A prostate cancer detection and analysis method based on image analysis

Abstract

This invention discloses a prostate cancer detection and analysis method based on image analysis, belonging to the field of prostate cancer analysis technology. The method includes: acquiring the patient's original MRI images and performing standardized preprocessing, followed by data augmentation of the processed MRI images; having a radiologist perform decoupled sign annotation on the preprocessed images and analyze the accuracy of the diagnostic results based on the annotation process; constructing a dual-path three-dimensional convolutional neural network model consisting of two encoders and a shared decoder; jointly training the dual-path three-dimensional convolutional neural network model using labeled data to optimize its performance; testing the performance of the optimized model using independent data samples to identify weaknesses and root causes of erroneous cases; and iteratively upgrading the model using augmented training data based on the test results. This invention, by constructing a three-dimensional convolutional neural network model, can improve diagnostic efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of prostate cancer analysis technology, and in particular to a prostate cancer detection and analysis method based on image analysis. Background Technology

[0002] Prostate cancer (PCa) is one of the most common types of cancer among middle-aged and elderly men. According to the latest epidemiological statistics, the incidence of PCa in my country has increased dramatically in recent years. In all cancer registry areas nationwide, prostate cancer currently ranks first among malignant tumors of the male genitourinary system in China, surpassing bladder cancer. Early diagnosis and treatment of prostate cancer are crucial for patient prognosis and quality of life. Accurate preoperative diagnostic assessment can not only reduce overtreatment of non-clinically significant cancers but also improve treatment outcomes by combining multiple therapies.

[0003] Magnetic resonance imaging (MRI) is a key imaging technique for the early detection of prostate cancer. To improve diagnostic efficiency and consistency, computer-aided detection and diagnosis (CAD) systems based on medical images have become an important research direction in this field. In recent years, deep learning technology, especially convolutional neural networks (CNNs), has gradually replaced traditional machine learning, becoming the core driving force for the development of CAD systems and demonstrating significant advantages in image analysis of various diseases.

[0004] However, current mainstream CNN-based MRI detection methods for prostate cancer still have several limitations. At the data processing level, most methods utilize multi-parameter MRI (such as T2WI, DWI, and ADC sequences) in a relatively simple manner, typically performing only basic registration and normalization. They lack preprocessing and enhancement strategies that can deeply integrate anatomical and functional information between sequences, limiting the model's ability to perceive essential pathological features. Secondly, at the model construction level, network design often aims to improve the performance of a single segmentation or classification task. Its internal decision-making process lacks interpretability and cannot be correlated with the clinical logic of radiologists making diagnoses based on specific imaging features, making its results difficult to fully trust. At the model optimization level, the training process usually stops at a one-time static training, lacking a closed-loop mechanism that can self-diagnose and iteratively evolve based on the specific weaknesses exposed by the model during testing. This makes it difficult for the model to continuously adapt to complex and ever-changing clinical case distributions.

[0005] To address the above shortcomings, this invention proposes a prostate cancer detection and analysis method based on image analysis. Summary of the Invention

[0006] This invention provides a prostate cancer detection and analysis method based on image analysis. By constructing a three-dimensional convolutional neural network model, it realizes automated analysis of image data to accurately identify early prostate cancer, shorten diagnosis time, and improve diagnostic efficiency and accuracy.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for prostate cancer detection and analysis based on image analysis, comprising: S1: Acquire the patient's raw MRI images and perform standardized preprocessing, and then perform data augmentation on the processed MRI images; S2: Radiologists decouple and annotate the signs in the preprocessed images, and analyze the authenticity of the diagnostic results based on the annotation process; S3: Construct a dual-path 3D convolutional neural network model consisting of two encoders and a shared decoder; S4: Use labeled data to jointly train the dual-path 3D convolutional neural network model and optimize its performance; S5: Use independent data samples to test the performance of the optimized model, and locate the weak points of the model and the root causes of the failure of error cases; S6: Iteratively upgrade the model using enhanced training data based on the test results.

[0008] The beneficial effects of the technical solution provided by this invention include at least the following: The method of this invention realizes intelligent identification of early prostate cancer lesions by constructing a computer-aided diagnostic system based on convolutional neural networks. It not only utilizes clinical multi-parameter MRI images, but also uses data augmentation technology to enable the model to have a strong generalization ability to identify atypical early lesions.

