Prostate volume intelligent measurement method and system based on multi-modal image

Abstract

The invention discloses an intelligent prostate volume measuring method and system based on a multi-modal image, and belongs to the technical field of medical images. The method comprises the following steps: acquiring an ultrasonic image and a nuclear magnetic resonance image of a prostate area of a patient; performing image preprocessing and registration based on the ultrasonic image and the nuclear magnetic resonance image; according to a registration result, fusing the ultrasonic image and the nuclear magnetic resonance image to generate an enhanced image; based on the enhanced image, utilizing a deep learning model to execute prostate region segmentation; calculating the prostate volume according to the segmentation result; according to the method, a diagnostic report containing a prostate volume measurement result is generated, through multi-modal image fusion, the real-time performance of ultrasonic imaging and the high-resolution characteristic of nuclear magnetic resonance imaging are fully utilized, the limitation of single-modal imaging is effectively overcome, the accuracy of prostate boundary recognition is improved through fusion, and the accuracy of prostate boundary recognition is improved. And the adaptability of the method to different imaging conditions is enhanced.

Description

Technical Field

[0001] The present invention relates to a method and system for intelligent measurement of prostate volume based on multi-modal images, belonging to the field of medical imaging technology. Background Art

[0002] Prostate diseases are common health problems among middle-aged and elderly men. The accurate measurement of prostate volume is crucial for diagnosis and treatment. Traditional methods for measuring prostate volume mainly rely on single-modal medical images, such as ultrasound or magnetic resonance imaging. However, these methods have their respective limitations and it is difficult to balance the accuracy and efficiency of measurement simultaneously.

[0003] Due to its real-time, non-invasive and relatively low-cost characteristics, ultrasound imaging is widely used in prostate volume measurement. However, the quality of ultrasound images is often affected by factors such as operator experience, equipment performance and patient body type, resulting in unsatisfactory repeatability and accuracy of measurement results. Especially in cases where the prostate boundary is blurred or there are calcifications, the error of ultrasound measurement may increase significantly.

[0004] On the other hand, magnetic resonance imaging can provide high-resolution soft tissue contrast, which is beneficial for the accurate identification of the prostate boundary. However, magnetic resonance examination takes a long time, has a high cost, and is not applicable to some patients (such as those with metal implants in the body). In addition, the analysis and volume calculation of magnetic resonance images usually require manual operation by professionals, which is not only time-consuming and laborious, but also may introduce subjective errors.

[0005] In recent years, with the development of computer vision and artificial intelligence technologies, some automated methods for measuring prostate volume have been proposed. These methods usually rely on deep learning algorithms, such as convolutional neural networks, to achieve automatic segmentation and volume calculation of the prostate region. Although these methods have improved the measurement efficiency to a certain extent, there are still some problems. First, most methods still rely on single-modal images and cannot fully utilize the complementary advantages of different imaging technologies. Second, existing deep learning models are often sensitive to changes in image quality and imaging conditions, and their robustness in clinical applications needs to be improved. Finally, these methods usually lack effective mechanisms for validating and optimizing measurement results, making it difficult to ensure reliable measurement results in different situations. Summary of the Invention

[0006] In order to improve the accuracy and efficiency of prostate volume measurement, the present invention provides a method and system for intelligent measurement of prostate volume based on multi-modal images, and the technical solutions are as follows:

[0007] The method for intelligent measurement of prostate volume based on multi-modal images of the present invention includes:

[0008] Step 1: Obtain the ultrasound image and magnetic resonance image of the prostate area of the subject to be measured and perform preprocessing;

[0009] Step 2: Register the preprocessed ultrasound image and magnetic resonance image. The registration process includes:

[0010] Step 21: Construct an adaptive similarity metric function, calculate the similarity between the ultrasound image and the magnetic resonance image. The calculation method is:

[0011] S = α·MI(I 1 , I 2 )+(1-α)·SSIM(I 1 , I 2 )

[0012] where MI represents mutual information, SSIM represents the structural similarity index, α is the weight coefficient, I 1 and I 2 respectively represent the two images to be registered;

[0013] Step 22: Use the adaptive similarity metric function as the objective function and adopt the particle swarm optimization algorithm to search for the optimal registration parameters;

[0014] Step 23: Based on the optimal registration parameters, perform spatial alignment on the ultrasound image and the magnetic resonance image;

[0015] Step 3: According to the registration result, fuse the ultrasound image and the magnetic resonance image to generate an enhanced image;

[0016] Step 4: Based on the enhanced image, use a deep learning model to segment the prostate area;

[0017] Step 5: According to the segmentation result, calculate the prostate volume, perform multiple measurements to obtain a volume measurement sequence, and perform statistical analysis on the volume measurement sequence to obtain an estimated value of the prostate volume.

