Method and system for synthesizing pathological digital image

Abstract

The invention relates to a digital image synthesis technology, and discloses a pathology digital image synthesis method and system, and the method comprises the steps: obtaining a pathology digital image, and obtaining the pathology digital image from a pathology digital image library; processing the pathological digital image: processing the acquired pathological digital image through a diffusion model; and pathology digital image synthesis: synthesizing the processed pathology digital image into an output pathology digital image. For the synthesis of pathological digital images, the clinical sample collection time is shortened, and a large number of pathological images with specific disease features are quickly generated. The time required for traditional sample collection is greatly shortened, so that research can be promoted more quickly. As the generated images are digital, the generated images can be directly used for analysis and research, and a physical scanning process is not needed. Therefore, the service time of the scanning equipment is saved, and errors and damage possibly occurring in the scanning process are avoided.

Description

Technical Field

[0001] The invention relates to digital image synthesis technology, and in particular to a synthesis method and system for pathological digital images. Background Art

[0002] In the practice and research of pathology, the collection and processing of pathology slides is a crucial link, which is directly related to the accuracy of disease diagnosis and the effectiveness of subsequent scientific research analysis. This process usually covers a delicate sampling step, that is, accurately selecting representative samples from the patient's body or diseased tissue; then, through a series of chemical fixative treatments to stabilize cell morphology, maintain the integrity of tissue structure, prevent autolysis and corruption, and lay the foundation for subsequent analysis.

[0003] The slicing operation requires extremely high technical precision. A microtome is used to cut the fixed tissue samples into extremely thin slices, usually with a thickness of several microns to tens of microns, so that they can be clearly observed under a microscope. After the slicing is completed, it is necessary to go through a complex preparation process, including staining, dehydration, transparency, and sealing steps to enhance the contrast of cell structures and facilitate pathologists' identification and diagnosis. However, traditional pathology slides have certain limitations in preservation, such as being easily affected by environmental factors, causing fading and deformation, and having a relatively short preservation period, which is not conducive to long-term archiving and remote consultation.

[0004] In order to meet these challenges, the application of pathology scanners came into being. It uses high-precision scanning technology to convert traditional pathology slides into high-resolution digital pathology slides. This process not only perfectly preserves the original information of the slides, such as cell morphology, tissue structure details and staining characteristics, but also greatly extends the preservation time of pathology data and avoids the risk of physical wear and chemical degradation. At the same time, digital pathology slides are easy to transmit and share in computer networks, which promotes the in-depth development of telemedicine, multidisciplinary consultation, and teaching and scientific research.

[0005] The high-resolution characteristics of pathology scanners provide unprecedented image details, but also bring challenges to scanning efficiency. A pathology digital slide image has a resolution of 100,000*100,000 pixels, which means that every subtle structural change can be accurately captured, which is crucial to improving the accuracy and reliability of pathology diagnosis. However, such a high pixel density requires the scanner to be meticulous during the scanning process, scanning and recording the color and brightness information of each pixel point by point, which naturally leads to a significant increase in scanning time.

[0006] Even when the scanned object is a piece of tissue that is only a few millimeters in size, although its physical size is not large, the number of pixels that the scanner needs to cover is extremely large due to the requirement of high resolution. In this case, even if the scanning speed is fast, it will inevitably take a long time to complete the entire scanning process. Digital pathology slides are often accompanied by a huge amount of data due to their high resolution and detailed display of details. The digital file of a single slide may reach several GB or even more than ten GB, which puts higher requirements on the capacity and performance of storage devices. Therefore, while promoting the use of digital pathology slides, it is also necessary to continuously optimize data storage solutions, such as using efficient compression algorithms, building distributed storage systems, or using cloud storage technology to effectively alleviate the pressure on hard disk space and ensure the long-term safe storage and efficient use of massive pathology data.

[0007] Pathological research in teaching and clinical research, especially the collection and digital processing of samples of rare diseases, does face multiple challenges. First, it is extremely difficult to obtain clinical samples of rare diseases because the incidence of such diseases is low, the number of patients is limited, and they are often distributed in different medical institutions, making it difficult to collect them centrally.

