A prostate tumor classification method and system based on multi-scale training strategy

Abstract

The present invention discloses a method for classifying prostate images by transfer learning based on a multi-scale strategy. The present invention uses transfer learning to improve the performance of backnone network feature extraction and can speed up the training speed. In addition, a multi-scale learning strategy is used to train multiple images with different resolutions on the same network model, so that the model can fully learn the correlation between key areas and the environment. Finally, global average pooling is used to fuse spatial features, and a specific scale is selected, and Sigmoid is used to obtain the final prediction result. The present invention can obtain reliable classification results on prostate cancer nuclear magnetic resonance images.

Description

Technical Field

[0001] The present invention belongs to the field of medical image classification, and specifically relates to a prostate tumor image classification system Background Art

[0002] Prostate cancer is the second most common cancer and the sixth leading cause of cancer death in men worldwide, with approximately 1.276 million new cancer cases and 359,000 deaths in 2018. Due to the aging population alone, the global burden of prostate cancer is expected to grow to 2.3 million cases and 740,000 deaths by 2040.

[0003] In recent years, with the advancement of imaging technology, the development of magnetic resonance imaging (MRI) has changed the previous habit of detecting prostate tumors by performing 10-12 biopsies on the gland to find tumors. The so-called multi-parameter MRI combines T2WI and functional pulse sequences, such as DWI or dynamic contrast-enhanced imaging, and has shown good results in PCa detection. Therefore, mp-MRI is widely used as a non-invasive detection method for pre-biopsy detection of prostate patients. This technology can reduce about 25% of unnecessary biopsies. However, although mp-MRI technology has gradually become a helper for radiologists, 15-30% of clinically significant cancers are still missed by doctors. Therefore, some groups have begun to devote themselves to the research of CAD (computer-aided diagnosis system), one of which is a very important module. Convolutional neural network. In the past, radiologists would do a slice-level analysis from 2D-MRI imaging and finally draw patient-level conclusions, but convolutional neural networks can analyze data at the pixel level, which is theoretically more accurate and objective than manual judgment.

[0004] But in fact, there are several challenges in the implementation of convolutional neural networks on medical datasets:

[0005] 1) Small amount of data. Most medical data sets involve personal privacy issues, making it difficult to collect sufficient data. Some institutions’ data are not publicly released, and a small amount of data will lead to overfitting after repeated training.

[0006] 2) The number of multiple samples is unbalanced. After the network receives information about the imbalanced data, it will cause the inter-class distinction curve to shift, and ultimately output a result that cannot correctly represent the accuracy;

[0007] 3) The data dimension is single, but in reality, doctors often need to consider the surrounding environment information of the lesion when diagnosing tumors;

[0008] 4) The time and hardware cost of training the network are too high. Convolutional neural networks often have millions of parameters and cannot be popularized to low-level devices.

[0009] 5) The network has poor transfer generalization and often performs well on one dataset but poorly on another. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a prostate tumor image classification method based on multi-scale strategy and transfer learning to solve the overfitting problem caused by the small amount of high-quality label data in medical data. A multi-scale training strategy is proposed to simulate the actual situation in which doctors make judgments based on MRI images when diagnosing tumors in combination with environmental information, so that the model can fully learn the relationship between lesions and environmental information to improve the accuracy of the model in classifying benign and malignant tumors. The present invention can obtain reliable classification results on prostate cancer magnetic resonance imaging, effectively assist doctors in tumor judgment, and improve the efficiency of doctors' diagnosis and treatment.

[0011] The object of the present invention is achieved through the following technical solution: a prostate tumor classification method based on a multi-scale training strategy, comprising the following steps:

[0012] S1, obtaining n prostate cancer MRI images, performing voxel spacing correction on the prostate cancer MRI images with different signals of the same patient according to the voxel spacing information, and obtaining images of lesion areas with different signals;

[0013] S2, performing multi-scale segmentation on the lesion area in each prostate cancer MRI image obtained in step S1, and finally obtaining images of lesion areas of different scales in prostate cancer MRI images with different signals;

[0014] S3, preprocessing the prostate cancer MRI images of the same scale obtained in S2 from different signals of the same patient to synthesize RGB pseudo-color images;

