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

Abstract

The application provides a prostate sub-region medical image segmentation method based on multi-perturbation consistency learning, comprising: acquiring a prostate MRI image dataset and performing preprocessing; determining slice indexes of labeled images and unlabeled images; performing strong and weak data enhancement processing on each prostate MRI image; inputting the strong enhanced prostate MRI image into a 3D network for feature extraction, and inputting the weak enhanced prostate MRI image into a 2D network for learning to obtain prediction results of the two networks; dividing the prediction results according to the determined slice indexes, and inputting the prediction results into LSUS strategy and USUS strategy respectively; inputting new prostate images and new labels generated by LSUS and new prostate images generated by USUS into the 3D network and the 2D network respectively according to different data enhancement modes for further learning to a specified number of iterations to obtain a trained model, and obtaining a prostate sub-region segmentation result according to the trained model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information, and particularly relates to a prostate sub-region medical image segmentation method based on multi-perturbation consistency learning. BACKGROUND

[0002] Prostate MRI (magnetic resonance imaging) images are widely used in the treatment of prostate diseases because they can show clear textures and contours of the prostate. They 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 has become a hot issue in this field. By segmenting the sub-regions of the prostate MRI image, the recognition and observation of the prostate structure can be improved, the doctor's reading and diagnosis can be accelerated, and a more comprehensive understanding of the state of the prostate can be obtained, providing more accurate information for clinical decision-making.

[0003] In order to use deep learning to segment medical images, accurate pixel-level annotation is essential. However, accurate annotation in medical images is very time-consuming and labor-intensive, and must be annotated by experts to ensure high-quality annotation. Therefore, it has wide application value to use some methods to reduce the workload of annotation without affecting the segmentation effect.

[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 synergy of multiple models, such as training multiple models using different views of medical 3D data sets, each model learns on different feature subsets, and enhances the learning process through mutual supervision. In 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 use 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 when facing perturbations. Combining co-training with consistency learning can fully leverage the advantages of multiple models and unlabeled data, further improving model performance. In this combination, co-training strengthens model learning through mutual supervision and pseudo-label generation between multiple models, while consistency learning forces the model to maintain consistent predictions under different perturbations, thereby enhancing the robustness of the model. The combined method can share information between multiple views, while ensuring the stability of the model under various perturbation conditions through consistency learning, significantly improving the generalization ability and robustness of the model. In particular, in scenarios where labeled data is scarce, this combination not only effectively reduces the dependence on labeled data, but also improves the performance of the model in diverse tasks and scenarios.

[0005] However, prostate sub-region segmentation has unique challenges compared to other medical image segmentation tasks, with fuzzy boundaries being one of the most prominent features. This fuzziness is due to the lack of clear signal intensity differences between sub-regions and the interference of individual patient differences, tumors or hyperplasia, further blurring the boundaries of normal tissue. In addition, the signal contrast of prostate sub-regions in MRI images is low, and the boundaries of different regions (such as the peripheral zone, transition zone and central zone) are difficult to accurately distinguish, so existing semi-supervised learning cannot be fully applied to prostate image sub-region segmentation. SUMMARY

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

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

[0008] S101, obtain a prostate MRI image dataset and pre-process a prostate MRI image in the prostate MRI image dataset; wherein the prostate MRI image comprises a labeled image and an unlabeled image;

[0009] S102, determine slice indexes of the labeled image and the unlabeled image;

[0010] S103, perform strong and weak data augmentation processing on each prostate MRI image to obtain a strong augmented prostate MRI image and a weak augmented prostate MRI image;

[0011] S104, input the strong augmented prostate MRI image into a 3D network for feature extraction and input the weak augmented 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;

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

[0013] S106, input the new prostate image and new labels generated by the LSUS and the new prostate image generated by the USUS into the 3D network and the 2D network respectively for further learning according to different data augmentation methods, repeat loading training update loss weights to a specified iteration number, and obtain a trained model to obtain a prostate sub-region segmentation result according to the trained model.

[0014] Preferably, in step S101, the pre-processing of the prostate MRI image comprises:

[0015] 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;

[0016] The resampled prostate MRI image is normalized by a maximum and minimum value normalization method to scale its gray value range to 0 to 1.