[0018] Optionally, step 5 further includes volume measurement optimization. The optimization process includes:

[0019] Step 51: Set a volume measurement error threshold and a maximum number of measurements;

[0020] Step 52: Calculate the relative error between two adjacent measurement results in the volume measurement sequence. The calculation formula is:

[0021]

[0022] where V i is the current measurement value, and V i-1 is the previous measurement value;

[0023] Step 53: Stop the measurement when the relative error is less than the volume measurement error threshold or the maximum number of measurements is reached;

[0024] Step 54: Eliminate the outliers;

[0025] Step 55: Calculate the weighted average of the remaining measurement values as the final volume estimate.

[0026] Optionally, in step 4, an improved U-Net model is used as the deep learning segmentation model, ResNet18 is used as the encoder of the U-Net model, and a batch normalization layer is introduced into each residual block of ResNet18; a spatial attention mechanism is introduced into each encoding block and decoding block of the U-Net model.

[0027] Optionally, the method uses a cross-modal loss function to train the deep learning model, and the cross-modal loss function is:

[0028] L = αL dice + βL focal + γL boundary

[0029] where L dice is the Dice loss, L focal is the Focal loss, L boundary is the boundary loss, which is used to improve the accuracy of the segmentation boundary, and α, β, and γ are weight coefficients.

[0030] The intelligent prostate volume measurement system based on multi-modal images of the present invention is characterized in that the system includes:

[0031] An image acquisition module for acquiring ultrasonic images and nuclear magnetic resonance images of the patient's prostate region;

[0032] An image preprocessing module for preprocessing the ultrasonic images and nuclear magnetic resonance images;

[0033] An image registration module for performing spatial alignment of the ultrasonic images and nuclear magnetic resonance images;

[0034] The registration process includes:

[0035] Construct an adaptive similarity metric function to calculate the similarity between the ultrasonic image and the nuclear magnetic resonance image. The calculation method is:

[0036] S = α·MI(I 1 , I 2 ) + (1 - α)·SSIM(I 1 , I 2 )

[0037] Among them, MI represents mutual information, SSIM represents structural similarity index, α is a weight coefficient, and I 1 and I 2 respectively represent two images to be registered;

[0038] Taking the adaptive similarity metric function as the objective function, a particle swarm optimization algorithm is used to search for the optimal registration parameters;

[0039] Based on the optimal registration parameters, spatial alignment of the ultrasound image and the magnetic resonance image is performed;

[0040] An image fusion module is used to fuse the registered ultrasound image and magnetic resonance image to generate an enhanced image;

[0041] An image segmentation module is used to perform prostate region segmentation on the enhanced image;

[0042] A volume calculation module is used to calculate the prostate volume according to the segmentation result, perform multiple measurements to obtain a volume measurement sequence, and perform statistical analysis on the volume measurement sequence to obtain an estimated prostate volume value.

[0043] Optionally, the system further includes a measurement optimization module, and the optimization process includes:

[0044] Setting a volume measurement error threshold and a maximum number of measurements;

[0045] Calculating the relative error between two adjacent measurement results in the volume measurement sequence, and the calculation formula is:

[0046]

[0047] where V i is the current measurement value, and V i-1 is the previous measurement value;

[0048] When the relative error is less than the volume measurement error threshold or the maximum number of measurements is reached, stop the measurement;

[0049] Eliminate the outliers;

[0050] Calculate the weighted average of the remaining measurement values as the final estimated volume value.

[0051] Optionally, the image segmentation module uses an improved U-Net model as the deep learning segmentation model, uses ResNet18 as the encoder of the U-Net model, and introduces a batch normalization layer in each residual block of ResNet18; a spatial attention mechanism is introduced in each encoding block and decoding block of the U-Net model.

[0052] Optionally, the system further includes: a report generation module for generating a diagnostic report including the prostate volume measurement result.

[0053] The present invention provides a prostate volume intelligent measurement device based on multi-modal imaging, including a memory and a processor;

[0054] The memory is used for storing a computer program;

[0055] The processor is used for implementing the prostate volume intelligent measurement method based on multi-modal imaging as described in any one of the above when executing the computer program.

[0056] The present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the prostate volume intelligent measurement method based on multi-modal imaging as described in any one of the above is implemented.

[0057] The beneficial effects of the present invention are as follows:

[0058] First, through multi-modal image fusion, the present invention makes full use of the real-time property of ultrasonic imaging and the high-resolution feature of nuclear magnetic resonance imaging, effectively overcoming the limitations of single-modal imaging. This fusion not only improves the accuracy of prostate boundary recognition but also enhances the adaptability of the method to different imaging conditions.

[0059] Second, the present invention adopts an improved deep learning model, especially introducing a spatial attention mechanism and a multi-scale feature extraction technology, significantly improving the accuracy of prostate region segmentation. This improvement enables the model to better handle individual differences in prostate shape and size, thus improving the accuracy of volume measurement.

[0060] Third, the present invention innovatively introduces an iterative optimization and outlier processing mechanism. Through multiple measurements and statistical analysis, the influence of random errors is effectively reduced. This not only improves the reliability of measurement results but also provides more comprehensive data support for clinicians.

