A medical image segmentation method based on conditional Bernoulli diffusion

Abstract

This invention belongs to the field of medical image analysis technology, specifically a medical image segmentation method based on conditional Bernoulli diffusion. This method uses Bernoulli noise as a diffusion kernel to enhance the segmentation capability of the diffusion model, thereby producing more accurate segmentation maps. By leveraging the stochastic nature of the diffusion model, the method generates a series of diverse segmentation maps by repeatedly randomly sampling the initial Bernoulli noise and intermediate latent variables, highlighting salient regions of interest and providing valuable reference for radiologists. Furthermore, based on the principle of DDIM, the method efficiently samples subsequences from the overall trajectory of inverse diffusion, thereby accelerating the segmentation process. Extensive experiments on real-world data demonstrate the superior performance of the proposed segmentation method, both qualitatively and quantitatively.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image analysis, and particularly relates to a medical image segmentation method. Background Art

[0002] Medical image segmentation plays a vital role in enabling better diagnosis, surgical planning, and image-guided surgery

[0003] The inherent ambiguity and high uncertainty of medical images pose a significant challenge to accurate segmentation [2]. This ambiguity and uncertainty come from many factors: for example, the unclear boundaries of tumors in brain magnetic resonance imaging (MRI) images and the existence of multiple possible annotations in computed tomography (CT) images of lung nodules.

[0004] Existing medical image segmentation methods typically only provide a single, deterministic, most likely hypothesis segmentation map, which may lead to misdiagnosis or suboptimal treatment. Therefore, in clinical practice, it is crucial to provide radiologists with accurate and diverse segmentation maps as valuable references [3].

[0005] Recently, diffusion models[4] have shown strong capabilities in various visual generation tasks[5][6]. However, how to better integrate with discrete segmentation tasks still requires further consideration. Although many studies have combined diffusion models with segmentation tasks and made some improvements[7][8], all of these methods do not fully consider the discrete characteristics of segmentation tasks and still use Gaussian noise as their diffusion kernel.

[0006] In order to achieve accurate and diverse segmentation, the present invention proposes a novel medical image segmentation method based on the conditional Bernoulli diffusion model. Unlike the use of Gaussian noise, the present invention uses Bernoulli noise as the diffusion kernel to enhance the segmentation ability of the diffusion model, thereby producing a more accurate segmentation map. Secondly, by utilizing the randomness of the diffusion model, the present invention generates a series of diverse segmentation maps by randomly sampling the initial Bernoulli noise and intermediate latent variables multiple times, thereby highlighting the significant areas of interest and providing valuable reference for radiologists. In addition, based on the principle of the denoised diffusion implicit model (DDIM) [9], the present invention efficiently samples subsequences from the overall trajectory of the reverse diffusion, thereby accelerating the segmentation process. The present invention has demonstrated its superior performance both qualitatively and quantitatively through a large number of experiments on real data. Summary of the Invention

[0007] The purpose of the present invention is to propose a medical image segmentation method based on conditional Bernoulli diffusion, which can effectively improve the diversity and accuracy of segmentation maps, thereby improving the accuracy of medical image analysis.

[0008] First, give the problem definition: Assume Represents an input medical image with a resolution of H×W and C channels. The corresponding ground-truth segmentation map y0∈{0,1} H×W , where 0 represents the background and 1 represents the target to be segmented. Inspired by diffusion models such as the denoised diffusion probability model (DDPM) and DDIM, this paper proposes a new conditional Bernoulli diffusion model, which can be expressed as:

[0009] P θ (y01x)=∫ Pθ (y 0:T |x)dy 1:T ,

[0010] Where T is the number of latent variables in the diffusion model, that is, the number of times Bernoulli noise is added in the forward process; y1,…y T is a latent variable with the same size as the segmentation map y0. For the medical binary segmentation task, the diversified inverse process of the present invention starts from the initial Bernoulli noise It starts with , and transitions through an intermediate latent variable state constrained by the input medical image x to produce a segmentation map, where 1 represents an all-one matrix of size H × W.

[0011] The medical image segmentation method provided by the present invention uses Bernoulli noise as a diffusion kernel to enhance the segmentation capability of the diffusion model, thereby generating a more accurate segmentation map.