[0017] Preferably, in step S102, the determination of the slice indexes of the labeled image and the unlabeled image comprises:

[0018] All slices of each image in the prostate MRI image dataset are traversed;

[0019] If the current slice has a corresponding expert label, its index is determined as a labeled slice index, otherwise its index is determined as an 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] Predictions output by the 3D network 3D and Predictions output by the 2D network 2D As follows:

[0025]

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

[0027] Preferably, in step S105, after the two networks output the predictions, respectively, the loss weight is updated first, and then the predictions of the two networks are divided into four kinds, strong augmentation labeled prediction slice S_A, weak augmentation labeled prediction slice W_A, strong augmentation unlabeled prediction slice S_U and weak augmentation unlabeled prediction slice W_U, using the determined labeled slice index and unlabeled slice index; the strong augmentation labeled prediction slice and the weak augmentation labeled prediction slice input the reliable prostate training image NewI and the new corresponding label NewGT generated in the LSUS strategy to generate a new reliable prostate image 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 a supervised loss L sup and an unsupervised loss L unsup of unlabeled data, defined as:

[0035]

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

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

[0038]

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

[0040]

[0041] The unsupervised loss includes and ; represents the cross-pseudo-label loss of calculating the original sample of the unlabeled slice, represents the cross-pseudo-label loss of calculating the newly generated sample of the unlabeled slice:

[0042]

[0043]

[0044] Preferably, the performance of the model is evaluated by four indexes of Dice similarity coefficient, Jaccard similarity coefficient, HD95 and ASD.

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

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

[0047] 2. For labeled data, the embodiment designs a multi-patch bidirectional displacement strategy (LSUS) based on a region of interest (ROI) guide. For unlabeled data, the embodiment designs a multi-patch bidirectional displacement strategy (USUS) based on a JS divergence guide.

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

[0049] Figure 1 The flowchart of the prostate sub-region medical image segmentation method based on multi-perturbation consistency learning provided by the embodiment of the application is shown.

[0050] Figure 2 The principle diagram of the prostate sub-region medical image segmentation method based on multi-perturbation consistency learning provided by the embodiment of the application is shown.

[0051] Figure 3 The overall architecture diagram of the model provided by the embodiment of the application is shown.

[0052] Figure 4 The basic flowchart of the LSUS strategy for labeled slices is shown.

[0053] Figure 5 The basic flowchart of the USUS strategy for unlabeled slices is shown.

[0054] Figure 6 The principle diagram of the LSUS strategy is shown.

[0055] Figure 7 The principle diagram of the USUS strategy is shown.

[0056] Figure 8 The comparison effect diagram of the prostate sub-region data set and other methods is shown.

[0057] Figure 9 The comparison effect diagram of the LA data set and other methods is shown.

[0058] Figure 10 The comparison effect diagram of the ACDC data set and other methods is shown. DETAILED DESCRIPTION

[0059] In order for those skilled in the art to better understand the technical solutions in the specification, the technical solutions in the specification will be clearly and completely described below in conjunction with the drawings in the embodiments of the specification. Obviously, the described embodiments are only part of the embodiments of the specification, not all. Based on the embodiments in the specification, all other embodiments obtained by those of ordinary skill in the art without creative labor should be within the scope of protection of the specification.

[0060] As shown in Figures 1-3 The embodiment of the application provides a prostate sub-region medical image segmentation method based on multi-disturbance consistency learning, which comprises:

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

[0062] Specifically, during preprocessing:

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

[0064] The resampling formula is as follows:

[0065]

[0066] Wherein:

[0067] old size[i]: the size of the input image in the ith dimension.

[0068] old spacing[i]: the voxel spacing of the input image in the ith dimension.

[0069] new size[i]: the size of the target image in the ith dimension.

[0070] This formula ensures that the image maintains its physical size after resampling, while adapting to the new size specification.

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

[0072] The normalization formula is as follows:

[0073]

[0074] Wherein:

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

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

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

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

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

[0080] x': normalized data value, ranging in [a, b].

[0081] This formula is applicable to scaling 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 index of the labeled image and the unlabeled image.

[0084] Specifically, according to the labeling information of the images in the preprocessed dataset, determine whether each prostate MRI image slice is labeled. The labeled image slices and the unlabeled image slices are numbered in an index manner to obtain the labeled image slice index and the unlabeled image slice index.

[0085] S103, for each prostate MRI image, perform strong and weak data augmentation processing to obtain strong augmented prostate MRI images and weak augmented prostate MRI images.

