Prostate subregion medical image segmentation method based on multi-disturbance consistency learning

Abstract

The invention provides a prostate sub-region medical image segmentation method based on multi-disturbance consistency learning, and the method comprises the steps: obtaining a prostate MRI image data set, and carrying out the preprocessing of the data set; determining slice indexes of the labeled image and the unlabeled image; performing strong and weak data enhancement processing on each prostate MRI image; inputting the strongly enhanced prostate MRI image into a 3D network for feature extraction, and inputting the weakly enhanced prostate MRI image into a 2D network for learning to obtain prediction results of the two networks; performing data division on a prediction result according to the determined slice index, and respectively inputting an LSUS strategy and a USUS strategy; and respectively inputting a new prostate image and a new label generated by the LSUS and a new prostate image generated by the USUS into a 3D network and a 2D network according to different data enhancement modes for further learning to a specified number of iterations to obtain a trained model, and obtaining a prostate subregion segmentation result according to the trained model.

Description

Technical Field

[0001] The present invention relates to the field of medical information, and particularly to a method for segmenting prostate sub-region medical images based on multi-disturbance consistency learning. Background Art

[0002] Prostate MRI (Magnetic Resonance Imaging) images are widely used in the treatment of prostate diseases because they can display clear textures and contours of the prostate, and play an important role in clinical diagnosis and treatment. However, due to the influence of operators and different imaging modalities and parameters, there are differences in the expression of target structures in prostate MRI images. Therefore, the segmentation of prostate MRI images has attracted more and more researchers and become a hot issue in this field. By segmenting the sub-regions of prostate MRI images, the ability to identify and observe the prostate structure can be improved, the doctor's film reading and diagnosis can be accelerated, the state of the prostate can be more comprehensively understood, and more accurate information can be provided for clinical decision-making.

[0003] When using deep learning to segment medical images, accurate pixel-level annotation is crucial. However, it is very time-consuming and laborious to make accurate annotations in medical images, and it is necessary to have experts for annotation to ensure high-quality annotations. Therefore, using some methods to reduce the annotation workload without affecting the segmentation effect has broad application value.

[0004] Currently, semi-supervised learning provides a potential solution. Semi-supervised learning can make full use of unlabeled data and labeled data to improve model performance. Existing semi-supervised learning methods mainly combine co-training and consistency learning for method design. Co-training aims to improve learning efficiency and performance through the collaborative effect among multiple models. For example, by using data from different perspectives of a medical 3D dataset to train multiple models, each model learns on different feature subsets and enhances the learning process through mutual supervision. During training, labeled data is used to train the model, while unlabeled data is used to assist model training by generating pseudo-labels. The output of the model is optimized through mutual feedback and pseudo-labels, thereby maximizing the utilization of unlabeled data. Co-training is particularly suitable for scenarios where labeled data is scarce and can effectively reduce the dependence on labeled data. Consistency learning aims to improve the robustness and generalization ability of the model by forcing the model to maintain consistent predictions in the face of perturbations. Combining co-training and consistency learning can give full play to the advantages of multiple models and unlabeled data and further improve model performance. In this combination, co-training strengthens the model's learning through mutual supervision and pseudo-label generation among multiple models, while consistency learning forces the model to maintain consistent predictions under different perturbations, thereby enhancing the model's robustness. The combined method can share information among multiple perspectives and ensure the stability of the model under various perturbation conditions through consistency learning, significantly improving the model's generalization ability and robustness. Especially in scenarios where labeled data is scarce, this combination can not only effectively reduce the dependence on labeled data but also improve the model's performance in diverse tasks and scenarios.

[0005] However, prostate sub-region segmentation poses unique challenges compared to other medical image segmentation tasks, and one of the most prominent features is the blurred boundary. This blurriness stems from the tissue transition between sub-regions and the unobvious difference in signal intensity, and is further interfered by individual patient differences, lesions such as tumors or hyperplasia, which further blur the boundaries of normal tissues. In addition, the signal contrast of prostate sub-regions in MRI images is low, and the boundaries between different regions (such as the peripheral zone, transitional zone, and central zone) are difficult to accurately distinguish. Therefore, existing semi-supervised learning is not fully applicable to the sub-region segmentation of prostate images. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a method for prostate sub-region medical image segmentation based on multi-perturbation consistency learning to improve the above problems.