[0012] Secondly, by utilizing the randomness of the diffusion model, the present invention generates a series of diverse segmentation maps by randomly sampling the initial Bernoulli noise and intermediate latent variables multiple times, thereby highlighting the salient regions of interest and providing valuable reference for radiologists.

[0013] Furthermore, based on the principle of DDIM, the present invention efficiently samples subsequences from the overall trajectory of reverse diffusion, thereby accelerating the segmentation process. The present invention has demonstrated its superior performance both qualitatively and quantitatively through extensive experiments on real data.

[0014] The medical image segmentation method based on conditional Bernoulli diffusion provided by the present invention has the following specific steps:

[0015] (1) First, a Bernoulli forward noise adding process is defined to gradually add more and more Bernoulli noise to the true segmentation map y0 of the medical image x. The Bernoulli forward process q(y 1:T |y0) belongs to a Markov chain of length T:

[0016]

[0017]

[0018] in, The probability parameter is Bernoulli distribution, β is the diffusion rate, which is used to define the noise schedule of the entire forward noise addition process. Specifically, T = 1000, a linear noise schedule, that is, β t =0.0001+0.0000199t; denote α t =1-β t , The noisy segmentation map y can be obtained from time step t in the form of a closed-form solution. t :

[0019]

[0020] In order to ensure efficient calculation of the objective function in the specific implementation process, the present invention proposes a novel reparameterization technique, which is to first The Bernoulli noise ε is obtained by sampling Get y t ,in Indicates the logical operator XOR. In addition, let ⊙ denote element-wise multiplication, and Norm(·) denotes normalizing the data along the channel dimension before outputting the second channel. The specific Bernoulli posterior probability can be expressed as follows:

[0021]

[0022] in,

[0023] (2) Secondly, the present invention also defines a corresponding diversified reverse process to generate diversified segmentation maps. This process also belongs to the Markov chain and starts from the Bernoulli noise The intermediate state transition is constrained by the input medical image x. The specific diversified reverse process can be expressed as:

[0024]

[0025]

[0026] Specifically, we use the estimated t Bernoulli noise To parameterize This parameterization process utilizes a correction function

[0027]

[0028] Where |·| represents the absolute value operation.

[0029] (3) In order to associate the Bernoulli forward noise addition process with the diversified reverse process, the present invention adopts the following method to train the model: in the training stage, given a medical image and the corresponding segmentation map {x, y0}, first randomly sample the time step t from the uniform distribution {`1,…,T}, and sample the corresponding Bernoulli noise ε and y t , and calculate the corresponding Bernoulli posterior probability q(y according to formula (4) t-1 |y t ,y0). We estimate the Bernoulli noise by To parameterize p θ (y t-1 |y t ,x).

[0030] Based on the negative log-likelihood variational upper bound in the previous diffusion model

[10] , the Kullback-Leibler (KL) divergence and binary cross entropy (BCE) loss are used to optimize the proposed method. The specific objective function form is:

[0031]

[0032]

[0033] in, Indicates the mathematical expectation of the distribution q(x,y0), Denotes the distribution q(y t |y0) to find the mathematical expectation, Represents the difference between the real Bernoulli noise ε and the predicted Bernoulli noise Find the mathematical expectation of the distribution formed, D KL represents the KL divergence, ∈ i,j and Represents the real Bernoulli noise ε and the predicted Bernoulli noise respectively The jth element in the i-th row of .

[0034] Finally, the overall objective function is expressed as:

[0035]

[0036] λ KL ,λ BCE is the weight coefficient. In all subsequent experiments, KL Set to 1, λ BCE Set to log(2).

[0037] After obtaining the loss according to formula (10), Perform gradient descent.

[0038] (4) Finally, the sampling algorithm of the present invention is as follows: In the sampling stage, the present invention first Sampling to obtain the initial latent variable y T , and then iteratively calculate y under different t t-1 The present invention is compatible with both DDPM and DDIM sampling strategies: under the DDPM sampling strategy, iterative sampling is performed using formula (6); under the DDIM sampling strategy, iterative sampling is performed using the following formula:

[0039]

[0040] Among them, σ t is a hyperparameter in the DDIM model. That is This makes the sampling process less random.

[0041] This paper proposes a new parameterization technique, namely the calibration function, to estimate y t This is the biggest difference from the previously proposed discrete state diffusion model

[10]

[11]

[12] .