[0061] In addition, the method of the present invention realizes the full-process automation from image acquisition to diagnostic report generation. This greatly improves work efficiency, reduces errors caused by manual operations, and at the same time, the standardized report format helps doctors quickly and accurately interpret the results.

[0062] Finally, the modular design and flexible parameter adjustment mechanism of the system of the present invention endow it with good scalability and adaptability. This means that the method can be easily applied to different medical devices and clinical environments and can be optimized and upgraded according to specific requirements.

[0063] Generally speaking, the prostate volume intelligent measurement method and system based on multi-modal images provided by the present invention not only solve the problems of accuracy, efficiency and reliability in the prior art, but also provide more comprehensive and reliable data support for the diagnosis and treatment of prostate diseases. This innovative solution is expected to significantly improve the diagnosis level and treatment effect of prostate diseases, bringing benefits to the majority of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0065] Figure 1 It is a flowchart of the prostate volume intelligent measurement method based on multi-modal images of the present invention.

[0066] Figure 2 It is a structural diagram of the prostate volume intelligent measurement system based on multi-modal images of the present invention.

[0067] Figure 3 It is a working flowchart of the image registration module of the present invention.

[0068] Figure 4 It is a working flowchart of the measurement optimization module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] To make the objectives, technical solutions and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the drawings.

[0070] Embodiment 1:

[0071] This embodiment provides a prostate volume intelligent measurement method based on multi-modal images, including:

[0072] Step 1: Obtain the ultrasonic image and nuclear magnetic resonance image of the prostate area of the person to be measured and perform preprocessing;

[0073] Step 2: Register the preprocessed ultrasonic image and nuclear magnetic resonance image. The registration process includes:

[0074] Step 21: Construct an adaptive similarity metric function, calculate the similarity between the ultrasonic image and the nuclear magnetic resonance image, and the calculation method is:

[0075] S = α·MI(I 1 , I 2 ) + (1 - α)·SSIM(I 1 , I2 )

[0076] Among them, MI represents mutual information, SSIM represents structural similarity index, α is a weight coefficient, I 1 and I 2 respectively represent two images to be registered;

[0077] Step 22: Taking the adaptive similarity measurement function as the objective function, use the particle swarm optimization algorithm to search for the optimal registration parameters;

[0078] Step 23: Based on the optimal registration parameters, perform spatial alignment on the ultrasound image and the nuclear magnetic resonance image;

[0079] Step 3: According to the registration result, fuse the ultrasound image and the nuclear magnetic resonance image to generate an enhanced image;

[0080] Step 4: Based on the enhanced image, use a deep learning model to segment the prostate region;

[0081] Step 5: According to the segmentation result, calculate the prostate volume, perform multiple measurements, obtain a volume measurement sequence, and perform statistical analysis on the volume measurement sequence to obtain an estimated value of the prostate volume.

[0082] Embodiment 2:

[0083] This embodiment provides a method for intelligent measurement of prostate volume based on multi-modal images. By fusing ultrasound images and nuclear magnetic resonance images, high-precision measurement of prostate volume is achieved. The method includes:

[0084] Step 1: Obtain the ultrasound image and the nuclear magnetic resonance image of the prostate region of the subject to be measured.

[0085] These two imaging modalities each have their own advantages. The ultrasound image has the characteristics of strong real-time performance and low cost, while the nuclear magnetic resonance image provides higher soft tissue contrast. By combining these two modalities, their complementary advantages can be fully utilized.

[0086] Step 2: Preprocess the ultrasound image and the nuclear magnetic resonance image.

[0087] For the nuclear magnetic resonance image, first perform grayscale processing to convert the color image into a grayscale image, which helps to reduce data complexity and improve processing efficiency. Preferably, the weighted average method can be used for grayscale processing, and the weights of the red, green, and blue channels can be set to 0.299, 0.587, and 0.114 respectively. These weight values are determined based on the sensitivity of the human eye to different colors.

[0088] Then the MRI image is subjected to denoising. Denoising is a key step in improving image quality, and a variety of algorithms can be used, such as Gaussian filtering, median filtering or wavelet transform. This embodiment uses an adaptive median filtering algorithm, which can effectively remove salt and pepper noise while retaining image details. The window size of the algorithm can be dynamically adjusted according to the degree of noise, usually between 3x3 and 7x7.

[0089] Normalization is the last step of preprocessing, which scales the image pixel values ​​to a specific range, usually [0, 1] or [-1, 1]. This helps improve the comparability between different images and facilitates the training of deep learning models. Normalization can use the following formula:

[0090]

[0091] Among them, I is the original pixel value, I min and I max are the minimum and maximum pixel values ​​of the image, respectively.