[0007] The present invention provides a method for prostate sub-region medical image segmentation based on multi-perturbation consistency learning, which includes:

[0008] S101. Obtain a prostate MRI image dataset, and preprocess the prostate MRI images in the prostate MRI image dataset; wherein, the prostate MRI images include labeled images and unlabeled images;

[0009] S102. Determine the slice indices of the labeled images and unlabeled images;

[0010] S103. Perform two types of data augmentation, strong and weak, on each prostate MRI image to obtain strongly augmented prostate MRI images and weakly augmented prostate MRI images;

[0011] S104. Input the strongly augmented prostate MRI images into a 3D network for feature extraction, and input the weakly augmented prostate MRI images into a 2D network for learning to obtain the prediction results output by the 3D network and the prediction results output by the 2D network;

[0012] S105. Divide the prediction results output by the 3D network and the 2D network according to the determined slice indices, and input them into the LSUS strategy and the USUS strategy respectively; wherein, the LSUS strategy is used to generate new reliable prostate images and new labels; the USUS strategy is used to generate new reliable prostate images;

[0013] S106. Input the new prostate images and new labels generated by LSUS, and the new prostate images generated by USUS, according to different data augmentation methods, into the 3D network and the 2D network for further learning, and repeatedly load and train to update the loss weights until the specified number of iterations to obtain a trained model, so as to obtain the prostate sub-region segmentation result according to the trained model.

[0014] Preferably, in step S101, preprocessing the prostate MRI images includes:

[0015] Resample the prostate MRI images by trilinear interpolation according to the preset target size and original size to make them isotropic;

[0016] Normalize the resampled prostate MRI images by the maximum-minimum normalization method to scale their gray value range to between 0 and 1.

[0017] Preferably, in step S102, determining the slice indices of the labeled images and unlabeled images includes:

[0018] Traverse all slices of each image in the prostate MRI image dataset;

[0019] If the current slice has a corresponding expert label, determine its index as the labeled slice index, otherwise determine its index as the unlabeled slice index.

[0020] Preferably, in step S103,

[0021] The strong data augmentation includes rotation, flipping, color jittering, and occlusion data;

[0022] The weak data augmentation includes rotation and flipping.

[0023] Preferably, in step S104,

[0024] The prediction results Predictions output by the 3D network 3D and the prediction results Predictions output by the 2D network 2D are as follows:

[0025]

[0026] where X represents the original data, and are the data augmentation operations of strong augmentation and weak augmentation respectively.

[0027] Preferably, in step S105, after the two networks respectively output the prediction results, first update the loss weights, and then use the determined labeled slice indices and unlabeled slice indices to divide the prediction results of the two networks into four types: strongly augmented labeled prediction slices S_A, weakly augmented labeled prediction slices W_A, strongly augmented unlabeled prediction slices S_U, and weakly augmented unlabeled prediction slices W_U; the strongly augmented labeled prediction slices and the weakly augmented labeled prediction slices are input into the LSUS strategy to generate new reliable prostate training images NewI and new corresponding labels NewGT, and the strongly augmented unlabeled prediction slices and the weakly augmented unlabeled prediction slices are input into the USUS strategy to generate new reliable prostate images NewI_U; the process is shown in the following formula:

[0028] S_A = Predictions 3D (labeled)

[0029] W_A = Predictions 2D (labeled)

[0030] S_U = Predictions 3D (Unlabeled)

[0031] W_U = Predictions 2D (Unlabeled)

[0032] NewI, NewGT = LSUS(S A , W A )

[0033] NewI_U = USUS(S_U, W_U).

[0034] Preferably, in step S106, when updating the loss weight, the loss function of the model includes the supervised loss L of the labeled data sup and the unsupervised loss L of the unlabeled data unsup , defined as:

[0035]

[0036] λ represents the weight for balancing different losses, which gradually increases with the increase of the number of iterations on the basis of Gaussian warming, expressed as: iter represents the current number of iterations, iter total represents the total number of iterations.

[0037] Preferably, the supervised loss includes and represents calculating the supervised loss of the original samples of the labeled slices, represents calculating the supervised loss of the newly generated samples of the labeled slices:

[0038]

[0039] wherein and are the cross - entropy loss and the dice loss respectively;

[0040]

[0041] The unsupervised loss includes and composed of; represents calculating the cross - pseudo - labeling loss of the original samples of the unlabeled slices, represents calculating the cross - pseudo - labeling loss of the newly generated samples of the unlabeled slices:

[0042]

[0043]

[0044] Preferably, four indexes, namely the Dice similarity coefficient, the Jaccard similarity coefficient, HD95 and ASD, are used to evaluate the performance of the model.

