μi and μj of these two feature vectors, as well as the standard deviations σi and σj, are calculated. The Pearson correlation coefficient rij is calculated using the formula rij=Σ[(Xi-μi)(Xj-μj)] / (N·σi·σj), where the summation is performed on all N sample points. The correlation coefficient ranges from -1 to 1. A value closer to 1 indicates a stronger positive correlation, a value closer to -1 indicates a stronger negative correlation, and a value close to 0 indicates that the two features are almost uncorrelated. To improve computational efficiency, a block-based computation strategy can be adopted. The prostate volume is divided into multiple sub-blocks, and the correlation coefficient is calculated separately within each sub-block. Then, a weighted average is taken, with the weight proportional to the number of sample points within the sub-block. This block-based strategy not only parallelizes the computation process but also captures local variations in feature correlation across different regions.

[0064] Based on the morphological feature description set MFD obtained in step S21 and the feature correlation matrix CM obtained in step S232, a curvature-related weight function is constructed to adjust the feature weights of different curvature regions. The specific operations are as follows: Extract the boundary curvature distribution information of each region from the MFD. For each region, calculate the average curvature and standard deviation of its boundary pixels. Then, based on the feature correlation matrix CM, determine the correlation between boundary curvature and other features. Extract the rows (or columns) corresponding to the boundary curvature from the CM to obtain a vector containing 5 elements, representing the correlation coefficients between boundary curvature and the other 5 features (tissue stiffness, ADC value, Ktrans value, shape factor, and region centroid distance). Use the absolute values ​​of these correlation coefficients as weights. The larger the weight, the stronger the correlation between the feature and the boundary curvature, and the more emphasis should be placed on it in the subsequent mapping function construction. Construct a curvature-related weight function CWF. CWF is a function whose input is the boundary curvature value and whose output is a weight vector. Each element of the weight vector corresponds to the weight of a feature. The specific form of CWF can be designed according to the actual situation. For example, a piecewise linear function can be designed: if the curvature value is less than the region's mean curvature minus the standard deviation, the weight vector is [w1, w2, w3, w4, w5]; if the curvature value is between the region's mean curvature minus the standard deviation and the region's mean curvature plus the standard deviation, the weight vector is linearly interpolated based on the curvature value; if the curvature value is greater than the region's mean curvature plus the standard deviation, the weight vector is [w1', w2', w3', w4', w5']. Here, w1-w5 and w1'-w5' are calculated based on the correlation coefficients between the boundary curvature and other features in the CM. The CWFs of each region are combined into a curvature weight function set CWFS.

[0065] Based on the anatomical prior constraint library APK obtained in step S22 and the curvature weight function set CWFS obtained in step S233, an anatomical penalty function is constructed to penalize samples that violate anatomical rules during the mapping function construction process. Specifically, constraint rules related to the morphology and location of the prostate region are extracted from the APK. Examples include "the shape factor of the peripheral zone should be greater than 0.6" and "the transition zone should be located between the peripheral and central zones." For each constraint rule, a corresponding penalty term is designed. The magnitude of the penalty term is proportional to the degree of rule violation. For example, for the rule "the shape factor of the peripheral zone should be greater than 0.6," a penalty term can be designed: Penalty = max(0, 0.6 - ShapeFactor_PZ). If the shape factor of the peripheral zone is less than 0.6, the penalty term is positive; otherwise, it is 0. For rules related to region location, penalty terms can be designed based on the relative position and orientation of the region's centroid. For example, if the centroid of the transition zone is located in front of the centroid of the peripheral zone, an additional penalty term is added. The penalty terms of all constraint rules are weighted and summed to obtain a total anatomical penalty function (APF). The APF takes the feature vector and region label of the current sample as input and outputs a scalar value representing the degree to which the sample violates the anatomical rules. The weights can be determined based on the importance of each rule. The APF is adjusted using a curvature weight function set (CWFS). Regions with greater boundary curvature are penalized more severely for violating anatomical rules, as these regions are often transitional areas between regions and are more prone to misclassification.

[0066] Based on the anatomical penalty function APF obtained in step S234 and the morphological feature description set MFD obtained in step S21, a boundary gradient penalty function is designed to constrain the gradient direction of the boundary region during the mapping function construction process, ensuring a smooth transition of physical properties at the boundary. The specific operation is as follows: For each boundary pixel of a region, calculate the rate of change of its physical property parameters (tissue stiffness, ADC value, Ktrans value) in the gradient direction. The gradient direction can be calculated using the method in step S132. Calculate the difference between the rate of change and the typical rate of change for that region. For example, for the boundary pixel between the peripheral zone and the transition zone, calculate the rate of change of its tissue stiffness in the gradient direction and compare it with the typical difference in tissue stiffness between the peripheral zone and the transition zone. If the rate of change is too large or too small, add a penalty term. The magnitude of the penalty term is proportional to the absolute value of the difference. Adjust the boundary gradient penalty term using the anatomical penalty function APF. For region boundaries that violate anatomical rules, increase the gradient penalty strength. For example, if a boundary pixel is classified as a peripheral zone but its shape factor is less than 0.6, increase its gradient penalty term. The penalty terms for all boundary pixels are weighted and summed to obtain a total boundary gradient penalty function (BGPF). The input of BGPF is the feature vector and region label of the current sample, and the output is a scalar value indicating whether the gradient change of the sample in the boundary region meets the expectation.