[0015] S4. Use the ResNet50V2 pre-trained model to build a transfer learning feature extraction module to obtain the decision maker modules corresponding to the lesion area images of different scales for use in the subsequent multi-scale data rotation training;

[0016] S5, adopting the multi-scale data rotation training method, using the RGB pseudo-color image obtained in S3 to train the transfer learning feature extraction module obtained in S4, and finally selecting a scale containing the richest lesion and environmental information according to the actual situation, and using the decision maker corresponding to this scale to splice the trained feature extraction module as the final prostate tumor classification model;

[0017] S6. Obtain an unlabeled prostate cancer MRI image, input the classification model obtained in S5, and output a judgment result of whether the lesion area in the prostate cancer MRI image is benign or malignant.

[0018] Furthermore, in step S1, the specific method for performing voxel spacing correction on different signal prostate cancer MRI images of the same patient according to voxel spacing information is: obtaining the voxel spacing value a of the image from the data information table, and using bilinear interpolation to expand the original image of size b*c to an image of size (b*a)*(c*a).

[0019] Furthermore, in step S2, the specific method for performing multi-scale segmentation on the lesion area in each prostate cancer MRI image is: six square grayscale images with side lengths of 32 pixels, 40 pixels, 48 ​​pixels, 56 pixels, 64 pixels, and 72 pixels are obtained according to cropping of the lesion center in the MRI image.

[0020] Furthermore, in step S3, the prostate cancer MRI images of the same scale of different signals from the same patient obtained in S2 are preprocessed to synthesize RGB pseudo-color images, and the grayscale images of three signals are extracted from the grayscale images of different scales of each patient's different signals obtained in S2 as the contents of three channels of an RGB image to generate a pseudo-color image.

[0021] Furthermore, the specific method of step S4 is:

[0022] The source network of transfer learning is a ResNet50V2 network pre-trained on the ImageNet dataset; in addition, according to the pseudo-color images of six scales obtained in step S3, six decision maker modules of different scales are prepared for the transfer learning feature extraction module. The parameters of the six decision maker modules are randomly initialized and generated. The architecture of the decision maker modules consists of a Flatten layer connected to the feature extractor module, a dense layer with ReLU activation function and l2 parameters, a dropout layer, and an output layer with sigmoid activation function and 2 neurons. The difference between the decision maker modules of different scales mainly lies in the different parameters of the Flatten layer connected to the feature extractor module.

[0023] Furthermore, in step S5, the specific steps of the multi-scale data rotation training strategy are as follows:

[0024] Step 1: Splice the decision maker module corresponding to 32 pixels onto the feature extraction module, and use the pseudo-color image of the corresponding scale obtained in step S3 to train the current network model. After training, store the current decision maker module in the stack structure for use in the next round.

[0025] Step 2: Regenerate a ResNet50V2 network model with empty parameters and a new input size, inherit the weights of the feature extraction module after the last training, and then splice the decision maker module corresponding to the next scale to the current feature extraction module;

[0026] Step 3: Repeat steps 1 and 2 until all sizes of data are trained on the feature extraction module;

[0027] Step 4: Repeat the training for n rounds for all size data until the accuracy reaches a stable period, and then obtain the trained feature extraction module;

[0028] Step 5: According to the actual situation, select a scale that contains the richest information about lesions and the environment, and use it as the scale corresponding to the final decision maker. Then, use it to splice the feature extraction module obtained in step 4 to obtain the final prostate tumor classification model.

[0029] A prostate tumor classification system based on a multi-scale training strategy, comprising:

[0030] A data acquisition and voxel spacing correction module is used to acquire data and correct the voxel spacing of different signal data;

[0031] A pseudo-color image synthesis module is used to synthesize the prostate cancer grayscale images of the same patient with different signals and the same scale into the RGB channel to obtain a pseudo-color image;

[0032] The transfer learning module includes a multi-scale strategy training module, a feature extraction module, and a decision maker module corresponding to each scale data, which is used to obtain the final prostate tumor classification model;

[0033] In the multi-scale strategy training module, the multi-scale rotation strategy training data is used, and the feature extraction module is repeatedly trained with images of different scales;

[0034] In the feature extraction module, the ResNet50V2 model trained with the ImageNet dataset is used as a pre-training model, and all layers are opened as the main body for training multi-scale data;

[0035] In the decision maker modules corresponding to the data of each scale, after each training of the multi-scale strategy training module is completed, the decision maker module of the original scale will be replaced by the decision maker module corresponding to the scale of the next training.