[0045] Compared with the prior art, the present invention has at least the following advantages:

[0046] 1. This embodiment proposes a semi-supervised medical image segmentation (SSMIS) framework, which not only solves the problem of the failure of multi-perturbation consistency learning, but also further improves the segmentation performance by introducing multi-dimensional information complementarity.

[0047] 2. For labeled data, this embodiment designs a multi-patch bidirectional displacement strategy (LSUS) guided by regions of interest (ROIs). For unlabeled data, this embodiment designs a multi-patch bidirectional displacement strategy (USUS) guided by JS divergence.

[0048] In summary, the present invention effectively solves the problem of the scarcity of labeled medical images, improves the segmentation accuracy and generalization ability of the model, and provides important support for the early diagnosis and treatment of prostate cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Schematic flow chart of the prostate sub-region medical image segmentation method based on multi-perturbation consistency learning provided by the embodiment of the present invention;

[0050] Figure 2 Schematic principle diagram of the prostate sub-region medical image segmentation method based on multi-perturbation consistency learning provided by the embodiment of the present invention;

[0051] Figure 3 Overall architecture diagram of the model provided by the embodiment of the present invention;

[0052] Figure 4 Basic flow chart of using the LSUS strategy for labeled slices;

[0053] Figure 5 Basic flow chart of using the USUS strategy for unlabeled slices;

[0054] Figure 6 Principle block diagram of the LSUS strategy;

[0055] Figure 7 Principle block diagram of the USUS strategy;

[0056] Figure 8 Comparison effect diagram of the prostate sub-region dataset with other methods;

[0057] Figure 9 Comparison effect diagram of the LA dataset with other methods;

[0058] Figure 10 Comparison effect diagram of the ACDC dataset with other methods. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this specification.

[0060] As Figures 1 - 3 shown, an embodiment of the present invention provides a method for segmenting prostate sub-region medical images based on multi-disturbance consistency learning, which includes:

[0061] S101, obtaining a prostate MRI image dataset, and preprocessing the prostate MRI images in the prostate MRI image dataset; wherein, the prostate MRI images include labeled images and unlabeled images.

[0062] Specifically, during preprocessing:

[0063] First, obtain a prostate MRI image training dataset. According to the size and resolution information of the images, use the trilinear interpolation algorithm to resample the images so that all images have a unified size and resolution, and obtain the resampled training dataset.

[0064] Among them, the resampling formula is as follows:

[0065]

[0066] Among them:

[0067] old size[i]: The size of the input image in the i-th dimension.

[0068] old spacing[i]: The voxel spacing of the input image in the i-th dimension.

[0069] new size[i]: The size of the target image in the i-th dimension.

[0070] This formula ensures that the physical size of the image remains consistent after resampling while adapting to the new size specifications.

[0071] Then, traverse each image in the resampled training dataset, and obtain the gray values of all pixel points in the image for normalization.

[0072] The normalization formula is as follows:

[0073]

[0074] Among them:

[0075] x: The value of the original data point.

[0076] x min : The minimum value in the original data.

[0077] x max : The maximum value in the original data.

[0078] a: The lower limit value of the normalized interval.

[0079] b: The upper limit value of the normalized interval.

[0080] x': The normalized data value, within the range of [a, b].

[0081] This formula is applicable to scaling the data to any specified range.

[0082] Repeat the above steps until all images in the prostate MRI image dataset are traversed to obtain the preprocessed dataset.

[0083] S102. Determine the slice indices of the labeled images and unlabeled images.

[0084] Specifically, according to the annotation information of the images in the preprocessed dataset, determine whether each slice of the prostate MRI image is annotated. The labeled image slices and unlabeled image slices are numbered using an indexing method to obtain the labeled image slice indices and unlabeled image slice indices.

[0085] S103. Perform two types of data augmentation processing, strong and weak, on each prostate MRI image to obtain strongly augmented prostate MRI images and weakly augmented prostate MRI images.

[0086] Among them, in order to create input perturbations, in this embodiment, two augmentation strategies are used to perform different data augmentation processing on the same data to expand the data, including strong data augmentation and weak data augmentation. The strong data augmentation includes rotation, flipping, color jittering, and occlusion data operations; the weak data augmentation includes rotation and flipping operations.