[0067] Based on the feature correlation matrix CM obtained in step S232, the curvature weight function set CWFS obtained in step S233, the anatomical penalty function APF obtained in step S234, and the boundary gradient penalty function BGPF obtained in step S235, they are integrated into a nonlinear mapping function. The specific operation is as follows (continued): Support Vector Machine (SVM) is used as the basis for the nonlinear mapping function. An SVM with a radial basis function (RBF) kernel is used. During the training process of the SVM, the feature correlation matrix CM is introduced. The kernel function of the SVM is adjusted according to the correlation between different features in CM. For features with strong correlation, their weight in the kernel function calculation is increased. For example, if the correlation coefficient between feature i and feature j is CM[i, j], then when calculating the kernel function K(x, x'), ||x-x'|| is increased. The algorithm is modified to Σ(CM[i,j]×(xi-x'i)×(xj-x'j)). A curvature weight function set (CWFS) is introduced. For each training sample, a corresponding weight vector is selected from the CWFS based on its region and boundary curvature value. This weight vector is applied to the sample's feature vector, weighting different features. An anatomical penalty function (APF) is introduced. A penalty term is added to the SVM's objective function, with its magnitude proportional to APF(x). Here, x is the current sample's feature vector and region label. Thus, if a sample's features and label violate anatomical rules, the SVM training process will tend to adjust the model parameters to reduce this violation. A boundary gradient penalty function (BGPF) is introduced. Another penalty term is added to the SVM's objective function, with its magnitude proportional to BGPF(x). Thus, if a sample's gradient change in the boundary region does not meet expectations, the SVM training process will tend to adjust the model parameters to reduce this discrepancy. Through the above method, feature association, curvature weights, anatomical constraints, and boundary gradient constraints are all integrated into the SVM training process. After training, a nonlinear mapping function set MPM is obtained. Similar to the MPM in step S23, the MPM contains multiple SVM models, each corresponding to the classification of a region. However, due to the introduction of more constraints and weights, the MPM can better map physical and morphological features to a unified feature space and ensure that the classification results conform to anatomical rules and the naturalness of boundary transitions. For a new voxel, its physical and morphological features (adjusted for curvature weights) are input into the MPM. Each SVM model outputs a probability value, representing the probability that the voxel belongs to the corresponding region. The region with the highest probability value is selected as the final classification result for that voxel. Preferably, the construction of the progressive constraint set of physical characteristics in step S3 includes:

[0068] Statistical analysis of regional physical property parameters was performed on the morphophysical joint feature map to obtain a set of regional parameter statistical features;

[0069] Using tissue stiffness, ADC value, and blood perfusion as three coordinate axes, a three-dimensional physical property parameter space was constructed to obtain a regional parameter space model;

[0070] Calculate the regional difference matrix of the regional parameter spatial model;

[0071] Determine the region segmentation priority table based on the region difference matrix;

[0072] Based on the regional parameter space model and the regional segmentation priority table, progressive constraint rules are designed to obtain a progressive constraint set of physical characteristics.

[0073] In this embodiment of the invention, statistical analysis of the physical characteristic parameters in the MP is required. Note that precise region segmentation has not yet been performed at this stage, so the "region" here refers to the region initially divided in step S21 (the preliminary division based on anatomical knowledge may not be very precise). For each voxel in the MP, three key physical characteristic parameters are extracted: tissue stiffness (corresponding to the first channel of RPT), ADC value (corresponding to the second channel of RPT), and Ktrans value (representing blood perfusion, corresponding to the third channel of RPT). For each region (PZ, TZ, CZ), the mean, standard deviation, median, minimum, and maximum values ​​of these three parameters are calculated respectively. These statistics are combined into a region parameter statistical feature set RPSS. RPSS is a matrix, RPSS[r, s], where r=1, 2, 3 correspond to the three regions PZ, TZ, and CZ respectively, and s=1, 2, ..., 15 correspond to the five statistics (mean, standard deviation, median, minimum, and maximum) of the above three parameters respectively.

[0074] Based on the Regional Statistical Feature Set (RPSS), a three-dimensional physical property parameter space is constructed. Tissue stiffness, ADC value, and Ktrans value are used as the three coordinate axes. For each region (PZ, TZ, CZ), a region is defined in this three-dimensional space based on its corresponding statistic in RPSS. This region can be represented by an ellipsoid. The center of the ellipsoid is determined by the mean of the three parameters, and the lengths of the three axes are determined by the standard deviations of the three parameters (e.g., the axis length can be set to twice the standard deviation). This yields a Regional Parameter Space Model (APSM). The APSM contains three ellipsoids, representing the distribution range of PZ, TZ, and CZ in the physical property parameter space, respectively.

[0075] Based on the Region Parameter Space Model (APSM), the pairwise differences between the three regions are calculated. Specifically, the overlap volume between each pair of the three ellipsoids is calculated. The smaller the overlap volume, the higher the distinguishability between the two regions in the physical property parameter space. The Monte Carlo method can be used to estimate the overlap volume. For example, a large number of points (e.g., 10,000 points) are randomly generated in the smallest cuboid containing the three ellipsoids, and the number of points falling within each ellipsoid and the number of points falling within both ellipsoids are counted. The overlap volume can be estimated by dividing the number of points falling within both ellipsoids by the total number of points. A 3×3 symmetric matrix is ​​obtained, where each element of the matrix represents the overlap volume between the corresponding two regions. This matrix is ​​the Region Difference Matrix (RDM). RDM[i,j] represents the overlap volume between region i and region j. The diagonal elements of RDM are all 0 (the overlap volume between a region and itself is defined as 0).

[0076] Based on the Region Discrimination Matrix (RDM), the priority of region segmentation is determined. Regions with the highest discrimination are segmented first. Specifically, the minimum value (excluding diagonal elements) is found in the RDM. The two regions corresponding to the minimum value are the regions with the highest discrimination. For example, if RDM[1, 2] is the minimum, then PZ and TZ are the region pairs with the highest discrimination, and these two regions are segmented first. Then, these two rows and two columns are deleted from the RDM. The minimum value is found among the remaining elements to determine the next region to be segmented. This process is repeated until all regions are sorted. A region segmentation priority table RSPT is obtained. RSPT is a list containing region indices indicating the order in which regions are segmented. For example, RSPT=[1, 3, 2] means segmenting PZ first, then CZ, and finally TZ.

[0077] Based on the Region Parameter Spatial Model (APSM) and the Region Segmentation Priority Table (RSPT), a series of progressive constraint rules are designed to guide subsequent hierarchical region segmentation. Specifically, for the first region in the RSPT, the range of its physical property parameters is determined according to its corresponding ellipsoid in the APSM. For example, if the first region in the RSPT is PZ, the ranges of tissue stiffness, ADC value, and Ktrans value are determined based on the ellipsoid corresponding to PZ in the APSM. The boundary of the ellipsoid can be used as the boundary of the parameter range. For subsequent regions in the RSPT, in addition to considering their own ellipsoids, the parameter ranges of the segmented regions must also be considered. For example, if the second region in the RSPT is CZ, when determining the parameter range of CZ, not only the ellipsoid of CZ but also the parameter range of PZ must be excluded. Specifically, the parameter range of CZ can be obtained by intersecting the complements of the ellipsoids of CZ and PZ. These constraint rules are combined into a progressive constraint set of physical properties (PPCS). PPCS is a set containing multiple rules. Each rule takes the form: IF(region=Ri)THEN(parameter1∈range1, parameter2∈range2, parameter3∈range3). Here, Ri is the i-th region in RSPT, and range1, range2, and range3 are parameter ranges determined based on APSM and the segmented regions.