[0009] The method of this invention uses a novel network that integrates a dual-path architecture and a feature decoupling learning mechanism. This not only automatically identifies lesions but also analyzes their internal key imaging features, significantly improving the reliability and clinical acceptance of diagnostic results. Furthermore, through automatic iterative optimization of the model, the system maintains high accuracy and stability over the long term. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1This is a flowchart of a prostate cancer detection and analysis method based on image analysis provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0013] Example An image analysis-based method for prostate cancer detection and analysis.

[0014] Please refer to Figure 1 This is a flowchart of the prostate cancer detection and analysis method based on image analysis provided in an embodiment of the present invention.

[0015] S1: Acquire the patient's raw MRI images and perform standardized preprocessing, and then perform data augmentation on the processed MRI images; S101 uses the RIAS MIT software tool to convert the original DICOM format MRI image data into NIFTI format image files, and then resamples the NIFTI format image files to an isotropic resolution of 1mm×1mm×1mm. It should be noted that the specific process of MRI raw image standardization is as follows: RIAS MIT is used to perform batch conversion on MRI images to generate intermediate NIFTI format image files, and the current pixel size information (including the voxel width, height, and interslice spacing of the image) is extracted; according to the target isotropic resolution, the resolution of 1 mm is divided by the voxel width, height, and interslice spacing in the current pixel size information to obtain the scaling factors in the three directions; and based on the resampling scaling factor and the preset interpolation algorithm, the intermediate NIFTI format images are resampled in three dimensions to generate the final NIFTI format image files of isotropic voxels (the preset interpolation algorithm is: for intermediate NIFTI images of T2-weighted imaging sequences, the trilinear interpolation algorithm is used for resampling; for intermediate NIFTI images of diffusion-weighted imaging sequences and apparent diffusion coefficient map sequences, the nearest neighbor interpolation algorithm is used for resampling).

[0016] S102, rigidly spatially register the resampled DWI sequence image and ADC sequence image with the T2WI sequence image respectively, so that the multi-sequence images achieve spatial anatomical alignment. It should be noted that the process of generating the registration image is as follows: The resampled T2WI sequence image is used as a fixed reference image, and the resampled DWI and ADC sequence images are used as the moving images to be registered. A rigid transformation model is constructed to determine the spatial transformation relationship between the reference image and the moving image (specifically, a six-DOF three-dimensional spatial transformation model, with six degrees of freedom including three translational degrees of freedom along the X, Y, and Z axes and three rotational degrees of freedom around the X, Y, and Z axes respectively). A mutual information similarity metric is defined to evaluate the degree of alignment between the reference image and the moving image during the transformation process. The parameters of the rigid transformation model are iteratively adjusted using a regularized gradient descent algorithm to maximize the mutual information similarity metric. The moving image is input into the optimized rigid transformation model, and a third-order B-spline interpolation algorithm is used to generate the final registration image aligned with the spatial anatomical position of the reference image.

[0017] The calculation process for mutual information similarity measurement is as follows:

[0018] in, Indicates a reference image; Indicates a moving image; and These represent the gray values ​​of corresponding voxels in the two images, respectively. Denotes the joint probability distribution; and These represent the edge probability distributions in the two images, respectively.

[0019] S103 performs intensity normalization on the spatially registered multi-sequence images and outputs the standardized MRI image data.

[0020] It should be noted that the normalization process is as follows: the N4ITK bias correction algorithm is used to correct the bias of the T2WI sequence images to eliminate low-frequency intensity artifacts caused by the inhomogeneity of the radio frequency coil; the voxel intensity statistical features of the corrected T2WI, DWI, and ADC sequence images are extracted respectively, and intensity scaling transformation is performed on each sequence image to standardize its intensity distribution (for T2WI sequence images, the statistical features are calculated based on the voxel intensity in the automatically segmented prostate gland region; for DWI and ADC sequence images, the statistical features are calculated based on the voxel intensity in the spatial location corresponding to the prostate gland region in the T2WI sequence); the standardized multi-sequence images are associated with patient identifiers and sequence metadata, encapsulated according to a preset data structure, and a standard dataset is generated.

[0021] For each sequence of images in each case, calculate the mean voxel intensity within the selected ROI. and standard deviation ; Using formula Intensity values ​​of all voxels in the entire image Perform the transformation to obtain the normalized intensity value. This makes the intensity distribution of the image sequence approximate a standard normal distribution.