[0042] Through a large number of experiments on the LIDC-IDRI dataset and the BRATS 2021 dataset, this invention has demonstrated its superior performance both qualitatively and quantitatively. That is, it can generate diversified segmentation maps that are both diverse and accurate in medical image segmentation tasks with more complex conditional distributions; in more deterministic medical image segmentation tasks, by generating multiple segmentation maps, it can also accurately segment the boundaries of the target to be segmented. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is the overall framework of the model of the present invention.

[0044] Figure 2 This is a visual comparison of the segmentation effects of the present invention and other methods on the LIDC-IDRI dataset.

[0045] Figure 3 This is a visual comparison of the segmentation effects of the present invention and other methods on the BRATS 2021 dataset. DETAILED DESCRIPTION

[0046] After introducing the algorithm principle and specific steps of the present invention, the following shows the segmentation effect of the present invention on real data and the comparison with other methods, including the segmentation quality and quantitative indicators.

[0047] The data used in this experiment are from the LIDC-IDRI[13,14] and BRATS 2021

[15] datasets. LIDC-IDRI contains 1018 lung CT scans annotated by four radiologists. We use the same standard preprocessing pipeline and training-validation-testing partitioning of lung CT scans as previous work[2]

[16]

[17] . BRATS 2021 consists of four different sequences (T1, T2, FlAIR, T1CE) of MRI images for each patient. All 3D scans are cut into slices along the axial direction, and the bottom 80 slices and the top 26 slices are discarded. We treat the original four types of brain tumors as one type, thereby converting the multi-target segmentation problem into a foreground segmentation problem, which follows the previous method[9]. Our training set consists of 55,174 2D images scanned from 1126 patients, and the test set consists of 3991 2D images scanned from 125 patients. Finally, the images from LIDC-IDRI and BRAST 2021 are resized to the resolutions of 128×128 and 224×224, respectively.

[0048] We implement all methods using the PyTorch library and train the models on NVIDIA V100 GPUs. All networks are trained using the AdamW

[18] optimizer with a batch size of 32. For the BRATS 2021 dataset, the initial learning rate is set to 1e-4, and for the LIDC-IDRI dataset, the initial learning rate is set to 5e-5. Figure 1 The Bernoulli noise estimation network architecture in

[19] is the same as that of the previous diffusion-based model. For all diffusion models, a linear noise schedule with T = 1000 time steps is used. The subsequence sampling strategy of DDIM is also adopted to accelerate the segmentation process. In mini-batch training on the LIDC-IDRI dataset, the present invention learns different expertise by randomly sampling one from the four annotated segmentation maps of each image.

[0049] In the experiments, three metrics are used for performance evaluation, including generalized energy distance (GED), Hungarian matching intersection over union (HM-IoU), and Dice coefficient. We use different numbers of split samples (1, 4, 8, and 16) to calculate GED, and 16 samples to calculate HM-IoU.

[0050] Example 1: Algorithm Performance on the LIDC-IDRI Dataset

[0051] Table 1: Quantitative comparison of different methods on the LIDC-IDRI test set

[0052] method GED(16) HM-IoU(16) Prob.U-net 0.320±0.0 0.500±0.03 Hprob.U-net 0.270±0.01 0.530±0.01 CAR 0.264±0.00 0.592±0.01 JPro.U-net 0.260±0.00 0.585±0.00 PixelSeg 0.260±0.00 0.587±0.01 SegDiff 0.248±0.01 0.585±0.00 MedSegDiff 0.420±0.03 0.413±0.03 BerDiff 0.238±0.01 0.596±0.00

[0053] BerDiff is the method of the present invention. The best results are in bold and the second best results are in italic bold. Table 1 shows that the method of the present invention has obvious advantages over other existing methods. Probabilistic U-net (Prob.U-net)

[16] , Hierarchical Prob.U-net (Hprob.U-net)

[20] , and Joint Prob.U-net (JPro.U-net)

[21] use conditional variational autoencoders (cVAE) to complete diverse segmentation tasks. Calibrated adversarial optimization (CAR)

[22] uses generative adversarial networks (GAN) to refine the segmentation. PixelSeg

[23] is based on an autoregressive model, while SegDiff [7] and MedSegDiff [8] are diffusion-based models. We have the following two observations: 1) For discrete segmentation tasks, diffusion-based methods have significant advantages over traditional methods based on VAE, GAN, and autoregressive models. 2) Our BerDiff outperforms other diffusion models that use Gaussian noise as the diffusion kernel. At the same time, we Figure 2 Compared to other models, our BerDiff effectively learns diverse expertise, resulting in more diverse and accurate segmentation maps. In particular, for potentially ambiguous small nodules, such as the lung nodule example on the left, our BerDiff method produces segmentation maps that better align with the true segmentation map distribution.