[0092] For the preprocessing of ultrasound images, denoising is also performed, but considering the characteristics of ultrasound images, different denoising algorithms may need to be used. For example, a non-local mean denoising algorithm can be used, which performs well in retaining image edge and texture information. In addition, this embodiment also performs contrast enhancement on ultrasound images to improve the recognizability of the image. Techniques such as histogram equalization or adaptive histogram equalization can be used. In a preferred embodiment, a contrast-limited adaptive histogram equalization (CLAHE) algorithm can be used, which can effectively enhance local contrast while avoiding excessive amplification of noise.

[0093] Step 3: Register the preprocessed ultrasound image and MRI image.

[0094] Step 31: Construct an adaptive similarity measurement function to calculate the similarity between the ultrasound image and the magnetic resonance image.

[0095] The adaptive similarity metric function is the key to achieve accurate registration. Considering the different characteristics of ultrasound and magnetic resonance images, this embodiment adopts a hybrid metric function combining mutual information and structural similarity. The function can be expressed as:

[0096] S = α·MI(I 1 ,I 2 )+(1-α)·SSIM(I 1 ,I 2 )

[0097] Among them, MI represents mutual information, SSIM represents structural similarity index, α is the weight coefficient, and I 1 and I 2They respectively represent two images to be registered. The weight coefficient α can be dynamically adjusted according to the specific application scenario, and its common value range is [0.4, 0.6]. In a preferred embodiment, α can be set to 0.5 to balance the contributions of mutual information and structural similarity.

[0098] Calculate the similarity between the ultrasound image and the magnetic resonance image based on this similarity metric function. During the calculation process, a multi-resolution strategy can be adopted, that is, gradually transitioning from low resolution to high resolution, which helps to improve the registration efficiency and robustness.

[0099] Step 32: Determine the optimal registration parameters.

[0100] Adopt an improved particle swarm optimization algorithm, which can efficiently search for the optimal solution by simulating swarm intelligence behavior. The objective function of the algorithm is the above similarity metric function. The position of the particle represents the registration parameters, including translation, rotation, and scaling, etc. The iterative process of the algorithm can be expressed as:

[0101]

[0102] Among them, among them, and respectively represent the velocity and position of the i-th particle at the t-th iteration, w is the inertia weight, c 1 and c 2 are acceleration constants, r 1 and r 2 are random numbers between [0, 1], is the individual optimal position of the particle, and g t is the global optimal position. In this embodiment, the number of particles is set to 50, and the maximum number of iterations is 100. The inertia weight w can decrease linearly, with an initial value of 0.9 and a final value of 0.4. The acceleration constants c 1 and c 2 can both be set to 2.0. These parameter settings can achieve good convergence performance in most cases.

[0103] Step 33: Based on the optimized optimal registration parameters, perform spatial alignment on the ultrasound image and the magnetic resonance image.

[0104] The alignment process usually involves image interpolation, and a bicubic interpolation algorithm can be adopted, which achieves a good balance between maintaining image quality and computational efficiency.

[0105] Through the above steps, this embodiment realizes the precise registration of the ultrasound image and the magnetic resonance image, laying a foundation for subsequent image fusion and segmentation. This multi-modal registration method not only improves the accuracy of prostate region recognition but also helps to make full use of the complementary information of the two imaging modalities, thereby improving the final volume measurement accuracy.

[0106] Step 4: According to the registration result, fuse the ultrasound image and the magnetic resonance image to generate an enhanced image.

[0107] Step 5: Based on the enhanced image, use a deep learning model to perform prostate region segmentation.

[0108] In this embodiment, an improved U-Net model is adopted as the deep learning segmentation model. The U-Net model is well-known for its excellent image segmentation performance, and the present invention further improves it to meet the special requirements of prostate volume measurement.

[0109] First, in this embodiment, ResNet18 is selected as the encoder of the U-Net. ResNet18 is a deep residual network, and its unique skip connection structure can effectively alleviate the gradient vanishing problem of deep networks, thus allowing the network to be further deepened to extract richer features. In this embodiment, each convolutional layer of ResNet18 is used as the downsampling path of the U-Net, and this design significantly improves the feature extraction ability of the model.

[0110] Preferably, in this embodiment, a batch normalization layer is introduced into each residual block of ResNet18. Batch normalization can accelerate network convergence and improve the generalization ability of the model at the same time. The mathematical expression of batch normalization is as follows:

[0111]

[0112] where x is the input, E[x] and Var[x] are the mean and variance of the mini-batch respectively, ∈ is a very small constant (usually 10 -5 ) to prevent division by zero, and γ and β are learnable scaling and translation parameters.

[0113] Secondly, in this embodiment, a spatial attention mechanism is introduced into the U-Net model. The spatial attention mechanism can help the model focus on important regions in the image, which is especially effective for organs such as the prostate with relatively fixed shapes and positions. The spatial attention module adopted in this embodiment can be expressed as:

[0114] M(F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0115] where F is the input feature map, f 7×7Denotes a 7x7 convolutional operation. AvgPool and MaxPool represent average pooling and max pooling respectively, and σ is the sigmoid activation function. This attention module is inserted into each encoding and decoding block of the U-Net, significantly improving the model's perception ability of the prostate region. This attention module is inserted into each encoding and decoding block of the U-Net, significantly improving the model's perception ability of the prostate region.