[0008] Secondly, the time-consuming issue of data scanning cannot be ignored. As mentioned earlier, although high-resolution pathology scanners can provide fine image details, the corresponding scanning time will also increase significantly. This problem is particularly prominent when a large number of samples of rare diseases need to be scanned, which may seriously affect the progress of research.

[0009] Furthermore, storage space is also an important consideration. High-resolution digital pathology images are large in size, and a single sample can take up a lot of hard disk space. For large-scale, long-term research projects, storage costs and management difficulties will rise sharply. Summary of the invention

[0010] The present invention aims to solve the problems in the prior art that large-size digital pathology images will increase the scanning time and a single slide digital pathology image will occupy a large amount of memory, and provides a method and system for compensating for the gap of a detection plate in a security inspection image.

[0011] In order to solve the above technical problems, the present invention is solved by the following technical solutions: A method for synthesizing a pathological digital image, the method comprising: Acquisition of pathological digital images, obtaining pathological digital images from a pathological digital image library; Processing of pathological digital images: processing the acquired pathological digital images through diffusion model; The synthesis of the pathological digital image is to synthesize the processed pathological digital image into an output pathological digital image.

[0012] Preferably, the pathological digital image is processed by performing forward diffusion and reverse denoising on the acquired pathological digital image through a diffusion model.

[0013] Preferably, the forward diffusion constructs a series of noise-contaminated pathological digital image data with different noise levels by gradually adding random Gaussian noise to the pathological digital image data.

[0014] Preferably, the reverse denoising is performed by training Unet to gradually remove noise in the pathological digital image data, thereby generating a clear pathological digital image.

[0015] Preferably, the forward diffusion of pathological digital image data includes: Segmentation of digital pathology images, segmenting the acquired digital pathology images into different digital pathology image blocks; Pollution of the digital pathology image block after segmentation processing, adding a first noise η(t) to the digital pathology image block after segmentation processing to perform pollution processing; For the cleaned image x(0) input during the training process, the first noise is added to this image in each step of the forward diffusion. The image after adding noise in the tth step is recorded as x(t), and the image after adding noise in the t+1th step is recorded as x(t+1). Then the noise adding process from x(t) to x(t+1) is expressed as: x(t+1) = x(t) + η(t) Where, η(t) is the first noise; The contaminated digital pathology image block is transferred to the Unet model to obtain the second noise x(t+1)–x(t); Calculation of noise loss: Calculation of noise loss for the first noise η(t) and the second noise x(t+1)–x(t) to obtain noise loss Loss(t); The calculated noise loss is transmitted to the Unet model, and the Unet model is updated; Loss_t= MSE((x_(t+1)–x(t)),η(t)}) = 1 / m Σ(x_(t+1)–x(t)),η(t)})^2 Here, m is the number of pixels in the image.

[0016] Preferably, the Unet model includes an encoder and a decoder, and the input digital pathology image is reduced in dimension into an embedding vector by the encoder, and then reconstructed into a new digital pathology image by the decoder.

[0017] Preferably, the reverse denoising process includes predicting the noise that needs to be removed in the current step t through the trained Unet; the noisy digital pathology image of the current step is x(t), and x(t) is the noisy digital pathology image of step number t; after predicting the noisy digital pathology image, the noisy digital pathology image x(t) of the current step is subtracted from the predicted noisy digital pathology image to obtain the digital pathology image x(t-1) of the previous step.

[0018] In order to solve the above technical problems, the present invention also provides a digital pathological image synthesis system, which includes: a pathological digital image acquisition module, which acquires the pathological digital image in the pathological digital image library; A pathological digital image processing module processes the acquired pathological digital image through a diffusion model; The synthesis module of the pathological digital image synthesizes the processed pathological digital image into an output pathological digital image.

[0019] The present invention has significant technical effects due to the adoption of the above technical solution: The present invention synthesizes pathological digital images, which reduces the time for clinical sample collection: by generating models, researchers can quickly generate a large number of pathological images with specific disease characteristics without relying on actual clinical samples. This greatly shortens the time required for traditional sample collection, allowing research to proceed faster.