[0078] Preferably, the peripheral zone preferential segmentation in step S3 includes:

[0079] Based on the progressive constraint set of physical properties, the morphological-physical joint feature map is enhanced with peripheral zone features to obtain the peripheral zone feature response map.

[0080] Adaptive threshold determination is performed based on the peripheral zone characteristic response map and the progressive constraint set of physical properties to obtain an adaptive threshold distribution map.

[0081] The adaptive threshold distribution map is used to perform preliminary segmentation and candidate region extraction on the peripheral zone feature response map to obtain the peripheral zone candidate region map.

[0082] Based on the candidate region map of the outer periphery zone, physical parameter constraints are used for screening to obtain the outer periphery zone constraint screening map;

[0083] Morphological optimization and contour refinement are performed based on the peripheral zone constraint screening map to obtain the peripheral zone segmentation mask map.

[0084] In this embodiment of the invention, the range of physical property parameters of the peripheral zone (PZ) is determined according to the rules in the PPCS. For example, the PPCS has the following rule: IF(region=PZ)THEN(tissue stiffness∈[a1,b1], ADC value∈[a2,b2], Ktrans value∈[a3,b3]). Here, a1, b1, a2, b2, a3, and b3 are the parameter ranges of the PZ determined according to the region parameter space model APSM. Then, these parameter ranges are used to enhance the features of the MP. Specifically, for each voxel in the MP, the degree of matching between its three key physical property parameters (tissue stiffness, ADC value, and Ktrans value) and the PZ parameter range is calculated. A matching function can be defined as: Match(x,[a,b])=exp(-(x-(a+b) / 2)). / (2×((ba) / 2) ()). Here, x is the parameter value, and [a, b] is the parameter range. The function takes a maximum of 1 when the parameter value is equal to the center of the range, and gradually decreases as the parameter value deviates from the center of the range. For each voxel, the product of its three parameters' matching functions is calculated as the probability that the voxel belongs to PZ: P_PZ = Match(tissue stiffness, [a1, b1]) × Match(ADC value, [a2, b2]) × Match(Ktrans value, [a3, b3]). The P_PZ values ​​of all voxels are combined into a peripheral zone characteristic response map PZR. PZR[x, y, z] represents the probability that the voxel with coordinates (x, y, z) belongs to PZ.

[0085] Based on PZR, an adaptive threshold is determined. Specifically, the Otsu method is used to determine the adaptive threshold. The Otsu method is an automatic method for determining the optimal threshold, and its principle is to maximize the inter-class variance. For each two-dimensional slice in PZR, its histogram is calculated, and the threshold that maximizes the inter-class variance is selected as the adaptive threshold for that slice. The adaptive thresholds of all slices are combined into an adaptive threshold distribution map (ATD).

[0086] Based on PZR and ATD, PZR is binarized to classify voxels into PZ candidate voxels and non-PZ voxels. Specifically, for each voxel in PZR, its value is compared with its corresponding adaptive threshold (the value at the corresponding position in ATD). If PZR[x, y, z] > ATD[x, y, z], the voxel is marked as a PZ candidate voxel; otherwise, it is marked as a non-PZ voxel. All PZ candidate voxels are combined into a peripheral zone candidate region map PZC. PZC[x, y, z] = 1 indicates that the voxel is a PZ candidate voxel, and PZC[x, y, z] = 0 indicates that the voxel is not a PZ candidate voxel.

[0087] Based on the PZC, PZ candidate voxels are further screened, and voxels that do not conform to the range of PZ physical property parameters are removed. Specifically, for each PZ candidate voxel in the PZC (i.e., the voxel with PZC[x, y, z] = 1), its three key physical property parameters (tissue stiffness, ADC value, and Ktrans value) are checked to see if they are all within the range of PZ parameters specified in the PPCS. If all three parameters are within the range, the voxel is retained; otherwise, the voxel is marked as a non-PZ voxel. The screened voxels are combined into a peripheral zone constraint screening map PZCS. PZCS[x, y, z] = 1 indicates that the voxel is a PZ candidate voxel after physical parameter constraint screening, and PZCS[x, y, z] = 0 indicates that the voxel is not.

[0088] Based on PZCS, morphological optimization is performed on PZ candidate regions to remove small isolated areas, fill small holes, and smooth boundaries. Specifically, the following operations are performed: First, an opening operation is performed. The opening operation is an erosion-dilation operation that can remove small isolated areas. A 3×3×3 cube structuring element is used, with two iterations of erosion followed by two iterations of dilation. Then, a closing operation is performed. The closing operation is a dilation-erosion operation that can fill small holes. A 3×3×3 cube structuring element is used, with two iterations of dilation followed by two iterations of erosion. Finally, boundary smoothing is performed. A Gaussian filtering method is used to smooth the PZCS. A 5×5×5 Gaussian kernel is used, with a standard deviation set to 1. The morphologically optimized result is used as the final peripheral zone segmentation mask image PZM. PZM[x, y, z] = 1 indicates that the voxel belongs to a PZ, and PZM[x, y, z] = 0 indicates that the voxel does not belong to a PZ.

[0089] Preferably, the transition zone conditional segmentation in step S3 includes:

[0090] Based on the peripheral zone segmentation mask map, the remaining region of the morphological-physical joint feature map is extracted and preprocessed to obtain the processed remaining region map.

[0091] Based on the progressive constraint set of physical characteristics, a transition zone feature scoring map is constructed from the processed remaining region map to obtain the transition zone feature scoring map;

[0092] Seed points are selected and initialized based on the transition zone feature scoring map to obtain the transition zone seed point set;

[0093] A conditional growth algorithm is constructed based on the seed point set of the transition zone.

[0094] A conditional growth algorithm is used to perform region growth on the seed point set of the transition zone to obtain a transition zone segmentation mask map.

[0095] In this embodiment of the invention, it is necessary to exclude the segmented peripheral zone (PZ) region from the median buffer (MP). Specifically, for each voxel in the MP, it is checked whether it is marked as PZ in the peripheral zone segmentation mask PZM (i.e., whether PZM[x, y, z] equals 1). If so, the values ​​of all channels of that voxel in the MP are set to 0; otherwise, the values ​​in the MP are kept unchanged. This yields a residual region map (RR). Then, the RR is preprocessed. Due to the exclusion of the PZ region, there will be some isolated noise points or small holes in the RR. Median filtering is used to remove these noise points. A 3×3×3 cubic filter is used to perform median filtering on each channel of the RR. The filtered result is used as the processed residual region map (PRR).