[0116] Finally, this embodiment uses a cross-modal loss function for model training. Considering the different characteristics of ultrasound images and magnetic resonance images, a single loss function may be difficult to take into account the characteristics of both modalities. Therefore, this embodiment designs a hybrid loss function:

[0117] L = αL dice + βL focal + γL boundary

[0118] where L dice is the Dice loss, which is used to measure the overall segmentation accuracy; L focal is the Focal loss, which can better handle the class imbalance problem; L boundary is the boundary loss, which is used to improve the accuracy of the segmentation boundary. α, β, and γ are weight coefficients used to balance the contributions of each loss term. In this embodiment, α = 0.5, β = 0.3, and γ = 0.2 are set. This improved U-Net model greatly improves the accuracy of prostate region segmentation, laying a solid foundation for subsequent volume calculation.

[0119] The prostate region segmentation process of the present invention includes multiple steps, aiming to achieve high-precision segmentation results. First, this method performs multi-scale feature extraction on the enhanced image. Multi-scale feature extraction can capture image information at different scales, which is beneficial for processing prostates of different sizes and shapes. In one embodiment of the present invention, dilated convolution can be used to achieve multi-scale feature extraction. The mathematical expression of dilated convolution is as follows:

[0120]

[0121] where F is the input feature map, k is the convolutional kernel, and l is the dilation rate. By setting different dilation rates, the receptive field can be expanded without increasing the number of parameters, thereby capturing multi-scale information.

[0122] Next, this embodiment performs pixel-level classification based on the extracted features. This step uses a fully convolutional network (FCN) structure to map the feature map to a segmentation map of the same size as the input image. During the classification process, the softmax function can be used to convert the feature vector of each pixel into a class probability distribution:

[0123]

[0124] Among them, y i is the predicted category of pixel i, x i is the feature vector of this pixel, and j and k are category indices.

[0125] Then, this embodiment uses morphological prior constraints to optimize the classification result. The prostate has relatively fixed morphological characteristics, and using this prior knowledge can effectively improve the segmentation accuracy. In a preferred embodiment, morphological opening and closing operations can be used to remove small misclassified regions and fill small holes. These operations can be expressed as:

[0126]

[0127] Among them, A is the segmentation result, B is the structuring element, and represent erosion and dilation operations respectively.

[0128] Finally, this embodiment performs edge pruning and boundary fusion to obtain the final segmentation result. Edge pruning aims to refine the segmentation boundary and can be implemented using an Active Contour Model. The Active Contour Model adjusts the segmentation boundary by minimizing the energy function:

[0129] E = ∫ 0 1 [α|C′(s)| 2 + β|C″(s)| 2 + E ext (C(s))]ds

[0130] Among them, C(s) is the parameterized contour curve, α and β are parameters controlling the curve smoothness, and E ext is the external energy term, reflecting the image features.

[0131] Boundary fusion is to integrate segmentation results of different scales or different modalities. A weighted average or voting mechanism can be used to fuse multiple segmentation results to obtain a more robust final segmentation.

[0132] Through this series of carefully designed segmentation steps, the present invention can achieve high-precision segmentation of the prostate region, providing a reliable basis for subsequent volume calculation.

[0133] In addition, this embodiment also includes a data augmentation step, which is a key technology for improving the performance of deep learning models. Data augmentation can expand the scale and diversity of the training dataset, thereby improving the generalization ability and robustness of the model.

[0134] First, in this embodiment, the region of interest of the prostate region is determined based on the segmentation result. This step aims to focus on the most relevant image regions and avoid introducing unnecessary background noise.

[0135] In this embodiment, the region of interest is defined as the minimum bounding rectangle of the prostate contour and is appropriately expanded by a certain proportion (such as 10%) in each direction to include more context information. The determination of the region of interest can be achieved through the following steps:

[0136] 1. Calculate the bounding box of the prostate segmentation mask.

[0137] 2. Expand the bounding box by 10% in each direction.

[0138] 3. Crop the original image to obtain the region of interest.

[0139] Next, the method performs transformations such as rotation, scaling, and translation on the region of interest. These geometric transformations can simulate different imaging angles and scales, increasing the model's adaptability to various situations. In the rotation transformation, the rotation angle can be randomly selected within the range of [-15°, 15°]. The rotation transformation can be achieved through the following matrix:

[0140]

[0141] where θ is the rotation angle.

[0142] For the scaling transformation, the scaling factor can be randomly selected within the range of [0.9, 1.1]. The scaling transformation can be achieved through the following matrix:

[0143]

[0144] where s x and s y are the scaling factors in the x and y directions respectively.