[0086] To create input perturbations, the embodiment uses two augmentation strategies 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 are specific data operation methods:

[0088] 1. Rotation

[0089] Rotation operation is achieved by angle transformation of the image, and the formula is as follows:

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

[0091] Where:

[0092] (x',y') is the position coordinate after rotation.

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

[0094]

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

[0096] θ is the rotation angle.

[0097] 2. Flip

[0098] Image flipping generally has horizontal flipping and vertical flipping, whose formulas are as follows:

[0099] 2.1 Horizontal Flip:

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

[0101] w is the image width.

[0102] 2.1 Vertical Flip:

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

[0104] h is the image height.

[0105] 3. Color Shaking

[0106] Color shaking is achieved by adjusting brightness, contrast, saturation and hue. The specific formula is 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 selected β ∈ [β min ,β max ].

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

[0115] S' = S · γ (9)

[0116] γ is a saturation factor, randomly chosen from the interval γ min , γ max ].

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

[0118] H' = (H + δ) mod 360 (10)

[0119] δ is a random offset.

[0120] 4 Random Cropping

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

[0122]

[0123] Where:

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

[0125] OcclusionRegion = {(x, y) | x1 ≤ x ≤ x2, y1 ≤ y ≤ y2}

[0126] (x1, y1) and (x2, y2) are the coordinates of the top-left and bottom-right corners of the cropping rectangle.

[0127] S104, input the strong enhancement prostate MRI image into a 3D network for feature extraction, and input the weak enhancement 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.

[0128] Wherein, the prediction result Predictions 3D output by the 3D network and the prediction result Predictions 2D output by the 2D network are as follows:

[0129]

[0130] Wherein, X represents the original data, and are data augmentation operations for strong enhancement and weak enhancement, respectively.

[0131] S105, the prediction results output by the 3D network and the 2D network are divided according to the determined slice index, and are respectively input into the LSUS strategy and the USUS strategy; wherein the LSUS strategy is used to generate a new reliable prostate image and a new label; and the USUS strategy is used to generate a new reliable prostate image.

[0132] After the two networks respectively output the prediction results, loss weight updating is performed, and then the prediction results of the two networks are divided into four kinds, strong enhancement containing labeled prediction slices (S_A), weak enhancement containing labeled prediction slices (W_A), strong enhancement not containing labeled prediction slices (S_U), and weak enhancement not containing labeled prediction slices (W_U), as shown in the figure. The strong enhancement containing labeled prediction slices and the weak enhancement containing labeled prediction slices input the LSUS strategy to generate a new reliable prostate training image (NewI) and a new corresponding label (NewGT), and the strong enhancement not containing labeled prediction slices and the weak enhancement not containing labeled prediction slices input the USUS strategy to generate a new reliable prostate image (NewI_U) process can be as follows Figures 4-7 and 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] Wherein, LSUS and USUS are different slice utilization strategies.

[0140] The principle of the LSUS strategy is as shown in Figure 4 and Figure 6 In the LSUS strategy, by comparing the real label of the labeled image with the prediction result, reliable prediction results with high confidence are screened out to generate a new reliable prostate image and a corresponding new label.

[0141] Specifically, for the prediction results generated by the 2D network, it includes two parts of images and labels. For the image part, it is cropped into K prediction patch blocks, and for the label part, it is cropped into K patch blocks and the label per-patch block mean value (non-0 is considered as the region of interest) is calculated, then the label region of interest patch position is mapped to the relative position of the prediction patch block of the image, and the confidence mean value of the region of interest of the label corresponding to the prediction patch block is calculated, then according to the confidence mean value from large to small, the patch block corresponding to the last K / 2 confidence mean values is selected.

[0142] For the prediction results generated by the 3D network, the same processing is performed, and then the patch blocks generated by the 2D network and the patch blocks generated by the 3D network are bidirectionally displaced, a new reliable prostate image and a new label are generated by randomly displacing a certain pixel distance in the horizontal and vertical directions, that is, a strongly enhanced label-containing prediction slice (S_A) and a weakly enhanced label-containing prediction slice (W_A).

[0143] The principle of the USUS strategy is shown in Figure 5 and Figure 7 In the USUS strategy, the confidence of the prediction result of the unlabeled image is calculated, and the slices with the highest confidence are selected to generate a new reliable unlabeled prostate image.

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

[0145] The new images generated through 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 loss weight is updated again.