[0036] In the data acquisition and voxel spacing correction module, the specific method of performing voxel spacing correction on the prostate cancer MRI images with different signals of the same patient according to the voxel spacing information is as follows: the voxel spacing value a of the image is obtained from the data information table, and the image with the original size of b*c is expanded to an image with the size of (b*a)*(c*a) by using the bilinear interpolation method;

[0037] The specific method for multi-scale segmentation of the lesion area in each prostate cancer MRI image is as follows: six square grayscale images with side lengths of 32 pixels, 40 pixels, 48 ​​pixels, 56 pixels, 64 pixels, and 72 pixels are obtained according to the center of the lesion in the MRI image.

[0038] In the pseudo-color image synthesis module, the specific method of preprocessing is: from the grayscale images of each scale of different signals of each patient obtained in S2, the grayscale images of three signals are extracted as the contents of three channels of an RGB image to generate a pseudo-color image.

[0039] In the transfer learning module, the specific process of the multi-scale training strategy is as follows:

[0040] Step 1: Splice the decision maker module corresponding to 32 pixels onto the feature extraction module, and use the pseudo-color image and the pseudo-color image of the corresponding scale obtained in the module to train the current network model. After the training is completed, store the current decision maker module in the stack structure for use in the next round;

[0041] Step 2: Regenerate a ResNet50V2 network model with empty parameters and a new input size, inherit the weights of the feature extraction module after the last training, and then splice the decision maker module corresponding to the next scale to the current feature extraction module;

[0042] Step 3: Repeat steps 1 and 2 until all sizes of data are trained on the feature extraction module;

[0043] Step 4: Repeat the training for n rounds for all size data until the accuracy reaches a stable period, and then obtain the trained feature extraction module;

[0044] Step 5: According to the actual situation, select a scale that contains the richest information about lesions and the environment, and use it as the scale corresponding to the final decision maker. Then use it to splice the feature extraction module obtained in step 4 to obtain the final prostate tumor classification model.

[0045] The advantages of the present invention compared with the prior art are:

[0046] (1) The present invention proposes a prostate tumor classification method and system based on a multi-scale training strategy. The problem of lack of high-quality labeled data often seen in medical imaging data is solved by pre-training the ResNet50V2 model with the ImageNet dataset and the discard layer and regularized l2 parameter in the decision maker. With less data, a model with better performance can be obtained through training more quickly.

[0047] (2) The present invention adopts a multi-scale strategy to train the network model, so that the same feature extraction subject can fully learn the relationship between the lesion area of ​​different scales and the environment. Compared with the prior art that uses interpolation to expand all scale images to the same resolution, according to the characteristics of the convolutional neural network architecture, we use the original size data to avoid the problem of adding unnecessary information to the source data due to interpolation, thereby making the network model more accurate in extracting features;

[0048] (3) The present invention synthesizes the image data of three different signals into the RGB channel to generate a pseudo-color image, thereby solving the overfitting problem caused by the fact that the pre-trained model of the transfer learning network can only input RGB channel color images while the original MRI image is a grayscale image and the same image must be copied to the RGB channel. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of a prostate tumor image classification method according to an embodiment of the present invention;

[0050] Figure 2 A block diagram of a prostate tumor image classification system according to an embodiment of the present invention;

[0051] Figure 3 This is a diagram of the data extraction & processing module architecture in an embodiment of the present invention;

[0052] Figure 4 This is a partial architecture diagram of the decision maker according to an embodiment of the present invention;

[0053] Figure 5 Schematic diagram of a multi-scale training strategy in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the technical solution of the present disclosure, the present disclosure is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0055] like Figure 1 As shown, a prostate tumor image classification method based on multi-scale strategy and transfer learning includes the following steps:

[0056] S1, obtaining n prostate cancer MRI images, performing voxel spacing correction on the prostate cancer MRI images with different signals of the same patient according to the voxel spacing information, and obtaining images of lesion areas with different signals;