[0096] Based on the PRR and PPCS, the probability that each voxel in the PRR belongs to the transition zone (TZ) is calculated. Specifically, the range of physical property parameters of the TZ is determined according to the rules in the PPCS. For example, the PPCS has the following rule: IF(region=TZ)THEN(tissue stiffness∈[c1,d1], ADC value∈[c2,d2], Ktrans value∈[c3,d3]). Here, c1, d1, c2, d2, c3, and d3 are the parameter ranges of the TZ determined according to the region parameter space model APSM and the segmented region (PZ). Then, these parameter ranges are used to perform feature scoring on the PRR. A method similar to peripheral zone feature enhancement is adopted. For each voxel in the PRR, the degree of matching between its three key physical property parameters (tissue stiffness, ADC value, and Ktrans value) and the TZ parameter range is calculated. The matching function is: Match(x,[a,b])=exp(-(x-(a+b) / 2)). / (2×((ba) / 2) Calculate the product of the matching functions for the three parameters of each voxel, which is the probability that the voxel belongs to the TZ: P_TZ = Match(tissue stiffness, [c1, d1]) × Match(ADC value, [c2, d2]) × Match(Ktrans value, [c3, d3]). Combine the P_TZ values ​​of all voxels into a transition zone feature score map TZS. TZS[x, y, z] represents the probability that the voxel with coordinates (x, y, z) belongs to the TZ.

[0097] Based on the TZS (Transition Zone Scaling), suitable seed points are selected for subsequent region growing. Specifically, the following strategy is adopted: First, a seed point selection threshold Tseed is set. For example, Tseed = 0.9 × max(TZS), where max(TZS) represents the maximum value in TZS. Then, voxels in TZS with values ​​greater than Tseed are searched as candidate seed points. To avoid over-concentration of seed points, non-maximum suppression is applied to the candidate seed points. Specifically, for each candidate seed point, if other candidate seed points exist within its 5×5×5 neighborhood, only the seed point with the largest TZS value is retained. The remaining candidate seed points after non-maximum suppression are used as the final seed points. The coordinates of these seed points are combined into a transition zone seed point set TZSD.

[0098] Based on TZSD, a conditional region growth algorithm is constructed to progressively expand the TZ region starting from a seed point. Specifically, the algorithm includes the following elements: Growth conditions: For a voxel to be included in a TZ region, two conditions must be met: (1) The voxel must be adjacent to an existing TZ region (26 neighbors); (2) The voxel's TZS value must be greater than a growth threshold Tgrow. Tgrow can be set to a value lower than Tseed, for example, Tgrow = 0.7 × max(TZS). Stopping condition: The algorithm stops when no new voxels can be included in the TZ region. Algorithm flow: First, all seed points in TZSD are marked as TZ regions. Then, the 26 neighbors of the boundary voxels of the TZ region are iteratively checked. For each neighboring voxel, if it meets the growth conditions, it is included in the TZ region and marked as a new boundary voxel. This process is repeated until no new voxels can be included in the TZ region.

[0099] Based on the constrained region growing algorithm and TZSD, region growing is performed to obtain the final TZ segmentation result. Specifically, following the algorithm flow, starting from the seed point in TZSD, the TZ region is gradually expanded until no new voxels can be included in the TZ region. All voxels included in the TZ region are marked as 1, and other voxels are marked as 0, resulting in a binary image, namely the transition band segmentation mask image TZM. TZM[x, y, z] = 1 indicates that the voxel belongs to TZ, and TZM[x, y, z] = 0 indicates that the voxel does not belong to TZ.

[0100] Preferably, the determination of the central zone boundary and the mapping of physical property distribution in step S3 includes:

[0101] The central zone boundary is determined and integrated based on the peripheral zone segmentation mask, the transition zone segmentation mask, and the morphological and physical joint feature map to obtain the region segmentation label map;

[0102] Physical property distribution mapping is performed on the region segmentation label map and the morphological-physical joint feature map to obtain the region physical property distribution map.

[0103] In this embodiment of the invention, since the three main regions of the prostate (PZ, TZ, CZ) are adjacent to each other, and PZ and TZ have already been segmented, the boundary of CZ can be determined by excluding the PZ and TZ regions. Specifically, for each voxel in the morphophysical joint feature map MP, it is checked whether it is marked as PZ in PZM (i.e., whether PZM[x,y,z] equals 1) or marked as TZ in TZM (i.e., whether TZM[x,y,z] equals 1). If neither is true, the voxel is marked as CZ. In this way, a preliminary CZ region is obtained. Then, morphological optimization is performed on this preliminary CZ region. Since there may be some errors in the segmentation of PZ and TZ, the boundary of the CZ region is not smooth, or there are some small isolated regions or holes. A morphological optimization method (opening and closing operations) similar to that used in peripheral zone segmentation is used to remove small isolated regions and holes, and smooth the boundary. Specifically, a 3×3×3 cube structuring element is used, performing two iterations of opening and two iterations of closing operations. The optimized CZ region is then integrated with the PZ and TZ regions. A new image RSL is created, where RSL[x, y, z] represents the region label of the voxel at coordinates (x, y, z). If PZM[x, y, z] = 1, then RSL[x, y, z] = 1 (representing PZ); if TZM[x, y, z] = 1, then RSL[x, y, z] = 2 (representing TZ); if the voxel belongs to neither PZ nor TZ, then RSL[x, y, z] = 3 (representing CZ). This yields a region segmentation label map RSL, which fully represents the segmentation results of the three regions of the prostate.

[0104] Based on RSL and MP, the physical characteristic parameters in MP are mapped to corresponding regions to generate a Regional Physical Characteristic Distribution Map (RPD). Specifically, for each voxel in MP, its physical characteristic parameters (the 8 physical feature channels in MP) are mapped to the corresponding region in RPD according to its region label in RSL. RPD is a multi-channel image, RPD[x, y, z, c], where c=1, 2, ..., 8 correspond to the 8 physical feature channels in MP. If RSL[x, y, z]=1 (representing PZ), then RPD[x, y, z, 1:8]=MP[x, y, z, 1:8]; if RSL[x, y, z]=2 (representing TZ), then RPD[x, y, z, 1:8]=MP[x, y, z, 1:8]; if RSL[x, y, z]=3 (representing CZ), then RPD[x, y, z, 1:8]=MP[x, y, z, 1:8]. If needed, each channel of the RPD can be pseudo-color encoded for easier visualization. For example, tissue stiffness can be mapped to red, ADC values ​​to green, and Ktrans values ​​to blue. This yields a regional physical property distribution map (RPD), which not only shows the segmentation results of the three regions of the prostate but also the distribution of physical property parameters within each region.