[0054] Example 2: Algorithm Performance on the BRATS 2021 Dataset

[0055] Table 2: Quantitative comparison of different methods on the BRATS 2021 test set

[0056]

[0057]

[0058] BerDiff is the method of the present invention. The best results are in bold and the second best results are in italic bold. Table 2 shows that the method of the present invention has obvious advantages over other existing methods. A comparative analysis was conducted on BerDiff and other models, such as nnU-net based on CNN

[24] , TransU-net based on transformer

[25] and Swin UNETR

[26] , and SegDiff based on diffusion model. First, we found that the diffusion-based method showed superior performance compared with the traditional U-net and transformer-based methods. In addition, the high performance achieved by the U-net that shares the same architecture with the noise estimation network of the present invention highlights the effectiveness of the backbone design based on the diffusion model. In addition, the BerDiff proposed by the present invention surpasses other diffusion models that use Gaussian noise as the diffusion kernel. Finally, from Figure 3 It can be seen that BerDiff segmentation is more accurate in parts that are difficult for the human eye to identify, such as the tumor in the third MRI image. At the same time, various segmentation maps can also be generated to generate saliency segmentation maps.

[0059] In summary, the conditional Bernoulli diffusion segmentation method proposed in this invention achieves the best performance in all data and performance indicators compared with other methods. Among them, the other algorithms compared are all very common algorithms in the field of medical image segmentation.

[0060] References

[0061] [1]Haque, IRI, Neubert, J.: Deep learning approaches to biomedical image segmentation. Informatics in Medicine Unlocked 18, 100297 (2020)

[0062] [2]Baumgartner,CF,Tezcan,KC,Chaitanya,K., A.M.,Muehlematter,U.J.,Schawkat,K.,Becker,A.S.,Donati,O.,Konukoglu,E.:PHISeg:Capturinguncertainty in medical image segmentation.In:Medical Image ComputingandComputer Assisted Intervention,2019,Proceedings,Part II 22.pp.119–127.Springer(2019)

[0063] [3]Lenchik,L.,Heacock,L.,Weaver,A.A.,Boutin,R.D.,Cook,T.S.,Itri,J.,Filippi,C.G.,Gullapalli,R.P.,Lee,J.,Zagurovskaya,M.,et al.:Automatedsegmentation of tissues using CT and MRI:a systematic review.Academicradiology 26(12),1695–1706(2019)

[0064] [4]Ho,J.,Jain,A.,Abbeel,P.:Denoising diffusion probabilisticmodels.Advances in Neural Information Processing Systems 33,6840–6851(2020)

[0065] [5]Ramesh,A.,Dhariwal,P.,Nichol,A.,Chu,C.,Chen,M.:Hierarchicaltext-conditional image generation with clip latents.arXiv:2204.06125(2022)

[0066] [6]Saharia,C.,Chan,W.,Saxena,S.,Li,L.,Whang,J.,Denton,E.,Ghasemipour,S.K.S.,Ayan,B.K.,Mahdavi,S.S.,Lopes,R.G.,et al.:Photorealistictext-to-imagediffusion models with deep language understanding.arXiv:2205.11487(2022)

[0067] [7]Amit,T.,Nachmani,E.,Shaharbany,T.,Wolf,L.:SegDiff:Imagesegmentation with diffusion probabilistic models.arXiv:2112.00390(2021)

[0068] [8]Wu,J.,Fang,H.,Zhang,Y.,Yang,Y.,Xu,Y.:MedSegDiff:Medicalimagesegmentation with diffusion probabilistic model.arXiv preprint arXiv:2211.00611(2022)

[0069] [9]Wolleb,J.,Sandkühler,R.,Bieder,F.,Valmaggia,P.,Cattin,P.C.:Diffusion models for implicit image segmentation ensembles.arXiv:2112.03145(2021)