[0057] S2, performing multi-scale segmentation on the lesion area in each prostate cancer MRI image obtained in step S1, and finally obtaining images of lesion areas of different scales in prostate cancer MRI images with different signals;

[0058] S3, preprocessing the prostate cancer MRI images of the same scale obtained in S2 from different signals of the same patient to synthesize RGB pseudo-color images. The specific processing flow details are as follows Figure 3 As shown;

[0059] S4. Use the ResNet50V2 pre-trained model to build a transfer learning feature extraction module to obtain the decision maker modules corresponding to the lesion area images of different scales for use in the subsequent multi-scale data rotation training. The specific structure of the decision maker module is as follows: Figure 4 As shown;

[0060] S5, such as Figure 5 As shown, a multi-scale data rotation training method is adopted, and the RGB pseudo-color image obtained in S3 is used to train the transfer learning feature extraction module obtained in S4. Finally, according to the actual situation, a scale containing the richest lesion and environmental information is selected, and the decision maker corresponding to this scale is used to splice the trained feature extraction module as the final prostate tumor classification model;

[0061] S6. Obtain an unlabeled prostate cancer MRI image, input it into the network model obtained in S5, and output a judgment result of whether the lesion area in the prostate cancer MRI image is benign or malignant.

[0062] A prostate tumor classification system based on a multi-scale strategy, including:

[0063] Data acquisition and voxel spacing correction module, used to acquire data and correct the voxel spacing of different signal data;

[0064] A pseudo-color image synthesis module is used to synthesize the prostate cancer grayscale images of the same patient with different signals and the same scale into the RGB channel to obtain a pseudo-color image;

[0065] The transfer learning module includes a multi-scale strategy training module, a feature extraction module, and a decision maker module; in the multi-scale strategy training module, multi-scale rotation strategy training data is used, and images of different scales are repeatedly trained on the network model; in the feature extraction module, a ResNet50V2 model trained with the ImageNet data set is used as a pre-training model, and all layers are opened as the main body for training multi-scale data; in the decision maker module corresponding to each scale data, after each training of the multi-scale strategy training module is completed, the decision maker module of the original scale will be replaced by the decision maker module corresponding to the next training scale.

[0066] Example:

[0067] A prostate tumor image classification method based on multi-scale strategy and transfer learning includes the following steps:

[0068] S1. Extract the original data, correct the voxel spacing, and segment the lesion area at multiple scales to obtain grayscale images of prostate tumors at six different scales, in preparation for subsequent synthetic input and training.

[0069] The original data of step S1 in this embodiment is based on the Prostate-Challenge dataset. The native data is a file in dicom format and needs to be manually extracted according to the provided information table. In addition, the dataset is a set of retrospective prostate MR studies, including T2-weighted (T2W), proton density-weighted (PD-W), dynamic contrast enhancement (DCE) and diffusion-weighted (DW) imaging. Therefore, each patient in the dataset has nine MRI images with different signals. Due to different imaging methods, the voxel spacing between different signal data is different. For example, the pixel spacing of data a is 1, and the pixel spacing of data b is 1.6. Although the sizes of a and b are the same, there is a difference in the actual sizes, so the voxel spacing information is needed to correct all signal data to the same voxel spacing to facilitate the subsequent RGB channel synthesis.

[0070] In this embodiment, in the data extraction stage, the original data of 9 different signals are finally obtained, namely adc, bval, dwi0, dwi1, dwi2, ktrans, Cor, Sag, and Tra. Then, the original data is divided into six square grayscale images with side lengths of 32 pixels, 40 pixels, 48 ​​pixels, 56 pixels, 64 pixels, and 72 pixels with the lesion as the center.

[0071] S2. Among the obtained data of different signals, select data of three different signals to form pseudo-color data, so that the processed data can meet the input requirements of the network.

[0072] In this embodiment, according to the statistics of the patient's lesion areas, the lesions are distributed in four different tissue areas, namely, SV area (spermatic vesicle zone), AS area (anterior fibromyoma zone), TZ area (transition zone), and PZ area (periglandular zone). According to experience, for different areas, the combination of different signal data may bring different effects. Therefore, after many experiments, the data of the three signals bval, dwi2, and ktrans are selected to synthesize the pseudo-color images of the PZ and AS areas, and the data of the three signals bval, dwi2, and tra are selected to synthesize the pseudo-color image of the TZ area. As for the SV area, it was not included in the experiment due to missing data.