[0105] Preferably, step S4 includes the following steps:

[0106] Step S41: Using the regional physical characteristic distribution map and the standardized multimodal image set as input, a lightweight convolutional neural network is constructed using deep learning technology to perform deep learning feature optimization and obtain a deep feature enhancement map;

[0107] Step S42: Perform microscale feature analysis on the depth feature enhancement map to obtain a microscale feature consistency map;

[0108] Step S43: Perform mesoscale structure evaluation based on the depth feature enhancement map and the microscale feature consistency map to obtain the mesoscale structure scoring matrix;

[0109] Step S44: Perform overall macro-scale consistency verification based on the deep feature enhancement map and meso-structure scoring matrix to obtain a set of macro-scale consistency indicators;

[0110] Step S45: Perform multi-scale consistency scoring on the micro-feature consistency map, meso-structure scoring matrix, and macro-coordination index set to obtain a multi-scale consistency scoring map;

[0111] Step S46: Calculate the classification confidence index based on the multi-scale consistency score map and the deep feature enhancement map to obtain the region classification confidence index.

[0112] In this embodiment of the invention, deep learning methods are used to further optimize features, particularly enhancing the features of region boundaries. The input data are the Region Physical Properties Distribution Map (RPD) obtained in step S3 and the Standardized Multimodal Image Set (SMI) obtained in step S11. A lightweight convolutional neural network (CNN) is constructed. This CNN employs an encoder-decoder architecture. The encoder part contains multiple convolutional and pooling layers for extracting multi-scale features. For example, three convolutional blocks can be used, each containing two 3×3 convolutional layers and one 2×2 max-pooling layer. The number of convolutional kernels can be increased block by block; for example, the first convolutional block uses 16 kernels, the second uses 32 kernels, and the third uses 64 kernels. The decoder part contains multiple upsampling and convolutional layers for restoring image resolution and generating the final feature map. The decoder part can use a structure symmetrical to the encoder part, i.e., three upsampling blocks, each containing one 2×2 upsampling layer and two 3×3 convolutional layers. The number of convolutional kernels can be decreased block by block. To preserve detail, skip connections are added between the encoder and decoder. Specifically, the output of each convolutional block of the encoder is concatenated with the input of the corresponding upsampled block of the decoder. The network's input layer receives the concatenation of RPD and SMI. Since RPD has 8 channels and SMI has 3 channels (elastic imaging, DWI, DCE-MRI), the input layer has 11 channels. The network's output layer is a single-channel image representing the Deep Feature Enhancement Map (DFE). The value of DFE[x, y, z] represents the intensity of the voxel in the deep feature space, reflecting the likelihood that the voxel belongs to a region boundary. The loss function consists of two parts: boundary loss and region consistency loss. The boundary loss, which can be Dice loss or cross-entropy loss, measures the difference between the DFE and the true boundary. The true boundary can be obtained by edge detection on the region segmentation label map RSL. The region consistency loss, which can be mean squared error loss, measures the smoothness of the DFE within each region. The total loss function is a weighted sum of these two losses. The network is trained using the Adam optimizer. The training dataset can be obtained by processing prostate images from multiple patients using steps S1-S3 as described above. After training, the RPD and SMI of new prostate images are input into the network to obtain the corresponding DFE.

[0113] For each voxel in the DFE, calculate the mean and standard deviation of its eigenvalues ​​within a 26-neighborhood. The mean reflects the average eigenvalue intensity around the voxel, and the standard deviation reflects the degree of variation in eigenvalue intensity. If a voxel's eigenvalue differs significantly from the eigenvalues ​​of its neighboring voxels (e.g., greater than the neighborhood mean plus twice the standard deviation, or less than the neighborhood mean minus twice the standard deviation), the voxel's eigenvalues ​​are considered inconsistent. Construct a Micro-Feature Consistency Graph (MFC). MFC[x, y, z] = 1 indicates that the voxel's eigenvalues ​​are consistent with the eigenvalues ​​of its neighboring voxels, and MFC[x, y, z] = 0 indicates inconsistency.

[0114] Divide the DFE into multiple sub-regions. Regular grid partitioning can be used, for example, dividing the DFE into multiple 3×3×3 sub-regions. Adaptive partitioning based on image content can also be used, for example, using a supervoxel segmentation algorithm. For each sub-region, calculate the following metrics: (1) the mean and standard deviation of the DFE value; (2) the proportion of voxels labeled as consistent in the MFC; and (3) the dominant gradient direction of the DFE value (if present). The dominant gradient direction can be obtained by calculating the gradients of all voxels within the sub-region and then summing the gradient vectors. If the dominant gradient direction is not obvious (e.g., the magnitude of the gradient vector is small), the sub-region is considered to have no obvious dominant gradient direction. Based on these metrics, evaluate whether the structural characteristics of each sub-region meet expectations. For example, for a sub-region located at the region boundary, its mean DFE value should be high, the proportion of voxels labeled as consistent in the MFC should be low, and there should be an obvious dominant gradient direction (perpendicular to the boundary). For a sub-region located inside the region, its mean DFE value should be low, the proportion of voxels labeled as consistent in the MFC should be high, and there should be no obvious dominant gradient direction. Based on the evaluation results, each sub-region is scored. For example, a score between 0 and 1 can be used to represent the degree to which the structural characteristics of the sub-region conform to the expectations. The scores of all sub-regions are combined into a mesostructural score matrix (MSE). MSE[i,j] represents the score of the i-th sub-region on the j-th indicator.

[0115] The overall consistency of the DFE was assessed on a prostate-wide scale. Specifically, the following aspects were evaluated: (1) whether the DFE clearly delineated the boundaries of the three main regions of the prostate (PZ, TZ, CZ); (2) whether the boundaries of different regions were smooth and continuous; and (3) whether the DFE conformed to known anatomical knowledge (e.g., the PZ is typically located posterolaterally to the prostate). To assess the clarity of the boundaries, the gradient magnitude of the DFE near the region boundaries could be calculated. A larger gradient magnitude indicates a clearer boundary. To assess the smoothness and continuity of the boundaries, the curvature of the boundaries could be calculated. A smaller curvature indicates a smoother boundary. To assess the degree of conformity with anatomical knowledge, the DFE could be registered with a standard prostate atlas, and the distance between the registered DFE and the region boundaries in the atlas could be calculated. A smaller distance indicates a higher degree of conformity. Based on the above assessment results, several macroscopic consistency indices were calculated. For example: (1) Boundary sharpness index: the average gradient magnitude of the DFE near the region boundary; (2) Boundary smoothness index: the average curvature of the region boundary; (3) Anatomical conformity index: the average distance between the DFE and the region boundary in the standard atlas. These indices are combined into a macro-consistency index set MGI.