[0070]

[10] Austin,J.,Johnson,D.D.,Ho,J.,Tarlow,D.,van den Berg,R.:Structureddenoising diffusion models in discrete state-spaces.Advances inNeuralInformation Processing Systems 34,17981–17993(2021)

[0071]

[11] Hoogeboom,E.,Nielsen,D.,Jaini,P.,Forré,P.,Welling,M.:Argmax flowsand multinomial diffusion:Learning categorical distributions.Advances inNeural Information Processing Systems 34,12454–12465(2021)

[0072]

[12] Sohl-Dickstein,J.,Weiss,E.,Maheswaranathan,N.,Ganguli,S.:Deepunsupervised learning using nonequilibrium thermodynamics.In:InternationalConference on Machine Learning.pp.2256–2265.PMLR(2015)

[0073]

[13] Armato III,S.G.,McLennan,G.,Bidaut,L.,McNitt-Gray,M.F.,Meyer,C.R.,Reeves,A.P.,Zhao,B.,Aberle,D.R.,Henschke,C.I.,Hoffman,E.A.,et al.:Thelung image database consortium(LIDC)and image database resource initiative(IDRI):a completed reference database of lung nodules on CT scans.Medicalphysics38(2),915–931(2011)

[0074]

[14] Clark,K.,Vendt,B.,Smith,K.,Freymann,J.,Kirby,J.,Koppel,P.,Moore,S.,Phillips,S.,Maffitt,D.,Pringle,M.,et al.:The cancer imaging archive(TCIA):maintaining and operating a public information repository.Journal of digitalimaging 26(6),1045–1057(2013)

[0075]

[15] Baid,U.,Ghodasara,S.,Mohan,S.,Bilello,M.,Calabrese,E.,Colak,E.,Farahani,K.,Kalpathy-Cramer,J.,Kitamura,F.C.,Pati,S.,et al.:TheRSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomicclassification.arXiv:2107.02314(2021)

[0076]

[16] Kohl,S.,Romera-Paredes,B.,Meyer,C.,De Fauw,J.,Ledsam,J.R.,Maier-Hein,K.,Eslami,S.,Jimenez Rezende,D.,Ronneberger,O.:A probabilistic U-Net forsegmentation of ambiguous images.Advances in neural informationprocessingsystems 31(2018)

[0077]

[17] Selvan,R.,Faye,F.,Middleton,J.,Pai,A.:Uncertainty quantificationin medical image segmentation with normalizing flows.In:Machine Learning inMedical Imaging.pp.80–90.Springer(2020)

[0078]

[18] Loshchilov,I.,Hutter,F.:Decoupled weight decayregularization.arXivpreprint arXiv:1711.05101(2017)

[0079]

[19] Nichol,A.Q.,Dhariwal,P.:Improved denoising diffusionprobabilistic models.In:International Conference on Machine Learning.pp.8162–8171.PMLR(2021)

[0080]

[20] Kohl,S.A.,Romera-Paredes,B.,Maier-Hein,K.H.,Rezende,D.J.,Eslami,S.,Kohli,P.,Zisserman,A.,Ronneberger,O.:A hierarchical probabilistic u-netfor modeling multi-scale ambiguities.arXiv preprint arXiv:1905.13077(2019)

[0081]

[21] Zhang,W.,Zhang,X.,Huang,S.,Lu,Y.,Wang,K.:A probabilistic modelfor controlling diversity and accuracy of ambiguous medical imagesegmentation.In:Proceedings of the 30-th ACM International Conference onMultimedia.pp.4751–4759(2022)

[0082]

[22] Kassapis,E.,Dikov,G.,Gupta,D.K.,Nugteren,C.:Calibratedadversarial refinement for stochastic semantic segmentation.In:Proceedings ofthe IEEE / CVF International Conference on Computer Vision.pp.7057–7067(2021)

[0083]

[23] Zhang,W.,Zhang,X.,Huang,S.,Lu,Y.,Wang,K.:PixelSeg:Pixel-by-pixelstochastic semantic segmentation for ambiguous medical images.In:Proceedingsof the 30-th ACM International Conference on Multimedia.pp.4742–4750(2022)

[0084]