[0022] S2 involves radiologists annotating decoupled signs on preprocessed images and analyzing the accuracy of diagnostic results based on the annotation process. S201: Radiologists use image segmentation software to simultaneously delineate lesions in multiple windows on registered T2WI and DWI sequences, generating overall lesion area annotations. It should be noted that the ITK-SNAP image segmentation software was used in the annotation process.

[0023] S202, In the same annotation task, the physician additionally delineates sub-annotations of suspicious microinvasive areas based on the capsule integrity on T2WI and the signal characteristics on the ADC map; It should be noted that during the annotation process, two radiologists with more than 5 years of experience in the imaging diagnosis of prostate diseases manually delineated the region of interest layer by layer on axial T2WI and DWI, and then copied the DWI ROI to the corresponding apparent diffusion coefficient map. When the patient has multiple lesions, the lesion with the larger diameter is selected. When delineating, the entire lesion should be covered as much as possible but not beyond the edge of the lesion, and the urethra, bleeding and calcification should be avoided as much as possible.

[0024] S203 involves another physician reviewing a randomly selected portion of the annotation results, calculating the intragroup correlation coefficient to assess annotation consistency, and generating an annotation quality report.

[0025] It should be noted that the process for judging the quality of the annotation report is as follows: a portion of the samples are randomly selected from the lesion area samples annotated by one physician to form a quality assessment dataset; another physician performs independent secondary annotation on the quality assessment subset under blinded conditions; and the correlation coefficient within the group is calculated based on the initial annotation results and the annotation results to quantify the consistency among annotators.

[0026] The process of calculating the intragroup correlation coefficient is as follows: the annotation results of each patient are regarded as an assessment object, and the annotation results of two physicians are regarded as two measurements of the object; the average measurement consistency calculation method based on the two-way random effects model is used to obtain the intragroup correlation coefficient value.

[0027] S3, construct a dual-path three-dimensional convolutional neural network model consisting of two encoders and a shared decoder; S301: Construct a network architecture with a dual encoder-shared decoder as the backbone based on 3D U-Net, and set up two independent encoder paths in the network architecture; Of the two encoder paths: the first path is used to process T2WI images and extract anatomical features, and the second path is used to process ADC images and extract functional information features. The shared decoder is used to output a segmentation prediction map of the overall lesion; It should be noted that the number of input channels of the first encoder path is configured to receive anatomical sequence images; the number of input channels of the second encoder path is configured to receive functional sequence images; the two paths have the same hierarchical structure, but the weights are not shared.

[0028] The decoding process of the shared decoder is as follows: the feature map space size is doubled by trilinear interpolation, and then channel concatenation is performed with the feature map from the corresponding layer fusion module; the concatenated features are fused and refined by two 3D convolutional layers (configured the same as the encoder), and the decoder finally outputs 1 channel, followed by a Sigmoid activation function to generate an overall lesion probability segmentation map with the same resolution as the input image.

[0029] S302, adaptive weighted feature fusion modules are set at the encoder bottleneck layer and the decoder jump connection of the backbone network, respectively.

[0030] It should be noted that the specific process of feature fusion is as follows: the feature maps of the same level from the two encoder paths are concatenated by channels, the number of channels is halved by a 1x1x1 convolutional layer, and then the ReLU activation function and another 1x1x1 convolutional layer are passed to finally output feature maps of two channels, which represent the weights of the two paths at this spatial location; the weight map is normalized by the Softmax function, and the fused feature map is output by summing the result after element-wise multiplication with the two original feature maps.

[0031] After the adaptive weighted feature fusion step is completed, a lightweight auxiliary decoder is added to the bottleneck layer to output a fine-grained image feature prediction map, thereby achieving feature decoupling learning. The lightweight auxiliary decoder is configured to receive and process only the fused feature map from the bottleneck layer as input.

[0032] It should be noted that the lightweight auxiliary decoder consists of two simple 3D transposed convolutional layers with a kernel size of 2×2×2, a stride of 2, and 128 and 64 channels respectively. By connecting a 1×1×1 convolutional layer and a sigmoid activation function, it outputs a fine-grained feature probability map with the same spatial size as the main segmentation map.