[0116] A weighted average method is used. For each voxel, its multiscale consistency score (MSC) is calculated using the following formula: MSC = w1 × MFC + w2 × MSE_score + w3 × MGI_score. Where MFC is the voxel's value (0 or 1) in the micro-feature consistency map; MSE_score is the average score of the sub-region to which the voxel belongs in the meso-structure score matrix MSE; MGI_score is a normalized score calculated based on the macro-coherence index set MGI (e.g., each index in MGI can be normalized to between 0 and 1, and then the average is taken); w1, w2, and w3 are weights representing the importance of the evaluation results at the three scales. The weights can be set empirically or determined experimentally. The MSC values ​​of all voxels are combined into a single multiscale consistency score map (MSC).

[0117] For each voxel, its Region Classification Confidence Index (RCI) is calculated using the following formula: RCI = MSC × (1 + DFE). Here, MSC is the multi-scale consistency score of the voxel, and DFE is the voxel's value in the deep feature enhancement map. This formula means that if a voxel's features are consistent with those of its neighboring voxels (higher MSC) and the voxel is located near the region boundary (higher DFE), its classification confidence is high; conversely, if a voxel's features are inconsistent with those of its neighboring voxels (lower MSC), or the voxel is far from the region boundary (lower DFE), its classification confidence is low. The RCI values ​​of all voxels are combined to form a Region Classification Confidence Index (RCI) map. A higher RCI value indicates a more reliable classification result for that voxel.

[0118] Of particular importance is that the deep learning feature optimization described in step S41 specifically involves:

[0119] Physical characteristic-guided feature channels are constructed from regional physical characteristic distribution maps and standardized multimodal image sets to obtain a physical guidance feature set;

[0120] An adaptive domain attention network is constructed on the physical guidance feature set to obtain a multi-domain feature model;

[0121] A physically constrained optimization model is obtained by training a network under physical constraints on a multi-domain feature model.

[0122] By using a physical constraint optimization model to perform boundary-region collaborative enhancement, a collaborative enhancement feature map is obtained.

[0123] Physiological verification and fine-tuning were performed based on the regional physical characteristic distribution map and the collaborative enhancement feature map to obtain the deep feature enhancement map;

[0124] In this embodiment of the invention, channel expansion is performed on RPD and SMI. RPD originally has 8 channels (corresponding to 8 physical features). These 8 channels are used as the first 8 channels of the physical guidance feature set. SMI originally has 3 channels (elastic imaging, DWI, DCE-MRI). For each channel of SMI, its gradient in three directions (x, y, z) is calculated, resulting in 3 gradient maps. The original 3 channels and 9 gradient maps (3 channels × 3 directions) are used as the last 12 channels of the physical guidance feature set. Thus, the physical guidance feature set PGF contains 20 channels (8 physical features + 3 original image channels + 9 gradient maps). PGF[x, y, z, c], where c = 1, 2, ..., 20. A convolutional neural network based on an attention mechanism is constructed. This network adopts an encoder-decoder structure. The encoder part contains multiple convolutional blocks. Each convolutional block contains two 3×3 convolutional layers and an adaptive domain attention (ADA) module. The ADA module is crucial. The input to the ADA module is the output feature map of the convolutional layer. The ADA module first performs average pooling on the feature map in the spatial dimension, obtaining a channel-dimensional vector. Then, this vector is passed through a fully connected layer to obtain a weight vector with the same number of channels. This weight vector represents the importance of each channel. The weight vector is multiplied by the original feature map to obtain a weighted feature map. Finally, the weighted feature map is passed through a 1×1 convolutional layer to obtain the output of the ADA module. The mathematical expression of the ADA module is as follows:

[0125] Input: X∈ ;

[0126] 1.Global Average Pooling: v=GAP(X)∈ ;

[0127] 2.Fully Connected Layer:w=FC(v)∈ ;

[0128] 3.Channel-wise Multiplication: Y=X⊙sigmoid(w);

[0129] 4. Output: Z = Conv1x1(Y) ∈ ;

[0130] Where H, W, and C represent the height, width, and number of channels of the feature map, respectively; GAP represents global average pooling; FC represents a fully connected layer; ⊙ represents channel-wise multiplication; and Conv1x1 represents a 1x1 convolution. The sigmoid function is used to convert the weight values ​​to a range between 0 and 1.

[0131] The ADA module's key feature lies in its ability to adaptively adjust the importance weights of different physical feature domains, enabling the network to dynamically focus on the most discriminative features across various cases and anatomical structures. To further enhance the ADA module's performance, we introduce a multi-head attention mechanism. Feature channels are grouped, with attention weights calculated independently for each group, and then the results are merged. Specifically, for C input channels, we evenly divide them into h groups (h is the number of heads, typically 4 or 8), each containing C / h channels. Channel attention is calculated for each group separately, resulting in h C / h-dimensional weight vectors, which are then concatenated into a complete C-dimensional weight vector. This multi-head mechanism captures the complex relationships between different feature combinations, improving the model's expressive power. Furthermore, we add residual connections to the ADA module, i.e., Y = X ⊙ sigmoid(w) + X, which helps alleviate the vanishing gradient problem and preserves original feature information. During training, we apply L1 regularization to the ADA module's weight vector, prompting the model to learn a sparse attention distribution, focusing more on the most critical feature channels while reducing the risk of overfitting.