[0087] The following specific data operation methods:

[0088] 1. Rotation

[0089] The rotation operation is achieved by performing an angle transformation on the image, and the formula is as follows:

[0090] I'(x', y') = I(x, y) (3)

[0091] Where:

[0092] (x', y') are the position coordinates after rotation.

[0093] (x, y) is the position coordinate before rotation and satisfies the following relationship:

[0094]

[0095] (x c , y c ) is the coordinate of the center point of the image.

[0096] θ is the rotation angle.

[0097] 2. Flipping

[0098] Image flipping usually includes horizontal flipping and vertical flipping, and their formulas are as follows:

[0099] 2.1 Horizontal Flip:

[0100] I'(x, y) = I(w - x - 1, y) (5)

[0101] w is the width of the image.

[0102] 2.1 Vertical Flip:

[0103] I'(x, y) = I(x, h - y - 1) (6)

[0104] h is the height of the image.

[0105] 3. Color Jitter

[0106] Color jitter is achieved by adjusting brightness, contrast, saturation, and hue. The specific formulas are as follows:

[0107] 3.1 Brightness Adjustment:

[0108] I'(x, y) = I(x, y) · α (7)

[0109] α is the brightness factor, usually a random value α ∈ [α min , α max

[0110] 3.2 Contrast Adjustment:

[0111] I'(x, y) = (I(x, y) - μ) · β (8)

[0112] μ is the mean value of the image;

[0113] β is the contrast factor, randomly taking values β ∈ [β min , β max .

[0114] ​3.3 Saturation adjustment (only for color images): After converting to the HSV space, adjust the saturation channel S':

[0115] S' = S · γ (9)

[0116] γ is the saturation factor, randomly taking values γ ∈ [γ min , γ max .

[0117] 3.4 Hue adjustment: Adjust the hue channel H' in the HSV space:

[0118] H' = (H + δ) mod 360 (10)

[0119] δ is the random offset.

[0120] 4 Random cropping

[0121] The random cropping operation is achieved by replacing a part of the pixels in the image with a fixed value (such as zero or random noise). The formula is as follows:

[0122]

[0123] Where:

[0124] OcclusionRegion is the cropped area, usually defined by a randomly generated rectangle, and its size and position can be described by the following parameters:

[0125] OcclusionRegion = {(x, y) | x 1 ≤ x ≤ x 2 , y 1 ≤ y ≤ y 2}

[0126] (x 1 , y 1 ) and (x 2 , y 2 ) are the coordinates of the upper left corner and the lower right corner of the cropping rectangle.

[0127] S104, input the strongly enhanced prostate MRI image into a 3D network for feature extraction, input the weakly enhanced prostate MRI image into a 2D network for learning, and obtain the prediction results output by the 3D network and the prediction results output by the 2D network.

[0128] Among them, the prediction results Predictions 3D output by the 3D network and the prediction results Predictions 2D output by the 2D network are as follows:

[0129]

[0130] Among them, X represents the original data, and are data augmentation operations of strong augmentation and weak augmentation respectively.

[0131] S105. Divide the prediction results output by the 3D network and the 2D network according to the determined slice indices, and input them into the LSUS strategy and the USUS strategy respectively; among them, the LSUS strategy is used to generate new reliable prostate images and new annotations; the USUS strategy is used to generate new reliable prostate images.

[0132] After the two networks output the prediction results respectively, first update the loss weights, and then use the labeled slice indices and unlabeled slice indices determined in the second step to divide the prediction results of the two networks into four types: strongly augmented labeled prediction slices (S_A), weakly augmented labeled prediction slices (W_A), strongly augmented unlabeled prediction slices S_U, and weakly augmented unlabeled prediction slices (W_U) as shown in the figure. The strongly augmented labeled prediction slices and the weakly augmented labeled prediction slices are input into the newly generated reliable prostate training image (NewI) and the new corresponding label (NewGT) in the LSUS strategy. The strongly augmented unlabeled prediction slices and the weakly augmented unlabeled prediction slices are input into the newly generated reliable prostate image (NewI_U) in the USUS strategy. The process can be as Figures 4 - 7 shown in Formulas 15 to 20:

[0133] S_A = Predictions 3D (labeled) (15)

[0134] W_A = Predictions 2D (labeled) (16)

[0135] S_U = Predictions 3D (Unlabeled) (17)

[0136] W_U = Predictions 2D (Unlabeled) (18)

[0137] NewI, NewGT = LSUS(S A , W A ) (19)

[0138] NewI_U = USUS(S_U, W_U) (20)

[0139] Among them, LSUS and USUS are different slice utilization strategies respectively.