[0146] S106, the new prostate images and new labels generated by LSUS, and the new prostate images generated by USUS are respectively input into the 3D network and the 2D network according to different data enhancement methods for further learning, the loss weight is repeatedly loaded and updated to a specified iteration number, and a trained model is obtained to obtain a prostate sub-region segmentation result according to the trained model.

[0147] Wherein, when updating the loss weight, the loss function of the model includes the supervision loss L of the labeled data supand the unsupervised loss L of unlabeled data unsup , which is defined as:

[0148]

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

[0150] The supervised loss The composition includes and represents the supervised loss of the original sample calculated by the labeled slice, represents the supervised loss of the newly generated sample calculated by the labeled slice. The specific calculation formula is:

[0151]

[0152] where and are cross-entropy loss and dice loss, respectively.

[0153]

[0154] The unsupervised loss is composed of and represents the cross-pseudo-label loss of the original sample calculated by the unlabeled slice, represents the cross-pseudo-label loss of the newly generated sample calculated by the unlabeled slice. The specific calculation formula is as follows:

[0155]

[0156] In summary, the 3D network and the 2D network are used to cooperatively train the expanded training data, and through multi-dimensional feature extraction and fusion, the segmentation ability of the model to the prostate sub-region is improved. In the training process, various perturbation methods are introduced, including data enhancement operations such as intensity transformation and adding noise to the input image, and random perturbation to the network structure, to improve the robustness and generalization ability of the model. Through consistency learning, the prediction results under different perturbations are kept consistent, reducing the supervision bias problem introduced by multiple perturbations, obtaining the final prostate sub-region segmentation model, and improving the segmentation accuracy of the prostate sub-region.

[0157] To illustrate the performance of the present embodiment, the application of the present application will be described with an actual example.

[0158] Experimental data set ​

[0159] The ProstateX dataset is a four-class prostate 3D MRI image dataset, containing 98 pairs of prostate 3D MRI images and expert hand-crafted segmentation masks, from the ProstateX challenge in 2017. For this dataset, we applied min-max normalization [0, 1] and resampled the images to a voxel size of 224x224x20. This embodiment randomly selects 70 images for training, 8 images for validation, and 20 images for testing.

[0160] The LA dataset is 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 0.625x0.625x0.625mm. This embodiment selects 80 images for training and 20 images for testing.

[0161] The ACDC dataset is a medical dataset containing three-class cardiac 3D MRI images, from the ACDC challenge in 2017. This dataset contains cardiac 3D MRI images and their corresponding expert hand-segmented masks. This embodiment performs min-max normalization [0, 1] on the 100 patient MRI images in the training data and resamples the images to a voxel size of 224x224x32. Then, 70 images are randomly selected for training, 10 images for validation, and 20 images for testing.

[0162] Evaluation indicators

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

[0164] Dice Similarity Coefficient

[0165] Dice is a statistical tool for quantifying the similarity between two samples, used to evaluate the performance of models in image segmentation and other forms of binary classification problems. The Dice coefficient is particularly suitable for handling cases where the number of positive and negative samples in the dataset is imbalanced, as it takes into account both true positives and false positives. The formula is:

[0166]

[0167] Here, X and Y represent two sets, the predicted segmented region and the actual segmented 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 measures the similarity between two sets, with values ​​ranging 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 standard

[0169] It is a metric for measuring the similarity between two sets, used to compare the similarity and differences between sample sets. It is a commonly used method for calculating the overlap between predicted and ground truth regions, especially in binary classification problems (such as image segmentation tasks). In computer vision tasks, such as object detection and segmentation, IoU is an important evaluation metric. Formula:

[0170]

[0171] Here, 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 measures the degree of similarity between two sets, with values ​​ranging from 0 (completely dissimilar) to 1 (completely identical).

[0172] 95% Hausdorff distance (HD95)

[0173] Hausdorff distance is a metric that measures the maximum distance between two sets of points. 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 of the Hausdorff distances between all pairs of points, used to reduce the influence of extreme values. The formula is:

[0174] H(A,B)=max(□(A,B),□(B,A)) (28)

[0175] in, HD95 is the 95th percentile of these distances.

[0176] Average Surface Distance ASD

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

[0178]

[0179] Where A represents the set of objects obtained from the segmentation, B represents the set of objects for the reference truth, S(A) and S(B) represent the sets of surface points of objects A and B, respectively, and d(a, S(B)) represents the Euclidean distance from surface point a to the nearest point in set S(B).