[0033] Fine-grained probabilistic maps are used to predict “suspicious microinvasive areas” within lesions. The training standard for these maps is simultaneously marked by radiologists when annotating the overall lesion.

[0034] S4. Joint training of a dual-path 3D convolutional neural network model is performed using labeled data to optimize its performance. S401: Pair the preprocessed MRI images into the model and set the corresponding supervision labels; It should be noted that the supervision labels include a global lesion binary mask for the main segmentation task, an image feature sub-region mask for the auxiliary task, and a pixel-level uncertainty weight map.

[0035] In MRI images: high-resolution anatomical structure sequences are input into the main encoder, and functional metabolic sequences are input into the auxiliary encoder; S402, based on the model output and supervision labels, calculate the joint objective function value consisting of weighted segmentation loss, auxiliary task loss and path consistency constraint loss; It should be noted that the calculation process of the joint objective function is as follows:

[0036] In the formula, The segmentation loss corresponding to the main encoder path is represented by a weighted sum of the Dice loss function and the cross-entropy loss function; This represents the auxiliary supervision loss corresponding to the auxiliary encoder path, and its form is determined according to the type of auxiliary task: if the auxiliary task is sub-region segmentation, then the Dice loss is used; if it is uncertainty estimation, then the mean squared error loss is used. The cross-path consistency constraint loss is used to align the high-level feature semantics output by the main encoder and the auxiliary encoder at the bottleneck layer. It is calculated using cosine similarity or mean square error. α, β, and γ are weighting coefficients that balance the various losses, where α > β to ensure that the optimization of the main task is dominant.

[0037] S403 employs a phased parameter update strategy, calculates gradients through backpropagation, and uses an optimizer to adjust all learnable parameters in the dual-path model. The phases include main path-first training and joint fine-tuning training; The process of the main path-first training phase is as follows: freeze all parameters of the auxiliary encoder, and train the main encoder, feature fusion module and decoder only using the weighted segmentation loss. The initial number of training epochs is T1. The joint fine-tuning training phase involves: unfreezing the auxiliary encoder parameters and using the complete joint objective function to train all parameters of the entire model end-to-end.

[0038] It should be noted that in the joint fine-tuning training, the initial learning rate of the auxiliary encoder is set higher than that of the main encoder to promote the supplementation of functional feature information to the main path.

[0039] During phased training: After each training cycle, the model's segmentation performance metric is calculated on an independent validation set. If the validation set segmentation performance metric does not improve within N consecutive training cycles, the optimizer's learning rate is reduced to half of its original value.

[0040] It should be noted that when the performance metric does not exceed the historical best value within M consecutive periods, training is terminated and the best-performing parameter copy is output.

[0041] S5 uses independent data samples to test the performance of the optimized model, and identifies the weak points of the model and the root causes of failures in erroneous cases. S501, construct a model performance evaluation test dataset and input it into the trained model. Based on the prediction confidence obtained by forward propagation of the model on the test set, stratify the test samples and evaluate the model performance in groups to locate its weak points. It should be noted that the model performance evaluation process is as follows: Test set samples are input into the trained dual-path model for forward propagation to obtain the lesion prediction probability map for each sample; by calculating the overall confidence index of each prediction probability map (the index is the average entropy of the probability values ​​of all voxels in the probability map), all test samples are divided into three confidence levels: high, medium, and low; performance evaluation indicators (including the Dessian similarity coefficient and lesion detection rate) are calculated for each confidence level; by comparing and analyzing the performance differences between different confidence levels, the model's performance on the low-confidence sample set is clearly diagnosed as a weak point in the current model.

[0042] S502 uses the spatial contribution score ranking method to analyze the root causes of failure in erroneous cases.

[0043] It should be noted that, for the erroneous prediction cases identified in step S501, the decoupled feature subspace vectors corresponding to the erroneous prediction decision layer are extracted from the dual-path model; the contribution score of each decoupled feature subspace vector to the final erroneous prediction result is calculated using the integral gradient method, and one or more feature subspaces that play a dominant role in the current erroneous decision are identified according to the contribution score; all erroneous cases are classified and statistically analyzed according to the dominant contribution feature subspace, and the root cause of the model's failure is attributed to a systematic defect in the ability to recognize specific types of features (such as global morphological features or local texture features).

[0044] The threshold for dividing the confidence level is dynamically determined based on the cluster analysis results of the overall confidence index or a predefined percentile.