[0140] The principle of the LSUS strategy is as Figure 4 andFigure 6 As shown, in the LSUS strategy, by comparing the true annotations and prediction results of the annotated images, reliable prediction results with high confidence are selected to generate new reliable prostate images and corresponding new annotations.

[0141] Specifically, for the prediction results generated by the 2D network, which include both an image and a label. For the image part, it is cropped into K predicted patch blocks. For the label part, it is cropped into K patch blocks and the mean value of each patch block of the label is calculated (if non-zero, it is considered an interesting region). Then, the position of the interesting region patch of the label is mapped to the relative position of the predicted patch block of the image, and the mean confidence of the interesting region of the label corresponding to the predicted patch block is calculated. After sorting according to the mean confidence from large to small, the patch blocks corresponding to the last K / 2 mean confidence values are selected.

[0142] The prediction results generated by the 3D network are processed in the same way. Then, the patch blocks generated by the 2D network and the patch blocks generated by the 3D network are displaced bidirectionally. By randomly displacing a certain pixel distance in the horizontal and vertical directions, new reliable prostate images and new annotations are generated, namely strongly enhanced annotated prediction slices (S_A) and weakly enhanced annotated prediction slices (W_A).

[0143] The principle of the USUS strategy is as Figure 5 and Figure 7 shown. In the USUS strategy, by calculating the confidence of the prediction results of the unannotated images, a part of the slices with the highest confidence is selected to generate new reliable unannotated prostate images.

[0144] Specifically, first, the prediction results of the unannotated slices (including the prediction results of the 3D network and the 2D network) are respectively cropped into K patch blocks, the mean confidence of each patch block is calculated, and the patch blocks are sorted in descending order. The first Sqrt(K) patch blocks with the largest mean confidence are selected as reliable unannotated regions. The selected reliable unannotated patch blocks are displaced bidirectionally. By randomly displacing a certain pixel distance in the horizontal and vertical directions, new reliable unannotated prostate images are generated, namely strongly enhanced unannotated prediction slices S_U and weakly enhanced unannotated prediction slices W_U.

[0145] The new images generated by these strategies are respectively input into the 3D network and the 2D network according to different enhancement methods, and finally the prediction results are output and the weights are updated again.

[0146] S106. Input the new prostate images and new annotations generated by LSUS, as well as the new prostate images generated by USUS, into a 3D network and a 2D network respectively for further learning according to different data augmentation methods. Repeat loading, training, and updating the loss weights until the specified number of iterations is reached to obtain a trained model, and use the trained model to obtain the prostate sub-region segmentation result.

[0147] Among them, when updating the loss weights, the loss function of the model includes the supervised loss L of the labeled data sup and the unsupervised loss L of the unlabeled data unsup , which is defined as:

[0148]

[0149] λ represents the weight for balancing different losses, which gradually increases with the increase in the number of iterations based on Gaussian warming and can be expressed as: iter represents the current iteration, and iter total represents the total number of iterations.

[0150] The supervised loss consists of and represents calculating the supervised loss of the original samples of the labeled slices, represents calculating the supervised loss of the newly generated samples of the labeled slices. The specific calculation formula is:

[0151]

[0152] Among them and are the cross-entropy loss and the dice loss respectively.

[0153]

[0154] The unsupervised loss consists of and . represents calculating the cross pseudo-labeling loss of the original samples of the unlabeled slices, represents calculating the cross pseudo-labeling loss of the newly generated samples of the unlabeled slices. The specific calculation formula is as follows:

[0155]

[0156] In summary, a 3D network and a 2D network are used to co-train the augmented training data. Through multi-dimensional feature extraction and fusion, the model's segmentation ability for prostate sub-regions is improved. During the training process, various perturbation methods are introduced, including data augmentation operations such as intensity transformation and noise addition to the input images, as well as random perturbation of the network structure, to improve the model's robustness and generalization ability. By consistency learning, the prediction results under different perturbations are constrained to be consistent, alleviating the supervision bias problem introduced by multiple perturbations, and obtaining the final prostate sub-region segmentation model, which improves the segmentation accuracy of prostate sub-regions.

[0157] To illustrate the performance of this embodiment, the application of the present invention will be described below with a practical example.