[0020] The invention reduces scanning time: since the generated images are digital, they can be directly used for analysis and research without going through the physical scanning process. This not only saves the use time of the scanning equipment, but also avoids errors and damage that may occur during the scanning process.

[0021] The present invention occupies little memory for images, and images can be generated instantly according to demand without pre-storing a large amount of data. When needed, the required images can be generated by simply running the model and specifying relevant parameters, thereby avoiding the space pressure caused by massive image storage.

[0022] The method of the present invention is flexible and customizable: by adjusting the input parameters of the model, researchers can control the characteristics of the generated images, such as the type and severity of the disease. This flexibility enables the model to meet the needs of different research scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of forward diffusion and reverse denoising of the present invention; Figure 2 is the original image of the gap of the detection plate of the present invention; Figure 3 is the original image of the present invention; Figure 4 An image for visualizing random noise of the present invention; Figure 5 Image 1 contaminated by random noise of the present invention; Figure 6 The second image is contaminated by random noise of the present invention; Figure 7 Image 3 is the random noise contaminated image of the present invention; Figure 8 The image after random noise pollution denoising of the present invention; Fig. 9 This is the image after sharpening according to the present invention. DETAILED DESCRIPTION

[0024] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0025] Example 1

[0026] A method for synthesizing a pathological digital image, the method comprising: Acquisition of pathological digital images, obtaining pathological digital images from a pathological digital image library; Processing of pathological digital images: processing the acquired pathological digital images through diffusion model; The synthesis of the pathological digital image is to synthesize the processed pathological digital image into an output pathological digital image.

[0027] Processing of pathological digital images: forward diffusion and reverse denoising are performed on the acquired pathological digital images through a diffusion model.

[0028] Forward diffusion constructs a series of noise-contaminated pathological digital image data with different noise levels by gradually adding random Gaussian noise to the pathological digital image data.

[0029] Reverse denoising gradually removes noise from pathological digital image data by training Unet, thereby generating clear pathological digital images.

[0030] Forward diffusion of pathology digital image data includes: Segmentation of digital pathology images, segmenting the acquired digital pathology images into different digital pathology image blocks; Pollution of the digital pathology image block after segmentation processing, adding a first noise η(t) to the digital pathology image block after segmentation processing to perform pollution processing; For the cleaned image x(0) input during the training process, the first noise is added to this image in each step of the forward diffusion. The image after adding noise in the tth step is recorded as x(t), and the image after adding noise in the t+1th step is recorded as x(t+1). Then the noise adding process from x(t) to x(t+1) is expressed as: x(t+1) = x(t) + η(t) Where, η(t) is the first noise; For an original image Figure 4 ,That Figure 5-Figure 7 The contaminated digital pathology image block is transferred to the Unet model to obtain the second noise x(t+1)–x(t); As t increases, the distribution of x(t) will gradually tend to the standard normal distribution and be completely contaminated into random noise. The training and implementation process of the Diffusion model is time-series, and each image will be iterated t times in the process. In practical applications, in order to control the amount of noise added in each step and ensure that the noise changes smoothly, we usually adjust the variance of the noise according to the number of steps t.

[0031] Calculation of noise loss: Calculation of noise loss for the first noise η(t) and the second noise x(t+1)–x(t) to obtain noise loss Loss(t); The calculated noise loss is transmitted to the Unet model, and the Unet model is updated; Loss_t= MSE((x_(t+1)–x(t)),η(t)}) = 1 / m Σ(x_(t+1)–x(t)),η(t)})^2 Here, m is the number of pixels in the image.

[0032] The Unet model includes an encoder and a decoder. The input digital pathology image is reduced to an embedding vector by the encoder. The encoder reduces the dimension of the embedding vector to remove redundant information. Then it is reconstructed into a new digital pathology image through the decoder.

[0033] Figure 8 In the process of reverse denoising, the noise that needs to be removed in the current step t is predicted by the trained Unet; the noisy digital pathology image of the current step is x(t), and x(t) is the noisy digital pathology image of step number t; after the noisy digital pathology image is predicted, the noisy digital pathology image x(t) of the current step is subtracted from the predicted noisy digital pathology image to obtain the digital pathology image x(t-1) of the previous step.