[0045] The evaluation test dataset consists of MRI case image data selected from an independent database; It should be noted that all MRI images must have a clear histopathological diagnosis.

[0046] MRI case image data were screened according to clinical and imaging standards; The screening process includes screening based on image quality and stratification based on case difficulty.

[0047] It should be noted that image quality screening is used to exclude images with severe motion artifacts, metallic artifacts, or excessively low signal-to-noise ratios that affect diagnosis. Case difficulty stratification is based on the imaging manifestations of the lesions, categorized according to the following dimensions: Typicality stratification: divided into typical lesions (clear boundaries, significant signal contrast) and atypical lesions (blurred boundaries, mixed signals); Size stratification: divided into large lesions (maximum diameter > 1.5 cm), medium lesions (1 cm ≤ maximum diameter ≤ 1.5 cm), and small lesions (maximum diameter < 1 cm); Location stratification: based on their location within the prostate gland (e.g., peripheral zone, transition zone, anterior fibromuscular matrix).

[0048] S6, Iteratively upgrade the model using enhanced training data based on the test results; S601 encodes the weak links and root causes of failures into structured generation conditions and synthesizes targeted augmented training data. It should be noted that the common imaging characteristics of the weak links identified in step S501 (such as "low-confidence group samples") are encoded into a conditional vector C1. These characteristics include, but are not limited to, lesion size range, average signal intensity, and boundary clarity. The typical failure sources analyzed in step S502 (such as "misclassifying glandular edges as lesions") are encoded into a conditional vector C2, specifying which image patterns are prone to causing specific misclassifications. By fusing conditional vectors C1 and C2, a structured condition C is formed to guide data generation.

[0049] The fusion process is as follows: C1 and C2 are projected onto a common intermediate dimension d through a fully connected layer, and the projected vectors are normalized to ensure that the values ​​of each dimension have zero mean and unit variance. The normalized vectors C1 and C2 are concatenated according to the following formula, and the concatenated vector is input into a two-layer feedforward neural network composed of ReLU activation functions to output the final structured conditional vector C.

[0050] S602 uses targeted augmentation training data to fine-tune the model in a targeted manner to optimize its weak points.

[0051] It should be noted that the synthesized targeted augmentation data is mixed with some of the original training data to form a fine-tuning training set; the model is trained again on the fine-tuning training set with a low initial learning rate, focusing on monitoring the improvement of the model's performance indicators in the weak links; when the model's loss on the targeted augmentation data converges and the performance on the original data does not show significant degradation, training is stopped and iterative optimization is completed.

[0052] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0053] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0054] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0056] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A method for detecting and analyzing prostate cancer based on image analysis, characterized in that, include: S1: Acquire the patient's raw MRI images and perform standardized preprocessing, and then perform data augmentation on the processed MRI images; S2: Radiologists decouple and annotate the signs in the preprocessed images, and analyze the authenticity of the diagnostic results based on the annotation process; S3: Construct a dual-path 3D convolutional neural network model consisting of two encoders and a shared decoder; S4: Use labeled data to jointly train the dual-path 3D convolutional neural network model and optimize its performance; S5: Use independent data samples to test the performance of the optimized model, and locate the weak points of the model and the root causes of the failure of error cases; S6: Iteratively upgrade the model using enhanced training data based on the test results.

2. The prostate cancer detection and analysis method based on image analysis as described in claim 1, characterized in that, S1 acquires the patient's raw MRI images and performs standardized preprocessing, and then performs data augmentation on the processed MRI images, wherein: S101 uses the RIAS MIT software tool to convert the original DICOM format MRI image data into NIFTI format image files, and then resamples the NIFTI format image files to an isotropic resolution of 1mm×1mm×1mm. S102, rigidly spatially register the resampled DWI sequence image and ADC sequence image with the T2WI sequence image respectively, so that the multi-sequence images achieve spatial anatomical alignment. S103 performs intensity normalization on the spatially registered multi-sequence images and outputs the standardized MRI image data.