[0158] Experimental dataset

[0159] The ProstateX dataset is a four-class prostate 3D MRI image dataset, containing 98 pairs of prostate 3D MRI images and expert handcrafted segmentation masks, from the 2017 ProstateX Challenge. For this dataset, we applied min-max normalization [0,1] and resampled the images to a voxel size of 224×224×20. In this embodiment, 70 images were randomly selected for training, 8 images for validation, and 20 images for testing.

[0160] The LA dataset comes from the left atrium segmentation challenge. This dataset contains 100 left atrium 3D MRI images and their corresponding expert hand-segmented masks, with an isotropic resolution of 0.625×0.625×0.625mm. In this embodiment, 80 images were selected for training and 20 images for testing.

[0161] The ACDC dataset is a medical dataset containing three classes of cardiac 3D MRI images, from the 2017 ACDC Challenge. This dataset contains cardiac 3D MRI images and their corresponding expert handcrafted segmentation masks. In this embodiment, min-max normalization [0,1] was performed on 100 patient MRI images in the training data, and the images were resampled to a voxel size of 224×224×32. Subsequently, 70 images were randomly selected for training, 10 images for validation, and 20 images for testing.

[0162] Evaluation metrics

[0163] In this embodiment, four objective evaluation metrics are adopted to quantitatively evaluate the performance of the proposed method, namely the Dice Similarity Coefficient, the Jaccard Similarity Coefficient, HD95, and ASD. Among them, the Dice Similarity Coefficient and the Jaccard Similarity Coefficient are positive metrics, that is, the higher the metric value, the better the performance of the model; while HD95 and ASD are negative metrics, that is, the smaller the metric value, the better the performance of the model.

[0164] Dice Similarity Coefficient

[0165] Dice is a statistical tool used to quantify the similarity between two samples and is used to evaluate the performance of models in image segmentation and other forms of binary classification problems. The Dice coefficient is particularly suitable for dealing with the situation of imbalance in the number of positive and negative samples in the dataset because it takes into account both true positives and false positives. The formula:

[0166]

[0167] where X and Y represent two sets, the predicted segmentation region and the true segmentation region respectively. |X∩Y| represents the number of elements in the intersection of X and Y, while |X| and |Y| represent the number of elements in X and Y respectively. In this way, the Dice coefficient can measure the similarity between two sets, and its value ranges from 0 (completely dissimilar) to 1 (completely identical). In fields such as medical image processing, the Dice coefficient is one of the important metrics for evaluating the performance of segmentation algorithms.

[0168] Jaccard Similarity Coefficient

[0169] It is a metric for measuring the similarity between two sets and is used to compare the similarity and difference between sample sets. Especially when evaluating the performance of binary classification problems (such as image segmentation tasks), it is a commonly used method to calculate the overlap degree between the predicted region and the true region. Especially in computer vision tasks such as object detection and segmentation, IoU is an important evaluation metric. The formula:

[0170]

[0171] where X and Y represent two sets, |X∩Y| represents the number of elements in the intersection of X and Y, and X∪Y represents the number of elements in the union of X and Y. The Jaccard Similarity Coefficient can measure the similarity degree between two sets, and its value ranges from 0 (completely dissimilar) to 1 (completely identical).

[0172] 95% Hausdorff Distance (HD95)

[0173] The Hausdorff distance is a measure of the maximum distance between two point sets. In medical image segmentation, HD95 is commonly used to measure the shape difference between the predicted boundary and the ground truth boundary. HD95 refers to the 95th percentile among the Hausdorff distances between all pairs of points, which is used to reduce the influence of extreme values. The formula is:

[0174] H(A,B)=max(□(A,B),□(B,A)) (28)

[0175] where and HD95 is the 95th percentile among these distances.

[0176] Average Surface Distance (ASD)

[0177] ASD evaluates the segmentation accuracy by measuring the average distance between the segmented and reference surfaces. The smaller the value, the better the segmentation result. The formula is:

[0178]

[0179] where A represents the set of segmented objects, B represents the set of reference ground truth objects, S(A) and S(B) represent the sets of surface points of objects A and B respectively, d(a,S(B)) represents the Euclidean distance from surface point a to the nearest point in set S(B),

[0180] and d(b,S(A)) represents the Euclidean distance from surface point b to the nearest point in set S(A).