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

[0181] Experimental Design and Results Analysis

[0182] Table 1

[0183]

[0184] As shown above, Table 1 compares the method of this embodiment with 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 also compared with CPS, CT, 3D2DCT, and ABD. Figure 8 As shown, the results demonstrate that when using 20% ​​labeled slices, the method of this embodiment significantly outperforms the state-of-the-art ABD method on multiple evaluation metrics: DSC increases from 68.60 to 69.28, and Jaccard increases from 54.94 to 55.54. Regarding edge sensitivity metrics, HD95 decreases from 5.25mm to 4.19mm, and ASD decreases from 1.55mm to 1.34mm. When using 50% labeled slices, the method of this embodiment outperforms the fully supervised UNet method in all major evaluation metrics. Specifically, DSC increases from 71.66 to 72.14, and Jaccard increases from 58.46 to 58.12. Regarding edge sensitivity, HD95 decreases from 4.63mm to 3.62mm, and ASD decreases from 1.24mm to 1.03mm. With these objective evaluation metrics and visualization results, these results demonstrate that the framework of this embodiment can effectively suppress the effects of mixed perturbations and improve consistency learning. The method described in this embodiment consistently outperforms existing state-of-the-art methods at both 20% and 50% labeled slices, demonstrating significant advantages, particularly in edge accuracy and consistency learning. This demonstrates the strong potential of the method in prostate medical image segmentation tasks, achieving excellent segmentation performance even in the presence of scarce labeled data.

[0185] Table 2

[0186]

[0187] Table 2 verifies that the method of the embodiment performs additional comparative experiments on the LA dataset. The comparison results show that, under the conditions of 5% and 10% labeled slices, the method of the embodiment performs better than the current most advanced segmentation methods (including DTC, SS-Net, MCF, BCP and PMT). Under the condition of 10% labeled slices, the method of the embodiment achieves a Dice Similarity Coefficient (DSC) of 91.33 and a Jaccard Coefficient of 83.82, which is higher than the PMT method and other methods, and shows higher segmentation accuracy. In addition, the method of the embodiment also performs well in edge sensitivity indicators, with HD95 of 3.82 mm and ASD of 1.14 mm, which is better than all the comparison methods. This shows that the model of the embodiment has obvious advantages in accurately capturing boundary details. When the labeling ratio is reduced to 5%, the method of the embodiment still performs well, although some indicators are slightly lower than PMT. Among them, DSC and Jaccard are 89.45 and 81.29 respectively, and the edge sensitivity indicators HD95 and ASD are reduced to 5.86 mm and 2.02 mm respectively, but are still better than most of the comparison methods. As shown in Figure 9 , it can be seen from the 3D visualization segmentation results that the model of the embodiment is closer to the true label than other models.

[0188] Table 3

[0189]

[0190] Table 3 verifies that the method of the embodiment performs additional comparative experiments on the ACDC dataset, and compares the method of the 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 the embodiment performs better than the CPS, CT, 3D2DCT and ABD methods. Compared with the current most advanced ABD method, the method of the embodiment improves in multiple indicators. Among them, the Dice Similarity Coefficient (DSC) is improved from 89.75 to 89.83, and the Jaccard Coefficient is improved from 81.81 to 81.98. In terms of edge sensitivity indicators, the method of the embodiment performs particularly outstanding: HD95 is reduced from 3.17 mm to 1.85 mm, and ASD is reduced from 1.21 mm to 0.70 mm. Further verification results show that, in comparison with the ABD method, the method of the embodiment continues to be superior in indicators. DSC is improved from 90.26 to 90.35, and the Jaccard Coefficient is improved from 82.63 to 82.83. At the same time, it is still significantly better than ABD in edge sensitivity: HD95 is reduced from 2.38 mm to 1.76 mm, and ASD is reduced from 0.74 mm to 0.70 mm. Combined with these objective evaluation indicators and visualization results (as shown in Figure 10As shown in FIG. 9, the edge accuracy of the model of the embodiment is obviously better than that of the model of the prior art (as shown in FIG. 8), which indicates that the model of the embodiment has obvious advantages in edge accuracy. In particular, the method of the embodiment can more accurately capture the boundary details of each region of the heart, and significant improvement is achieved in boundary accuracy and precision, which is of great significance for heart image segmentation.

[0191] Obviously, various modifications and variations of the embodiments of the present application can be made by those skilled in the art without departing from the spirit and scope of the embodiments of the present application. Thus, it is intended that the present application include modifications and variations of the embodiments of the present application provided they come within the scope of the appended claims and their equivalents.