[0145] For the translation transformation, the translation distance can be randomly selected within ±5% of the image size. The translation transformation can be achieved through the following matrix:

[0146]

[0147] where t x and t y are the translation distances in the x and y directions respectively.

[0148] Through these transformations, the method generates enhanced training samples. It should be noted that these transformations should be applied to both the image and the corresponding segmentation mask simultaneously to maintain consistency.

[0149] Finally, the present invention adds the enhanced training samples to the training dataset. Preferably, the size of the original dataset can be expanded by 2 to 3 times. For example, if the original dataset contains 1000 images, then after data augmentation, the training dataset may contain 2000 to 3000 images. This expansion can significantly improve the learning ability and generalization performance of the model.

[0150] Through various data augmentation steps, the present invention can effectively solve the common problem of data scarcity in the field of medical images, improve the performance and reliability of deep learning models, and thus provide a more solid foundation for the accurate measurement of prostate volume.

[0151] Step 6: Calculate the prostate volume according to the segmentation result.

[0152] Step 61: Perform three-dimensional reconstruction on the segmentation result.

[0153] In this embodiment, the Marching Cubes Algorithm is used to implement three-dimensional reconstruction. The basic idea of this algorithm is to traverse each cell in the voxel grid and generate a triangular mesh according to the intersection of the isosurface and the cell vertices. The key steps of the algorithm can be expressed as:

[0154] (x,y,z)=f(x,y,z)isovalue

[0155] where f(x,y,z) is the value of the voxel and isovalue is the isosurface threshold. When the sign of V(x,y,z) changes between the eight vertices of the cell, it means that the isosurface passes through the cell and corresponding triangular patches need to be generated.

[0156] Step 62: Calculate the prostate volume based on the reconstruction result.

[0157] The volume calculation can be achieved by voxelizing the three-dimensional model and then counting the number of voxels.

[0158] Another more accurate method is to use the Gauss divergence theorem to transform the volume calculation into a surface integral:

[0159]

[0160] where r is the vector from a point on the surface to the origin, n is the surface normal vector, and S is the prostate surface.

[0161] Step 63: Perform multiple measurements to obtain a volume measurement sequence.

[0162] To further improve the reliability of the measurement, this embodiment performs multiple measurements to obtain a volume measurement sequence. Multiple measurements help reduce the influence of random errors. Preferably, 5 to 10 measurements can be carried out, and this number can usually achieve a good balance between accuracy and efficiency.

[0163] Step 64: Perform statistical analysis on the volume measurement sequence to obtain a final volume estimate.

[0164] Statistical analysis may include outlier detection and central tendency estimation.

[0165] In outlier detection, this embodiment uses an improved Z-score method:

[0166]

[0167] where x i is a single measurement value, X is the measurement sequence, and MAD is the median absolute deviation. If Z i exceeds a certain threshold (such as 3.5), then this measurement value is considered an outlier.

[0168] In central tendency estimation, considering the possible non-normal distribution, the median or the trimmed mean can be used as the final volume estimate. The calculation formula for the trimmed mean is:

[0169]

[0170] where x (i) is the sorted measurement value, and k is the trimming quantity, usually taking 10% of the number of measurements.

[0171] Through this multi-step and multiple-measurement method, this embodiment can significantly improve the accuracy and reliability of prostate volume measurement, providing strong support for clinical diagnosis and treatment decisions.

[0172] In addition, this embodiment also adopts the following steps for volume measurement optimization.

[0173] (1) Set a volume measurement threshold and a maximum number of measurements.

[0174] The measurement threshold is used to judge whether the measurement result has reached sufficient accuracy, while the maximum number of measurements is used to prevent time waste caused by excessive measurement. In this embodiment, the volume measurement threshold is set to 1% of the previous measurement result, and the maximum number of measurements is set to 15 times. The selection of these parameters is based on a large amount of clinical practice experience and can achieve a good balance between accuracy and efficiency.

[0175] (2) Calculate the difference between two adjacent measurement results.

[0176] The difference calculation uses relative error, and the calculation formula is:

[0177]

[0178] Among them, V i is the current measured value, and V i-1 is the previous measured value.

[0179] (3) When the difference is less than the threshold or the maximum number of measurements is reached, stop the measurement.

[0180] (4) Based on the measurement sequence, perform outlier detection and rejection.

[0181] Outliers may be caused by various factors, such as image noise, segmentation errors, etc. To identify outliers, the modified Z-score method can be used:

[0182]

[0183] Among them, x i is the single measurement value, X is the measurement sequence, and MAD is the median absolute deviation. If |Z i | exceeds a certain threshold (usually taken as 3), then the measurement value is considered an outlier and is rejected.

[0184] (5) Calculate the weighted average of the remaining measurement values as the final volume estimate value.