[0181] Experimental Design and Result Analysis

[0182] Table 1

[0183]

[0184] As shown above, Table 1 shows the comparison of the method of this embodiment with the state-of-the-art semi-supervised medical image segmentation methods on the ProstateX dataset using 20% and 50% labeled slice ratios. The method of this embodiment is compared with CPS, CT, 3D2DCT, and ABD. As Figure 8As shown, the results indicate that when using 20% labeled slices, the method of this embodiment is significantly superior to the state-of-the-art ABD method in multiple evaluation metrics: the DSC increases from 68.60 to 69.28, and the Jaccard increases from 54.94 to 55.54. In terms of the edge-sensitive metrics, the HD95 decreases from 5.25 mm to 4.19 mm, and the ASD decreases from 1.55 mm to 1.34 mm. When using 50% labeled slices, the method of this embodiment outperforms the fully supervised UNet method in all major evaluation metrics. Specifically, the DSC increases from 71.66 to 72.14, and the Jaccard increases from 58.46 to 58.12. In terms of edge sensitivity, the HD95 decreases from 4.63 mm to 3.62 mm, and the ASD decreases from 1.24 mm to 1.03 mm. With these objective evaluation metrics and visualization results, these results show that the framework of this embodiment can effectively suppress the influence of mixed perturbations and improve consistency learning. The method of this embodiment consistently outperforms the existing state-of-the-art methods in 20% and 50% labeled slices, especially showing significant advantages in edge accuracy and consistency learning. This demonstrates the strong potential of the method of this embodiment in the application of prostate medical image segmentation tasks, and excellent segmentation performance can be achieved even in the presence of scarce labeled data.

[0185] Table 2

[0186]

[0187] Table 2 verifies that additional comparative experiments of the method of this embodiment were conducted on the LA dataset. The comparison results show that under the conditions of 5% and 10% labeled slices, the method of this embodiment performs superiorly compared to the current state-of-the-art segmentation methods (including DTC, SS-Net, MCF, BCP, and PMT). Under the condition of 10% labeled slices, the method of this embodiment achieved a Dice similarity coefficient (DSC) of 91.33 and a Jaccard coefficient of 83.82, exceeding methods such as PMT and showing higher segmentation accuracy. In addition, the method of this embodiment also performs excellently in the edge sensitivity metrics, where the HD95 is 3.82 mm and the ASD is 1.14 mm, superior to all comparative methods. This shows that the model of this embodiment has obvious advantages in accurately capturing boundary details. When the labeling ratio is reduced to 5%, the method of this embodiment still shows strong performance, although some metrics are slightly lower than PMT. Among them, the DSC and Jaccard reach 89.45 and 81.29 respectively, while the edge sensitivity metrics HD95 and ASD decrease to 5.86 mm and 2.02 mm respectively, but still superior to most comparative methods. As Figure 9 shown, it can be seen from the 3D visualization of the segmentation results that the model of this embodiment is closer to the real label results compared to other models.

[0188] Table 3

[0189]

[0190] Table 3 verifies that the method of this embodiment has conducted additional comparative experiments on the ACDC dataset, and compared the method of this embodiment with other latest semi-supervised medical image segmentation methods. The comparison results show that under the conditions of 16% and 25% labeled slices, the method of this embodiment performs better than the CPS, CT, 3D2DCT, and ABD methods. Compared with the current state-of-the-art ABD method, the method of this embodiment has improvements in multiple metrics. Among them, the Dice similarity coefficient (DSC) increases from 89.75 to 89.83, and the Jaccard coefficient increases from 81.81 to 81.98. In terms of the edge sensitivity metric, the method of this embodiment is particularly prominent: HD95 decreases from 3.17 mm to 1.85 mm, and ASD decreases from 1.21 mm to 0.70 mm. Further verification results show that in the comparison with the ABD method, the method of this embodiment continues to be superior in metrics. DSC increases from 90.26 to 90.35, and the Jaccard coefficient increases from 82.63 to 82.83. At the same time, it is still significantly better than ABD in terms of edge sensitivity: HD95 decreases from 2.38 mm to 1.76 mm, and ASD decreases from 0.74 mm to 0.70 mm. Combining these objective evaluation metrics and visualization results (such as Figure 10 shown in), it shows that the model of this embodiment has obvious advantages in edge accuracy. In particular, the method of this embodiment can more accurately capture the boundary details of each region of the heart, achieving significant improvements in boundary accuracy and precision, which is of great significance for cardiac image segmentation.

[0191] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these changes and modifications.