3. The prostate cancer detection and analysis method based on image analysis as described in claim 1, characterized in that, S2 involves a radiologist annotating the preprocessed image with decoupled signs and analyzing the accuracy of the diagnostic results based on the annotation process, wherein: S201: Radiologists use image segmentation software to simultaneously delineate lesions in multiple windows on registered T2WI and DWI sequences, generating overall lesion area annotations. S202, In the same annotation task, the physician additionally delineates sub-annotations of suspicious microinvasive areas based on the capsule integrity on T2WI and the signal characteristics on the ADC map; S203 involves another physician reviewing a randomly selected portion of the annotation results, calculating the intragroup correlation coefficient to assess annotation consistency, and generating an annotation quality report.

4. The prostate cancer detection and analysis method based on image analysis as described in claim 1, characterized in that, The S3 constructs a dual-path three-dimensional convolutional neural network model consisting of two encoders and a shared decoder, wherein: S301: Construct a network architecture with a dual encoder-shared decoder as the backbone based on 3D U-Net, and set up two independent encoder paths in the network architecture; Of the two encoder paths: the first path is used to process T2WI images and extract anatomical structural features, and the second path is used to process ADC images and extract functional information features. The shared decoder is used to output a segmentation prediction map of the overall lesion; S302, adaptive weighted feature fusion modules are set at the encoder bottleneck layer and the decoder jump connection of the backbone network, respectively.

5. The prostate cancer detection and analysis method based on image analysis as described in claim 4, characterized in that, The adaptive weighted feature fusion steps are respectively set at the encoder bottleneck layer and the decoder skip connection of the backbone network, wherein: After the adaptive weighted feature fusion step is completed, a lightweight auxiliary decoder is added to the bottleneck layer to output a fine-grained image feature prediction map, thereby achieving feature decoupling learning. The lightweight auxiliary decoder is configured to receive and process only the fused feature map from the bottleneck layer as input.

6. The method for prostate cancer detection and analysis based on image analysis as described in claim 1, characterized in that, S4 utilizes labeled data to jointly train the dual-path 3D convolutional neural network model, optimizing its performance, wherein: S401: Pair the preprocessed MRI images into the model and set the corresponding supervision labels; In the MRI images: high-resolution anatomical structure sequences are input into the main encoder, and functional metabolic sequences are input into the auxiliary encoder; S402, based on the model output and supervision labels, calculate the joint objective function value consisting of weighted segmentation loss, auxiliary task loss and path consistency constraint loss; S403 employs a phased parameter update strategy, calculates gradients through backpropagation, and uses an optimizer to adjust all learnable parameters in the dual-path model. The phases include main path priority training and joint fine-tuning training; The process of the main path priority training phase is as follows: freeze all parameters of the auxiliary encoder, and train the main encoder, feature fusion module and decoder only using weighted segmentation loss, with an initial training cycle number of T1. The process of the joint fine-tuning training phase is as follows: unfreeze the parameters of the auxiliary encoder and use the complete joint objective function to train all parameters of the entire model end-to-end.

7. The prostate cancer detection and analysis method based on image analysis as described in claim 6, characterized in that, The proposed phased parameter update strategy includes: During the training phase: after each training cycle, the segmentation performance index of the model is calculated on an independent validation set. If the segmentation performance index of the validation set does not improve within N consecutive training cycles, the learning rate of the optimizer is reduced to half of its original value.

8. The prostate cancer detection and analysis method based on image analysis as described in claim 1, characterized in that, The S5 method uses independent data samples to test the performance of the optimized model, pinpointing its weaknesses and the root causes of failed error cases. S501, construct a model performance evaluation test dataset and input it into the trained model. Based on the prediction confidence obtained by forward propagation of the model on the test set, stratify the test samples and evaluate the model performance in groups to locate its weak points. S502 uses the spatial contribution score ranking method to analyze the root causes of failure in erroneous cases.

9. The method for prostate cancer detection and analysis based on image analysis as described in claim 5, characterized in that, The model performance evaluation test dataset is input into the trained model, wherein: The evaluation test dataset consists of MRI case image data selected from an independent database; The MRI case image data were screened according to clinical and imaging standards; The screening process includes image quality screening and case difficulty stratification.

10. The method for prostate cancer detection and analysis based on image analysis as described in claim 1, characterized in that, S6 iteratively upgrades the model using enhanced training data based on the test results, wherein: S601 encodes the weak links and root causes of failures into structured generation conditions and synthesizes targeted augmented training data. S602, using the aforementioned targeted enhancement training data, the model is fine-tuned in a targeted manner to optimize its weak points.

Images