[0034] In the denoising process, the amount of noise removed at each step is not fixed, but is determined by the learning results of the model. Therefore, in the reverse process, the noise needs to be removed step by step and carefully to obtain a high-quality restored image.

[0035] Example 2

[0036] Based on Example 1, this example Fig. 9Image sharpening is also included in the image degeneration process. Image sharpening is performed by a sharpening filter to enhance edges and details by performing convolution operations on the image. Commonly used sharpening filters include Prewitt operator, Roberts operator and Laplace operator. These operators can make the contrast of the high-frequency part of the image with large gradient changes more obvious, thus making the image clearer.

[0037] Example 3

[0038] Based on the first embodiment, this embodiment is a digital pathological image synthesis system, which includes: a pathological digital image acquisition module, which acquires the pathological digital image in the pathological digital image library; A pathological digital image processing module processes the acquired pathological digital image through a diffusion model; The synthesis module of the pathological digital image synthesizes the processed pathological digital image into an output pathological digital image.

Claims

1. A method for synthesizing a pathological digital image, the method comprising: Acquisition of pathological digital images, obtaining pathological digital images from a pathological digital image library; Processing of pathological digital images: processing the acquired pathological digital images through diffusion model; The synthesis of the pathological digital image is to synthesize the processed pathological digital image into an output pathological digital image.

2. A method for synthesizing a pathological digital image according to claim 1, characterized in that: Processing of pathological digital images: forward diffusion and reverse denoising are performed on the acquired pathological digital images through a diffusion model.

3. A method for synthesizing a pathological digital image according to claim 2, characterized in that: Forward diffusion constructs a series of noise-contaminated pathological digital image data with different noise levels by gradually adding random Gaussian noise to the pathological digital image data.

4. The method for synthesizing a pathological digital image according to claim 2, characterized in that: Reverse denoising gradually removes noise from pathological digital image data by training Unet, thereby generating clear pathological digital images.

5. The method for synthesizing a pathological digital image according to claim 2, characterized in that: Forward diffusion of pathology digital image data includes: Segmentation of digital pathology images, segmenting the acquired digital pathology images into different digital pathology image blocks; Pollution of the digital pathology image block after segmentation processing, adding a first noise η(t) to the digital pathology image block after segmentation processing to perform pollution processing; For the cleaned image x(0) input during the training process, the first noise is added to this image in each step of the forward diffusion. The image after adding noise in the tth step is recorded as x(t), and the image after adding noise in the t+1th step is recorded as x(t+1). Then the noise adding process from x(t) to x(t+1) is expressed as: ； Where, η(t) is the first noise; The contaminated digital pathology image block is transferred to the Unet model to obtain the second noise x(t+1)–x(t); Calculation of noise loss: Calculation of noise loss for the first noise η(t) and the second noise x(t+1)–x(t) to obtain noise loss Loss(t); The calculated noise loss is transmitted to the Unet model, and the Unet model is updated; Loss_t= MSE((x_(t+1)–x(t)),η(t)}) = 1 / m Σ(x_(t+1)–x(t)),η(t)})^2; Here, m is the number of pixels in the image.

6. The method for synthesizing a pathological digital image according to claim 2, characterized in that: The Unet model includes an encoder and a decoder. The input digital pathology image is reduced into an embedding vector by the encoder and then reconstructed into a new digital pathology image by the decoder.

7. The method for synthesizing a pathological digital image according to claim 2, characterized in that: The reverse denoising process includes predicting the noise that needs to be removed in the current step t through the trained Unet; the noisy digital pathology image of the current step is x(t), and x(t) is the noisy digital pathology image of step number t; after predicting the noisy digital pathology image, the noisy digital pathology image x(t) of the current step is subtracted from the predicted noisy digital pathology image to obtain the digital pathology image x(t-1) of the previous step.

8. A digital pathological image synthesis system, characterized in that: include: A pathological digital image acquisition module is used to acquire the pathological digital image from the pathological digital image library; A pathological digital image processing module processes the acquired pathological digital image through a diffusion model; The synthesis module of the pathological digital image synthesizes the processed pathological digital image into an output pathological digital image.

Images