[0073] The transfer learning technology used in this embodiment requires that the input data size of the non-convolutional layer of the network model built based on this technology should be consistent with that of the source network. For example, most of the transfer learning models are pre-trained on the ImageNet dataset. The images on this database are all RGB three-channel color images with a size of 224*224 pixels. Therefore, it is necessary to ensure that the number of channels of the input data is consistent with the input data of the original model. The original data, regardless of the signal, is a single-channel grayscale image. If you want to input the network, you must copy the single channel to other channels to form a three-channel image. However, this operation will bring redundancy and eventually cause the network model to overfit. Therefore, in this embodiment, images of the same size of different signals are selected to be fused into the RGB channel to form a pseudo-color image, which solves the potential overfitting problem well and expands the feature dimension accepted by the network model in a single training, thereby improving the network accuracy.

[0074] S3. Use the ResNet50V2 pre-trained model to build a transfer learning feature extraction module, and splice the feature extraction module with the decision maker module of the corresponding scale for lesion area images of different scales.

[0075] The source network migrated by the present invention is a ResNet50V2 model pre-trained on the ImageNet dataset. The network adopts a novel and practical skip network structure. This residual structure solves the degradation problem caused by too deep neural networks. The network has been successfully applied in many fields, and some studies have also shown that the model has better results in transfer learning than other networks.

[0076] Since the transfer learning technology is used, the input data size of the pre-trained model ResNet50V2 is a 224*224 RGB three-channel image. Although the feature extraction layer part of the original model (before Flatten) can input data of different sizes, the final decision layer module must correspond to each input size, so the present invention adopts a stack structure to save the corresponding decision layer module.

[0077] For the architecture of the decision-making layer module, Figure 2 As shown, the present invention uses the Flatten layer to flatten the output of the feature extraction layer, and also serves as the connection part between the feature extraction layer and the decision layer. It is then connected to a dense layer with 256 units and an l2 parameter (parameter value is 0.1), and then connected to a drop layer with a drop rate of 0.5 to alleviate the overfitting phenomenon caused by a small amount of data. Finally, it is connected to a dense layer with one unit, activated by a sigmoid function, and used as the final output layer to output the prediction result.

[0078] S4. Use a multi-scale training strategy to train the model obtained in S3, and finally obtain a fixed-scale model, which is then spliced ​​with the corresponding decision-making layer module as the final benign and malignant tumor classification model.

[0079] In this embodiment, the Adam optimizer is uniformly used to optimize the network during the training phase, and only the learning rate is set to 1e-4, and other parameters use the default parameters in the Keras framework. The loss function is a binary cross entropy function (binary_crossentropy). In addition, the early-stopping method is used to avoid overtraining. The early-stopping method automatically monitors the indicators in the network training. If the indicators are not optimized within a certain period of time, the early-stopping method will think that the network model has been optimized and stop continuing training. This method is very good to avoid the overfitting problem caused by overtraining.

[0080] In this embodiment, a multi-scale training strategy is adopted. The training process is as follows: Figure 4 Specifically, first, use the output size of 32*32 to obtain the input layer from the Keras library, then splice the pre-trained model of the same size obtained from Keras, and finally splice the decision maker of the corresponding size, and then use 32*32 data for the first training. Because the early stopping method is used, there is no need to manually set the appropriate number of epochs. The upper limit of the number of epochs set in this embodiment is 100. After the first training is completed, the stack structure is used to save the current decision layer for subsequent training, and a ResNet50V2 model framework with an input size of 40*40 (next size) with empty parameters is obtained from Keras, and the weight parameters of the feature extraction layer after the first training are inherited, and then the decision layer of the corresponding size is spliced ​​to start the second training. This cycle is repeated until the training of data of size 72*72 is completed, and then the ResNet50V2 model framework of size 32*32 input is obtained again, and the decision layer of the corresponding size is taken out from the previously saved stack structure and spliced ​​with the pre-trained model to start the second round of training. This cycle is repeated until the second round of training of data of size 72*72 is completed, and then the ResNet50V2 model framework of size 64*64 is obtained again, and the corresponding decision maker in the stack structure is used to splice it and then the last training is carried out. The final result is the prostate tumor benign and malignant classification system proposed in the present invention.