Claims

1. A prostate subregion medical image segmentation method based on multi-perturbation consistency learning, characterized in that: include: S101, obtaining a prostate MRI image dataset, and preprocessing the prostate MRI images in the prostate MRI image dataset; wherein the prostate MRI images include labeled images and unlabeled images; S102, determining the slice indexes of the labeled image and the unlabeled image; S103, performing strong and weak data enhancement processing on each prostate MRI image to obtain a strongly enhanced prostate MRI image and a weakly enhanced prostate MRI image; S104, inputting the strongly enhanced prostate MRI image into a 3D network for feature extraction, and inputting the weakly enhanced prostate MRI image into a 2D network for learning, to obtain a prediction result output by the 3D network and a prediction result output by the 2D network; S105, dividing the prediction results output by the 3D network and the 2D network into data according to the determined slice index, and inputting them into the LSUS strategy and the USUS strategy respectively; wherein the LSUS strategy is used to generate a new reliable prostate image and a new annotation; and the USUS strategy is used to generate a new reliable prostate image; S106, the new prostate image and new annotation generated by LSUS, and the new prostate image generated by USUS, are respectively input into the 3D network and the 2D network for further learning according to different data enhancement methods, and the training loss weight is repeatedly loaded and updated to the specified number of iterations to obtain a trained model, so as to obtain the prostate sub-region segmentation result according to the trained model.

2. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to claim 1, characterized in that: In step S101, the prostate MRI image is preprocessed, including: According to a preset target size and an original size, the prostate MRI image is resampled by a trilinear interpolation method to make it isotropic; The maximum and minimum normalization method was used to normalize the resampled prostate MRI images and scale their grayscale values ​​to between 0 and 1.

3. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to claim 1, characterized in that: In step S102, determining the slice indexes of the labeled image and the unlabeled image includes: Traversing all slices of each image in the prostate MRI image dataset; If the current slice has a corresponding expert annotation, its index is determined as the labeled slice index, otherwise its index is determined as the unlabeled slice index.

4. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to claim 1, characterized in that: In step S103, The strong data enhancement includes rotation, flipping, color jittering and occlusion data; The weak data enhancement includes rotation and flipping.

5. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to claim 1, characterized in that: In step S104, Predictions output by the 3D network 3D And the prediction results output by the 2D network Predictions 2D As shown below: and Among them, X represents the original data, and They are data enhancement operations of strong enhancement and weak enhancement respectively.

6. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to claim 5, characterized in that: In step S105, after the two networks output prediction results respectively, the loss weight is updated first, and then the prediction results of the two networks are divided into four types using the determined labeled slice index and non-labeled slice index, namely, strongly enhanced labeled prediction slice S_A, weakly enhanced labeled prediction slice W_A, strongly enhanced non-labeled prediction slice S_U and weakly enhanced non-labeled prediction slice W_U; the strongly enhanced labeled prediction slice and the weakly enhanced labeled prediction slice are input into the LSUS strategy to generate a reliable prostate training image NewI and a new corresponding label NewGT, and the strongly enhanced non-labeled prediction slice and the weakly enhanced non-labeled prediction slice are input into the USUS strategy to generate a new reliable prostate image NewI_U; the process is shown in the following formula: S_A=Predictions 3D (labeled) W_A=Predictions 2D (labeled) S_U=Predictions 3D (Unlabeled) W_U=Predictions 2D (Unlabeled) NewI,NewGT=LSUS(S A ,W A ) NewI_U=USUS(S_U,W_U).

7. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to claim 6, characterized in that: In step S106, when updating the loss weight, the loss function of the model Including the supervision loss L of the labeled data sup and the unsupervised loss L for unlabeled data unsup , defined as: λ represents the weight for balancing different losses. On the basis of Gaussian warming, it gradually increases with the number of iterations, expressed as: iter indicates the current number of iterations. total Represents the total iterations.

8. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to claim 7, characterized in that: Supervised losses include and represents the supervision loss of the original sample with labeled slices. Represents the supervised loss for calculating newly generated samples with labeled slices: in and They are cross entropy loss and dice loss respectively; Unsupervised loss include and composition; represents the cross pseudo-annotation loss of the original sample of the unlabeled slice. Represents the cross-pseudo-annotation loss for newly generated samples of unlabeled slices:

9. The prostate sub-region medical image segmentation method based on multi-perturbation consistency learning according to any one of claims 1 to 8, characterized in that: The performance of the model was evaluated using four indicators: Dice similarity coefficient, Jaccard similarity coefficient, HD95 and ASD.

Images