Claims

1. A medical image segmentation method for the prostate subregion based on multi-perturbation consistency learning, characterized in that, include: S101, acquire a prostate MRI image dataset and preprocess the prostate MRI images in the dataset; wherein, the prostate MRI images include labeled images and unlabeled images; S102, determine the slice indices of the labeled and unlabeled images; S103, Perform 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; the strong data enhancement includes rotation, flipping, color jitter, and occlusion data; the weak data enhancement includes rotation and flipping. S104, the strongly enhanced prostate MRI image is input into a 3D network for feature extraction to obtain the prediction result output by the 3D network, and the weakly enhanced prostate MRI image is input into a 2D network for learning to obtain the prediction result output by the 2D network. S105, the prediction results output by the 3D network and the 2D network are divided into data according to the determined slice index, and the LSUS strategy and the USUS strategy are adopted 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. The LSUS strategy is a multi-patch bidirectional displacement strategy based on the region of interest (ROI) guided by the region of interest (ROI) designed for the labeled slices, and the USUS strategy is a multi-patch bidirectional displacement strategy that selects the slices with the highest confidence by calculating the confidence of the prediction results of the unlabeled images. S106. The new prostate image and new annotations generated by LSUS, as well as the new prostate image generated by USUS, are input into the 3D network and 2D network respectively for further learning, depending on the data augmentation method. The training and updating loss weights are repeatedly loaded to a specified number of iterations to obtain the trained model, so as to obtain the prostate subregion segmentation result based on the trained model.

2. The prostate subregion medical image segmentation method based on multi-perturbation consistency learning as described in claim 1, characterized in that, In step S101, the prostate MRI image is preprocessed, including: resampling the prostate MRI image using a trilinear interpolation method based on a preset target size and original size to make it isotropic; and normalizing the resampled prostate MRI image using a maximum-minimum normalization method to scale its grayscale value range to between 0 and 1.

3. The prostate subregion medical image segmentation method based on multi-perturbation consistency learning as described in claim 1, characterized in that, In step S102, determining the slice indexes of the labeled and unlabeled images 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 subregion medical image segmentation method based on multi-perturbation consistency learning as described in claim 1, characterized in that, In step S104, Prediction results output by 3D network The prediction results of the 2D network output They are shown below: ； ； ； Where X is a table These are strong and weak data augmentation operations, respectively. The input data consists of the original prostate MRI image after weak enhancement data manipulation; The input data consists of the original prostate MRI image after high-contrast data manipulation.

5. The prostate subregion medical image segmentation method based on multi-perturbation consistency learning as described in claim 4, characterized in that, In step S105, after the two networks output their prediction results, the loss weights are updated first, and then the prediction results of the two networks are divided into four types using the determined labeled and unlabeled slice indices: strongly enhanced and labeled prediction slices. Weakly enhanced labeled prediction slices Strongly enhanced unlabeled prediction slices The weakly enhanced unlabeled prediction slice W_U; the strongly enhanced labeled prediction slice and the weakly enhanced labeled prediction slice are used to generate new reliable prostate training images using the LSUS strategy. and new corresponding tags New reliable prostate images are generated using the USUS strategy for both strongly enhanced unlabeled prediction slices and weakly enhanced unlabeled prediction slices. The process is shown in the following formula: ； ； ； ； ； 。 6. The prostate subregion medical image segmentation method based on multi-perturbation consistency learning as described in claim 5, characterized in that, In step S106, when updating the loss weights, the model's loss function... Supervised loss including labeled data Unsupervised loss of unlabeled data Defined as: ； The weights representing the balancing of different losses gradually increase with the number of iterations, based on Gaussian heating, and are expressed as: , Indicates the current iteration number. Losses under supervision include , This indicates the calculation of the supervised loss for the original samples with labeled slices. This represents the supervised loss for calculating newly generated samples with labeled slices: ； in and These are cross-entropy loss and dice loss, respectively. ； Unsupervised loss include and composition; This represents the calculation of the cross-pseudo-label loss of the original samples from the unlabeled slice. Cross-labeling loss for newly generated samples from unlabeled slices: ； 。 7. The prostate subregion medical image segmentation method based on multi-perturbation consistency learning as described in any one of claims 1 to 6, characterized in that, The performance of the model was evaluated using four metrics: Dice similarity coefficient, Jaccard similarity coefficient, HD95, and ASD.

Images