[0185] The weighted average can consider the reliability of different measurement results, thus obtaining a more accurate estimate. The weights can be determined based on the position in the measurement sequence, and newer measurement results are usually given higher weights. The calculation formula for the weighted average is as follows:

[0186]

[0187] Among them, V i is the i-th valid measurement value, and w i is the corresponding weight. In this embodiment, an exponential decay weight is used: w i = e -λ(n-i) , where λ is the decay rate (such as 0.1) and n is the number of valid measurements.

[0188] Step 7: Generate a diagnostic report containing the prostate volume measurement results.

[0189] Step 71: Extract the prostate volume measurement results and historical measurement data.

[0190] It includes not only the current measurement results but also the patient's past measurement records (if any). The patient's historical data can be retrieved from the electronic health record system and integrated with the current measurement results. Data extraction can use Structured Query Language (SQL) or other appropriate database query methods.

[0191] Step 72: Calculate the volume change trend.

[0192] The volume change trend is crucial for evaluating disease progression or treatment effectiveness, and it can be calculated in the following ways:

[0193] 1. Absolute change: ΔV = V current - V previous ;

[0194] 2. Relative change:

[0195] 3. Annual growth rate:

[0196] where V current is the currently measured volume, and V previous is the volume of the previous measurement.

[0197] Step 93: Generate a preliminary diagnosis recommendation based on preset diagnosis rules.

[0198] The diagnosis rules can be formulated based on the absolute value of the prostate volume, the change rate, and other relevant clinical indicators. For example, the following rules can be adopted:

[0199] If the prostate volume < 30 mL, it is considered of normal size.

[0200] If 30 mL ≤ prostate volume < 50 mL, it is considered mildly enlarged.

[0201] If 50 mL ≤ prostate volume < 80 mL, it is considered moderately enlarged.

[0202] If the prostate volume ≥ 80 mL, it is considered severely enlarged.

[0203] Meanwhile, corresponding recommendations can also be given according to the volume change rate:

[0204] If the annual growth rate < 5%, routine follow-up is recommended.

[0205] If 5% ≤ annual growth rate < 10%, shortening the follow-up interval is recommended.

[0206] If the annual growth rate ≥ 10%, further clinical evaluation and possible treatment intervention are recommended.

[0207] Step 94: Integrate the measurement results, change trends, and diagnostic suggestions to generate a standardized diagnostic report.

[0208] Finally, this embodiment integrates the measurement results, change trends, and diagnostic suggestions to generate a standardized diagnostic report. The content of the report should include, but not be limited to:

[0209] 1. Basic patient information;

[0210] 2. Examination date and method;

[0211] 3. Current prostate volume measurement results;

[0212] 4. Comparison of historical measurement data (if any);

[0213] 5. Analysis of volume change trends;

[0214] 6. Preliminary volume-based diagnosis;

[0215] 7. Doctor's suggestions and follow-up plans;

[0216] Preferably, the diagnostic report can adopt a structured format for easy and quick reading and understanding by doctors. At the same time, visual elements, such as volume change trend charts, can be considered to intuitively display the patient's disease progression.

[0217] Through this comprehensive diagnostic report generation process, the present invention not only provides accurate prostate volume measurement results but also provides valuable diagnostic references and treatment suggestions for clinicians, thus improving the diagnosis and management levels of prostate diseases.

[0218] Embodiment 3:

[0219] This embodiment provides a multi-modal imaging-based intelligent prostate volume measurement system, including:

[0220] An image acquisition module for acquiring ultrasonic images and nuclear magnetic resonance images of the patient's prostate region;

[0221] An image preprocessing module for preprocessing the ultrasonic images and nuclear magnetic resonance images;

[0222] An image registration module for performing spatial alignment of the ultrasonic images and nuclear magnetic resonance images;

[0223] The registration process includes:

[0224] Construct an adaptive similarity metric function to calculate the similarity between the ultrasonic image and the nuclear magnetic resonance image. The calculation method is:

[0225] S = α·MI(I 1 ,I 2 )+(1-α)·SSIM(I1 , I 2 )

[0226] Among them, MI represents mutual information, SSIM represents structural similarity index, α is a weight coefficient, and I 1 and I 2 respectively represent two images to be registered;

[0227] Taking the adaptive similarity measurement function as the objective function, a particle swarm optimization algorithm is used to search for the optimal registration parameters;

[0228] Based on the optimal registration parameters, the ultrasound image and the nuclear magnetic resonance image are spatially aligned;

[0229] An image fusion module is used to fuse the registered ultrasound image and nuclear magnetic resonance image to generate an enhanced image;

[0230] An image segmentation module is used to perform prostate region segmentation on the enhanced image;

[0231] A volume calculation module is used to calculate the prostate volume according to the segmentation result, perform multiple measurements to obtain a volume measurement sequence, and perform statistical analysis on the volume measurement sequence to obtain an estimated prostate volume value.