[0081] In this example, two rounds of training were conducted at all scales, followed by the final training (64*64). The final 64*64 scale was chosen because in reality, instrument imaging will have a certain deviation due to the patient's body shape, so the 72*72 scale was selected as an error buffer. After separate testing of all size data, the 64*64 size was found to be the best size.

[0082] S5. Obtain an unlabeled prostate cancer MRI image, synthesize corresponding color data according to different regions, put it into the prediction model obtained in step S4, perform auxiliary judgment, and output the judgment result of the benign or malignant lesion area in the prostate cancer MRI image.

[0083] The above contents are further detailed descriptions of the present invention in combination with specific / preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can also make several substitutions or modifications to these described embodiments without departing from the concept of the present invention, and these substitutions or modifications should be regarded as belonging to the protection scope of the present invention.

[0084] Parts of the present invention that are not described in detail belong to the well-known technologies of those skilled in the art.

Claims

1. A prostate tumor classification method based on a multi-scale training strategy, It is characterized in that The following steps are involved: S1, obtaining n prostate cancer MRI images, performing voxel spacing correction on the prostate cancer MRI images with different signals of the same patient according to the voxel spacing information, and obtaining images of lesion areas with different signals; S2, performing multi-scale segmentation on the lesion area in each prostate cancer MRI image obtained in step S1, and finally obtaining images of lesion areas of different scales in prostate cancer MRI images with different signals; S3, preprocessing the prostate cancer MRI images of the same scale obtained in S2 from different signals of the same patient to synthesize RGB pseudo-color images; S4. Use the ResNet50V2 pre-trained model to build a transfer learning feature extraction module to obtain the decision maker modules corresponding to the lesion area images of different scales for use in the subsequent multi-scale data rotation training; S5, adopting the multi-scale data rotation training method, using the RGB pseudo-color image obtained in S3 to train the transfer learning feature extraction module obtained in S4, and finally selecting a scale containing the richest lesion and environmental information according to the actual situation, and using the decision maker corresponding to this scale to splice the trained feature extraction module as the final prostate tumor classification model; S6, obtaining an unlabeled prostate cancer MRI image, inputting the classification model obtained in S5, and outputting a judgment result of whether the lesion area in the prostate cancer MRI image is benign or malignant; In step S1, the specific method of performing voxel spacing correction on the prostate cancer MRI images with different signals of the same patient according to the voxel spacing information is as follows: obtaining the voxel spacing value a of the image from the data information table, and using the bilinear interpolation method to expand the image with the original size of b*c to an image with the size of (b*a)*(c*a); In step S5, the specific steps of the multi-scale data rotation training strategy are as follows: Step 1: Splice the decision maker module corresponding to 32 pixels onto the feature extraction module, and use the pseudo-color image of the corresponding scale obtained in step S3 to train the current network model. After training, store the current decision maker module in the stack structure for use in the next round. Step 2: Regenerate a ResNet50V2 network model with empty parameters and a new input size, inherit the weights of the feature extraction module after the last training, and then splice the decision maker module corresponding to the next scale to the current feature extraction module; Step 3: Repeat steps 1 and 2 until all sizes of data are trained on the feature extraction module; Step 4: Repeat the training for n rounds for all size data until the accuracy reaches a stable period, and then obtain the trained feature extraction module; Step 5: According to the actual situation, select a scale that contains the richest information about lesions and the environment, and use it as the scale corresponding to the final decision maker. Then, use it to splice the feature extraction module obtained in step 4 to obtain the final prostate tumor classification model.

2. A prostate tumor classification method based on a multi-scale training strategy according to claim 1, It is characterized in that In step S2, the specific method for multi-scale segmentation of the lesion area in each prostate cancer MRI image is: six square grayscale images with side lengths of 32 pixels, 40 pixels, 48 ​​pixels, 56 pixels, 64 pixels, and 72 pixels are obtained according to the lesion center in the MRI image.