[0132] The encoder part can contain multiple convolutional blocks, such as 3 convolutional blocks. The number of kernels in each convolutional block can increase progressively, for example, 16, 32, or 64. The decoder part contains multiple upsampling blocks. Each upsampling block contains one upsampling layer and two 3×3 convolutional layers. The structure of the decoder part can be symmetrical to that of the encoder part. To preserve detail, skip connections are added between the encoder and decoder. The output of each convolutional block of the encoder is concatenated with the input of the corresponding upsampling block of the decoder. The input of the network is the Physical Guided Feature Set (PGF) (20 channels). The output of the network is a Multi-Domain Feature Model (MDFM), which contains multiple feature maps, each corresponding to a specific region or feature combination. Define the loss function. The loss function consists of three parts: region segmentation loss, boundary loss, and physical consistency loss. Region segmentation loss: Using Dice loss or cross-entropy loss, it measures the difference between the segmentation result of MDFM and the true region label (RSL obtained in step S3). Boundary loss: Using Dice loss or cross-entropy loss, it measures the difference between the boundary prediction result of MDFM and the true boundary. The true boundary can be obtained by edge detection of RSL. Physical consistency loss: For each region of MDFM, calculate the mean and standard deviation of its physical property parameters (tissue stiffness, ADC value, Ktrans value). Then, calculate the difference between these statistics and the parameter range specified in PPCS obtained in step S3 "Constructing a progressive constraint set of physical properties". If the statistics are out of range, add a penalty term. The total loss function is the weighted sum of the three losses. Train the network using the Adam optimizer. The training dataset can be obtained by processing prostate images of multiple patients using the above steps S1-S3. After training, a physical constraint optimization model PCOM is obtained. Input the PGF of the new prostate image into PCOM. PCOM will output a multi-channel feature map. Post-process this feature map to obtain the final co-enhanced feature map CEF. Post-processing includes the following steps: (1) Region segmentation: Determine the region (PZ, TZ, CZ) to which each voxel belongs based on the region segmentation probability map output by PCOM. (2) Boundary extraction: Determine whether each voxel belongs to the boundary based on the boundary probability map output by PCOM. (3) Feature Fusion: For each voxel, the features of its region, boundary features, and original physical features are fused. A weighted average method can be used. The weights can be determined based on the probability values ​​output by PCOM. The fused features are used as the values ​​of the corresponding voxels in CEF. CEF is compared with the RPD obtained in step S3. For each region (PZ, TZ, CZ), the mean and standard deviation of the physical property parameters (tissue stiffness, ADC value, Ktrans value) of the corresponding region in CEF are calculated. These statistics are compared with the statistics of the corresponding region in RPD. If the difference is too large (e.g., greater than 2 standard deviations), CEF is adjusted.Adjustments can be made by locally smoothing the CEF or adjusting the weights of feature fusion. The CEF should be checked for obvious artifacts or discontinuities. If any are found, morphological filtering or other methods should be used to correct them. The validated and adjusted CEF should be used as the final Deep Feature Enhancement (DFE).

[0133] Of particular importance is that the multi-scale consistency scoring in step S45 specifically refers to:

[0134] The micro-feature consistency map, meso-structure scoring matrix, and macro-coordination index set are standardized in feature space to obtain a multi-scale feature standardization set;

[0135] Calculate the scale information entropy weight set of the multi-scale feature standardization set;

[0136] Deep features and multi-scale features are fused together with the scale information entropy weight set and the multi-scale feature standardization set to obtain a multi-scale fused feature map.

[0137] Construct a set of consistency scoring functions based on multi-scale fused feature maps;

[0138] A global scoring map is generated and optimized from the multi-scale fused feature map using a set of consistency scoring functions to obtain a multi-scale consistency scoring map.

[0139] In this embodiment of the invention, for MFC, since it is a binary image, no additional standardization processing is required. For MSE, each index needs to be normalized using the min-max normalization method to scale the value of each index to the [0, 1] interval. For MGI, since it contains multiple indices, and each index has different dimensions and value ranges, standardization processing is required using the z-score normalization method to convert the value of each index to a standard normal distribution. The standardized MFC, MSE', and MGI' are combined into a multi-scale feature normalization set MSNS. The weights are calculated using the information entropy method. The smaller the information entropy, the greater the information content of that scale, and the higher the weight. The specific operation is as follows: For each scale (micro, meso, macro) in MSNS, its information entropy is calculated. Based on the information entropy of the three scales, the weight of each scale is calculated, and the weight is inversely proportional to the information entropy. By normalizing the reciprocal of the information entropy, the weight of each scale is obtained, and they are combined into a scale information entropy weight set SIEW. DFE is then fused with MFC. Since MFC is a binary image and DFE is a continuous image, DFE needs to be binarized. A fixed threshold can be used, or the Otsu method can be used to adaptively determine the threshold. The binarized DFE is multiplied voxel-by-voxel by MFC to obtain the fused micro-feature map MFM. Then, MFM, MSE', and MGI' are weighted and averaged according to the weights calculated in SIEW. For each voxel, its multi-scale fusion feature value MSF is calculated by the following formula: MSF = w_micro × MFM + w_meso × MSE'_score + w_macro × MGI'_score, where MSE'_score is the average score of the sub-region to which the voxel belongs in the standardized MSE', and MGI'_score is a normalized score calculated based on the standardized MGI' (for example, each index in MGI' can be normalized to between 0 and 1, and then the average is taken). The MSF values ​​of all voxels are combined into a multi-scale fusion feature map MSFF. A simple linear function is used: Score = a × MSF + b. Here, a and b are constants, which can be set empirically or determined experimentally. Alternatively, a non-linear function, such as the sigmoid function: Score = 1 / (1 + exp(-c × (MSF - d))). Here, c and d are constants, which can be set empirically or determined experimentally. This function is used as the consistency scoring function (CSF). The MSF value of each voxel in the MSF is input into the CSF to calculate the consistency score for that voxel. The consistency scores of all voxels are combined into a global score map (GSC). The GSC is then optimized. Because the calculation of the MSF may contain some noise, some isolated low-scoring or high-scoring regions may appear in the GSC.Median filtering was used to smooth the GSC. A 3×3×3 cubic filter was used. The filtered result was used as the final multi-scale consensus score (MSC). MSC[x, y, z] represents the multi-scale consensus score of a voxel with coordinates (x, y, z). This score integrates feature information from the micro, meso, and macro scales, as well as depth feature information, reflecting the reliability of the classification result of this voxel at different scales.

[0140] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0141] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for constructing a prostate cancer image recognition and classification model based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain a standardized multimodal image set; construct an inter-regional characteristic ratio map based on the standardized multimodal image set; construct a regional physical feature tensor based on the inter-regional characteristic ratio map; Step S2: Divide the standardized multimodal image set into three regions: peripheral zone, transition zone, and central zone. Introduce anatomical prior knowledge as constraints. Construct a nonlinear mapping function based on the regional physical feature tensor to obtain a morphophysical mapping function set. Generate a morphophysical joint feature map based on the regional physical feature tensor and the morphophysical mapping function set. Step S3: Construct a progressive constraint set of physical properties based on the morphological-physical joint feature map; perform peripheral zone priority segmentation on the morphological-physical joint feature map based on the progressive constraint set of physical properties to obtain a peripheral zone segmentation mask map; perform transition zone conditional segmentation based on the peripheral zone segmentation mask map to obtain a transition zone segmentation mask map; determine the central zone boundary and map the physical property distribution based on the transition zone segmentation mask map to obtain a regional physical property distribution map. Step S4: Perform deep learning feature optimization on the regional physical characteristic distribution map to obtain a deep feature enhancement map; perform micro-medium-macro progressive verification and scoring on the deep feature enhancement map to obtain the regional classification credibility index.

2. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Acquire and preprocess prostate multimodal medical images to obtain a standardized multimodal image set, which includes elastography, DWI sequence and DCE-MRI sequence; Step S12: Calculate the regional feature parameters of the standardized multimodal image set to obtain the regional physical parameter matrix; Step S13: Construct a region characteristic ratio map based on the region physical parameter matrix; Step S14: Perform multi-parameter feature fusion and tensor construction on the regional physical parameter matrix and the inter-regional characteristic ratio map to obtain the regional physical feature tensor.

3. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 2, characterized in that, Step S13 includes the following steps: Step S131: Define the boundary transition zone based on the regional physical parameter matrix to obtain the boundary transition zone location map; Step S132: Calculate the multi-parameter directional gradient based on the boundary transition zone location map and the regional physical parameter matrix to obtain the directional parameter gradient set; Step S133: Calculate the local weighted ratio based on the regional physical parameter matrix and the gradient set of directional parameters to obtain the local parameter ratio set; Step S134: Perform regional feature enhancement processing on the local ratio set of parameters to obtain an enhanced ratio map set; Step S135: Perform multi-parameter ratio fusion on the enhanced ratio map to obtain the inter-regional characteristic ratio map.

4. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Based on anatomical knowledge, the standardized multimodal image set is divided into three regions: peripheral zone, transition zone, and central zone. Morphological features are extracted and quantified based on the regional physical feature tensor to obtain a morphological feature description set. Step S22: Encode the anatomical prior knowledge into the morphological feature description set to obtain the anatomical prior constraint library; Step S23: Construct a nonlinear mapping function based on the region physical feature tensor, morphological feature description set, and anatomical prior constraint library to obtain the morphophysical mapping function set; Step S24: Generate a morphophysical joint feature map based on the region physical feature tensor and the morphophysical mapping function set.

5. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 4, characterized in that, Step S23 includes the following steps: Step S231: Extract key physical property parameters from the region physical feature tensor, including tissue stiffness, water diffusion, and blood perfusion; extract morphological features from the morphological feature description set, including boundary curvature, shape factor, and relative position. Step S232: Perform feature correlation on key physical property parameters and morphological features to obtain the feature correlation matrix; Step S233: Extract the boundary curvature distribution from the morphological feature description set, construct the curvature-related weight function based on the feature correlation matrix, and obtain the curvature weight function set; Step S234: Extract anatomical constraint rules for the prostate region from the anatomical prior constraint library, construct anatomical constraint penalty terms based on the curvature weight function set, and obtain the anatomical penalty function; Step S235: Design the boundary gradient penalty term based on the anatomical penalty function and the morphological feature description set to obtain the boundary gradient penalty function; Step S236: Integrate nonlinear mapping functions based on the feature correlation matrix, curvature weight function set, anatomical penalty function, and boundary gradient penalty function to obtain the morphophysical mapping function set.

6. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 1, characterized in that, The construction of the progressive constraint set of physical properties in step S3 includes: Statistical analysis of regional physical property parameters was performed on the morphophysical joint feature map to obtain a set of regional parameter statistical features; Using tissue stiffness, ADC value, and blood perfusion as three coordinate axes, a three-dimensional physical property parameter space was constructed to obtain a regional parameter space model; Calculate the regional difference matrix of the regional parameter spatial model; Determine the region segmentation priority table based on the region difference matrix; Based on the regional parameter space model and the regional segmentation priority table, progressive constraint rules are designed to obtain a progressive constraint set of physical characteristics.

7. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 1, characterized in that, The peripheral zone priority segmentation mentioned in step S3 includes: Based on the progressive constraint set of physical properties, the morphological-physical joint feature map is enhanced with peripheral zone features to obtain the peripheral zone feature response map. Adaptive threshold determination is performed based on the peripheral zone characteristic response map and the progressive constraint set of physical properties to obtain an adaptive threshold distribution map. The adaptive threshold distribution map is used to perform preliminary segmentation and candidate region extraction on the peripheral zone feature response map to obtain the peripheral zone candidate region map. Based on the candidate region map of the outer periphery zone, physical parameter constraints are used for screening to obtain the outer periphery zone constraint screening map; Morphological optimization and contour refinement are performed based on the peripheral zone constraint screening map to obtain the peripheral zone segmentation mask map.

8. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 1, characterized in that, The transition zone conditional segmentation in step S3 includes: Based on the peripheral zone segmentation mask map, the remaining region of the morphological-physical joint feature map is extracted and preprocessed to obtain the processed remaining region map. Based on the progressive constraint set of physical characteristics, a transition zone feature scoring map is constructed from the processed remaining region map to obtain the transition zone feature scoring map; Seed points are selected and initialized based on the transition zone feature scoring map to obtain the transition zone seed point set; A conditional growth algorithm is constructed based on the seed point set of the transition zone. A conditional growth algorithm is used to perform region growth on the seed point set of the transition zone to obtain a transition zone segmentation mask map.

9. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 1, characterized in that, The determination of the central zone boundary and the mapping of physical property distribution in step S3 includes: The central zone boundary is determined and integrated based on the peripheral zone segmentation mask, the transition zone segmentation mask, and the morphological and physical joint feature map to obtain the region segmentation label map; Physical property distribution mapping is performed on the region segmentation label map and the morphological-physical joint feature map to obtain the region physical property distribution map.

10. The method for constructing a deep learning-based prostate cancer image recognition and classification model according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Using the regional physical characteristic distribution map and the standardized multimodal image set as input, a lightweight convolutional neural network is constructed using deep learning technology to perform deep learning feature optimization and obtain a deep feature enhancement map; Step S42: Perform microscale feature analysis on the depth feature enhancement map to obtain a microscale feature consistency map; Step S43: Perform mesoscale structure evaluation based on the depth feature enhancement map and the microscale feature consistency map to obtain the mesoscale structure scoring matrix; Step S44: Perform overall macro-scale consistency verification based on the deep feature enhancement map and meso-structure scoring matrix to obtain a set of macro-scale consistency indicators; Step S45: Perform multi-scale consistency scoring on the micro-feature consistency map, meso-structure scoring matrix, and macro-coordination index set to obtain a multi-scale consistency scoring map; Step S46: Calculate the classification confidence index based on the multi-scale consistency score map and the deep feature enhancement map to obtain the region classification confidence index.

Images