[24] Isensee,F.,Jaeger,P.F.,Kohl,S.A.,Petersen,J.,Maier-Hein,K.H.:nnU-Net:a self-configuring method for deep learning-based biomedical imagesegmentation.Nature methods 18(2),203–211(2021)

[0085]

[25] Chen,J.,Lu,Y.,Yu,Q.,Luo,X.,Adeli,E.,Wang,Y.,Lu,L.,Yuille,A.L.,Zhou,Y.:TransUNet:Transformers make strong encoders for medicalimagesegmentation.arXiv:2102.04306(2021)

[0086]

[26] Hatamizadeh,A.,Nath,V.,Tang,Y.,Yang,D.,Roth,H.R.,Xu,D.:SwinUNETR:Swin transformers for semantic segmentation of brain tumors in MRIimages.In:Brainlesion:Glioma,Multiple Sclerosis,Stroke and Traumatic BrainInjuries,Part I.pp.272–284.Springer(2022)。

Claims

1. A medical image segmentation method based on conditional Bernoulli diffusion, characterized in that: Use Bernoulli noise as the diffusion kernel to enhance the segmentation ability of the diffusion model. The specific steps are as follows: (1) Define a Bernoulli forward noise adding process, gradually adding more and more Bernoulli noise to the real segmentation map y0 of the medical image x; the Bernoulli forward process q(y 1:T |y0) belongs to a Markov chain of length T: in, Represents an input medical image with a resolution of H×W and C channels, and the corresponding true segmentation map y0∈{0,1} H×W , where 0 represents the background and 1 represents the target to be segmented; The probability parameter is Bernoulli distribution, β is the diffusion rate, which is used to define the noise schedule of the entire forward noise addition process; α t =1-β t , Obtain the noisy segmentation map y from time step t in the form of a closed-form solution t : In order to ensure efficient calculation of the objective function in the specific implementation process, reparameterization is performed, that is, first from the distribution The Bernoulli noise ε is obtained by sampling Get y t ,in Represents the logical operator XOR; let ⊙ represent element-wise multiplication, and Norm(·) represents normalizing the data along the channel dimension before outputting the second channel. The specific Bernoulli posterior probability is expressed as follows: in, (2) Define a corresponding diversified inverse process to generate a diversified segmentation map; this inverse process also belongs to the Markov chain and starts from the Bernoulli noise The intermediate state transition is constrained by the input medical image x; the specific diversified reverse process is expressed as: Specifically, using the estimated t Bernoulli noise To parameterize This parameterization process utilizes a correction function Among them, |·| represents the absolute value operation; (3) The model is trained in the following way, linking the Bernoulli forward noise addition process and the diversified reverse process: In the training phase, given a medical image and the corresponding segmentation map {x, y0}, first randomly sample the time step t from the uniform distribution {1,…,T}, and sample the corresponding Bernoulli noise ε and y t , and calculate the corresponding Bernoulli posterior probability q(y according to formula (4) t-1 |y t ,y0); by estimating the Bernoulli noise To parameterize p θ (y t-1 |y t ,x); KL divergence and binary cross entropy loss are used for optimization. The specific objective function form is: in, Indicates the mathematical expectation of the distribution q(x,y0), Denotes the distribution q(y t |y0) to find the mathematical expectation, Represents the difference between the real Bernoulli noise ε and the predicted Bernoulli noise The mathematical expectation of the distribution formed is D KL represents the KL divergence, ∈ i,j and Represents the real Bernoulli noise ε and the predicted Bernoulli noise respectively The jth element in the i-th row of ; The overall objective function is expressed as: λ KL ,λ BCE is the weight coefficient.

2. The medical image segmentation method based on conditional Bernoulli diffusion according to claim 1, characterized in that: The sampling algorithm is as follows: In the sampling phase, first Sampling to obtain the initial latent variable y T , and then iteratively calculate y under different t t-1 The probability parameters are iterated and sampled, which are compatible with both DDPM and DDIM sampling strategies: Under the DDPM sampling strategy, iterative sampling is performed using formula (6); Under the DDIM sampling strategy, iterative sampling is performed using the following formula: Among them, σ t is a hyperparameter in the DDIM model, let That is This makes the sampling process less random.

3. The medical image segmentation method based on conditional Bernoulli diffusion according to claim 1, characterized in that: Specifically, take T = 1000 and use a linear noise schedule, i.e., β t =0.0001+0.0000199t.

Images