3. A prostate tumor classification method based on a multi-scale training strategy according to claim 2, It is characterized in that In step S3, the prostate cancer MRI images of the same scale of different signals from the same patient obtained in S2 are preprocessed to synthesize RGB pseudo-color images. Among the grayscale images of different scales of different signals of each patient obtained in S2, the grayscale images of three signals are extracted as the contents of three channels of an RGB image to generate a pseudo-color image.

4. A prostate tumor classification method based on a multi-scale training strategy according to claim 3, It is characterized in that The specific method of step S4 is: The source network of transfer learning is a ResNet50V2 network pre-trained on the ImageNet dataset; in addition, according to the six scales of pseudo-color images obtained in step S3, six decision maker modules corresponding to different scales are prepared for the transfer learning feature extraction module. The parameters of the six decision maker modules are randomly initialized and generated. The architecture of the decision maker modules consists of a Flatten layer connected to the feature extractor module, a dense layer with ReLU activation function and l2 parameters, a discard layer, and an output layer with sigmoid activation function and 2 neurons; the difference between decision maker modules of different scales mainly lies in the different parameters of the Flatten layer connected to the feature extractor module.

5. A prostate tumor classification system based on a multi-scale training strategy, It is characterized in that include: Data acquisition and voxel spacing correction module, used to acquire data and correct the voxel spacing of different signal data; A pseudo-color image synthesis module is used to synthesize the prostate cancer grayscale images of the same patient with different signals and the same scale into the RGB channel to obtain a pseudo-color image; The transfer learning module includes a multi-scale strategy training module, a feature extraction module, and a decision maker module corresponding to each scale data, which is used to obtain the final prostate tumor classification model; In the multi-scale strategy training module, the multi-scale rotation strategy training data is used, and the feature extraction module is repeatedly trained with images of different scales; In the feature extraction module, the ResNet50V2 model trained with the ImageNet dataset is used as a pre-training model, and all layers are opened as the main body for training multi-scale data; In the decision maker modules corresponding to the data of each scale, after each training of the multi-scale strategy training module is completed, the decision maker module of the original scale will be replaced by the decision maker module corresponding to the scale of the next training; In the data acquisition and voxel spacing correction module, the specific method of performing voxel spacing correction on the prostate cancer MRI images with different signals of the same patient according to the voxel spacing information is as follows: the voxel spacing value a of the image is obtained from the data information table, and the image with the original size of b*c is expanded to an image with the size of (b*a)*(c*a) by using the bilinear interpolation method; The specific method of multi-scale segmentation of the lesion area in each prostate cancer MRI image is as follows: six square grayscale images with side lengths of 32 pixels, 40 pixels, 48 ​​pixels, 56 pixels, 64 pixels, and 72 pixels are obtained by cropping the lesion center in the MRI image; In the transfer learning module, the specific process of the multi-scale training strategy is as follows: Step 1: Splice the decision maker module corresponding to 32 pixels onto the feature extraction module, and use the pseudo-color image and the pseudo-color image of the corresponding scale obtained in the module to train the current network model. After the training is completed, store the current decision maker module in the stack structure for use in the next round; Step 2: Regenerate a ResNet50V2 network model with empty parameters and a new input size, inherit the weights of the feature extraction module after the last training, and then splice the decision maker module corresponding to the next scale to the current feature extraction module; Step 3: Repeat steps 1 and 2 until all sizes of data are trained on the feature extraction module; Step 4: Repeat the training for n rounds for all size data until the accuracy reaches a stable period, and then obtain the trained feature extraction module; Step 5: According to the actual situation, select a scale that contains the richest information about lesions and the environment, and use it as the scale corresponding to the final decision maker. Then use it to splice the feature extraction module obtained in step 4 to obtain the final prostate tumor classification model.

6. A prostate tumor classification system based on a multi-scale training strategy according to claim 5, It is characterized in that In the pseudo-color image synthesis module, the specific method of preprocessing is: in the grayscale images of each scale of different signals of each patient obtained in the data acquisition and voxel spacing correction module, the grayscale images of three signals are extracted as the contents of the three channels of an RGB image to generate a pseudo-color image.

Images