[0232] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0233] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for intelligent measurement of prostate volume based on multimodal imaging, characterized in that: The method comprises: Step 1: Obtain ultrasound images and magnetic resonance images of the prostate area of ​​the subject and perform preprocessing; Step 2: Register the preprocessed ultrasound image and MRI image. The registration process includes: Step 21: Construct an adaptive similarity measurement function to calculate the similarity between ultrasound images and magnetic resonance images. The calculation method is: S=α·MI(I1,I2)+(1-α)·SSIM(I1,I2) Among them, MI represents mutual information, SSIM represents structural similarity index, α is the weight coefficient, I1 and I2 represent the two images to be registered respectively; Step 22: Using the adaptive similarity measurement function as the objective function, a particle swarm optimization algorithm is used to search for optimal registration parameters; Step 23: Based on the optimal registration parameters, spatially aligning the ultrasound image and the magnetic resonance image; Step 3: According to the registration result, the ultrasound image and the magnetic resonance image are fused to generate an enhanced image; Step 4: Based on the enhanced image, segment the prostate region using a deep learning model; Step 5: Calculate the prostate volume according to the segmentation result, perform multiple measurements to obtain a volume measurement sequence, perform statistical analysis on the volume measurement sequence, and obtain a prostate volume estimation value.

2. The method according to claim 1, characterized in that The step 5 also includes volume measurement optimization, and the optimization process includes: Step 51: Setting the volume measurement error threshold and the maximum number of measurements; Step 52: Calculate the relative error between two adjacent measurement results in the volume measurement sequence, using the following calculation formula: Among them, V i is the current measured value, V i-1 is the last measured value; Step 53: When the relative error is less than the volume measurement error threshold or the maximum number of measurements is reached, stop measuring; Step 54: Eliminate outliers; Step 55: Calculate the weighted average of the remaining measurements as the final volume estimate.

3. The method according to claim 1, characterized in that In step 4, an improved U-Net model is used as a deep learning segmentation model, ResNet18 is used as an encoder of the U-Net model, and a batch normalization layer is introduced in each residual block of ResNet18; a spatial attention mechanism is introduced in each encoding block and decoding block of the U-Net model.

4. The method according to claim 1, characterized in that: The method uses a cross-modal loss function to perform model training on a deep learning model, and the cross-modal loss function is: L=αL dice +βL focal +γL boundary Among them, L dice is the Dice loss, L focal is the focal loss, L boundary is the boundary loss, which is used to improve the accuracy of the segmentation boundary, and α, β and γ are weight coefficients.

5. A prostate volume intelligent measurement system based on multimodal imaging, characterized in that: The system comprises: An image acquisition module, used to acquire ultrasound images and magnetic resonance images of the patient's prostate region; An image preprocessing module, used for preprocessing the ultrasound image and the magnetic resonance image; An image registration module, used for performing spatial alignment of the ultrasound image and the magnetic resonance image; The registration process includes: An adaptive similarity measurement function is constructed to calculate the similarity between ultrasound images and magnetic resonance images. The calculation method is: S=α·MI(I1,I2)+(1-α)·SSIM(I1,I2) Among them, MI represents mutual information, SSIM represents structural similarity index, α is the weight coefficient, I1 and I2 represent the two images to be registered respectively; Taking the adaptive similarity metric function as the objective function, a particle swarm optimization algorithm is used to search for optimal registration parameters; Based on the optimal registration parameters, spatially aligning the ultrasound image and the magnetic resonance image; An image fusion module is used to fuse the registered ultrasound image and magnetic resonance image to generate an enhanced image; An image segmentation module, used for performing prostate region segmentation on the enhanced image; The volume calculation module is used to calculate the prostate volume according to the segmentation result, perform multiple measurements to obtain a volume measurement sequence, and perform statistical analysis on the volume measurement sequence to obtain a prostate volume estimation value.

6. The system according to claim 5, characterized in that The system also includes a measurement optimization module, and the optimization process includes: Set the volume measurement error threshold and maximum number of measurements; The relative error between two adjacent measurement results in the volume measurement sequence is calculated using the following formula: Among them, V i is the current measured value, V i-1 is the last measured value; When the relative error is less than the volume measurement error threshold or the maximum number of measurements is reached, stop measuring; Remove outliers; The weighted average of the remaining measurements was calculated as the final volume estimate.

7. The system according to claim 5, characterized in that The image segmentation module adopts an improved U-Net model as a deep learning segmentation model, adopts ResNet18 as the encoder of the U-Net model, and introduces a batch normalization layer in each residual block of ResNet18; and introduces a spatial attention mechanism in each encoding block and decoding block of the U-Net model.

8. The system according to claim 5, characterized in that The system further comprises: a report generating module, which is used to generate a diagnosis report including the prostate volume measurement result.

9. An intelligent prostate volume measurement device based on multimodal imaging, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the intelligent prostate volume measurement method based on multimodal imaging as described in any one of claims 1 to 4 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the intelligent prostate volume measurement method based on multimodal imaging as described in any one of claims 1 to 4 is implemented.

Images