A method for predicting cell nucleus size from polarized scattering spectra based on Bayesian neural network

Abstract

The present invention relates to a method for predicting cell nucleus size based on the combination of polarized scattering spectroscopy and Bayesian deep learning. By measuring the single scattering spectrum of diseased tissue and using a trained Bayesian neural network to predict the input single scattering spectrum, the tissue's cell nucleus size can be determined. The method includes the following steps: establishing a physical model for single scattering spectroscopy, generating a single scattering spectrum training set, constructing a Bayesian neural network, training the neural network using the generated single scattering spectrum training set, and evaluating the model; and using the trained Bayesian network model to predict the single scattering spectrum of the sample to be tested, providing the cell nucleus size and the prediction uncertainty. Based on Bayesian deep learning, the present invention achieves rapid inversion of cell nucleus size, improving the inversion speed and improving the prediction uncertainty.

Description

Technical Field

[0001] The present invention belongs to the technical field of early gastrointestinal cancer detection, and in particular relates to a method for predicting cell nucleus size using polarization scattering spectra based on a Bayesian neural network. Background Art

[0002] Early diagnosis and treatment of gastrointestinal malignancies can significantly improve patients' long-term survival and quality of life. However, due to their location in the human body, malignant tumors are often diagnosed only at a late stage. Currently, the diagnosis and treatment of gastrointestinal malignancies are performed through endoscopy. Endoscopists detect lesions and perform biopsies, while pathologists perform microscopic examination and diagnosis of the biopsied tissue. These examinations aim to accurately diagnose and locate residual cancer. However, this is a complex task that requires extensive knowledge and experience from both endoscopists and pathologists, and the diagnostic process is time-consuming. Magnifying endoscopy combined with narrow-band imaging (ME-NBI) is typically used for the endoscopic diagnosis of early-stage gastrointestinal cancer. ME-NBI generally diagnoses early-stage malignancies based on abnormal glandular and capillary morphology. In contrast, pathological diagnosis relies on abnormal nuclear morphology, such as nuclear size, shape, and density. This morphological information can be obtained from single-shot scattering spectroscopy. Therefore, light scattering spectroscopy (LSS) is a promising method that can non-invasively obtain abnormal nuclear morphological information in vivo from the single scattered light collected from epithelial cells [Mourant, JR; Bigio, IJ; Boyer, J.; Conn, RL; Johnson, T.; Shimada, T., Spectroscopic diagnosis of bladder cancer with elastic light scattering. Lasers Surg. Med. 1995, 17(4), 350-357.]. After Backman et al. pioneered the polarized light scattering spectroscopy (PLSS) technology in U.S. Patent No. 6624890B2, researchers have developed many novel PLSS systems that have shown good diagnostic effects in clinical practice. In existing PLSS systems, either a mechanically rotatable polarizer is used, such as in U.S. Patent No. 6624890B2, or two parallel orthogonal polarizers are used, such as in U.S. Patent No. 9788728B2, to separate the single scattered spectrum. In order to simplify the PLSS system and improve detection efficiency, we proposed a snapshot PLSS technology [Tuniyazi A, Mu T, Jiang X, Han F, Li H, Li Q, Gong H, Wang W, Qin B. Snapshot polarized light scattering spectroscopy using spectrally-modulated polarimetry for early gastric cancer detection. J Biophotonics. 2021 Sep; 14(9): e202100140.].

[0003] Regardless of the PLSS system used, nuclear morphology should be inferred from the scattering spectrum. Generally, nuclear size is inferred by fitting experimental and theoretical spectra using a model-driven approach. Although such methods can estimate the size distribution of epithelial cell nuclei with relatively high precision and accuracy, fitting a scattering spectrum requires a long time, which is a challenge for large-scale, real-time early cancer diagnosis. To achieve real-time diagnosis, Le Qiu et al. proposed a fast semi-empirical algorithm [L. Qiu, R. Chuttani, D.K. Pleskow, V. Turzhitsky, U. Khan, Y.N. Zakharov, L. Zhang, T.M. Berzin, E.U. Yee, M.S. Sawhney, Y. Li, E. Vitkin, J.D. Goldsmith, I. Itzkan, and L.T. Perelman, "Multispectral light scattering endoscopic imaging of esophageal precancer," Light Sci Appl 7, 17174 (2018)]. This algorithm does not require inverting the size of epithelial cell nuclei. Instead, it provides an empirical parameter based on the statistical properties of the spectrum and uses this empirical parameter to diagnose cancer. Although this algorithm is fast, it relies heavily on physical priors about tissue characteristics and is a semi-quantitative method. In pathological diagnosis, the determination of tissue cancer is primarily based on the morphological information of the cell nucleus. Cancer cell nuclei are generally larger than normal cell nuclei. Therefore, in practice, the size distribution of epithelial cell nuclei is very valuable for quantitatively diagnosing early-stage cancer. To achieve the real-time inversion performance of PLSS technology for cell nuclear size, a faster method capable of real-time inversion of cell nuclear size is urgently needed. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a polarization scattering spectrum cell nucleus size prediction method based on Bayesian neural network to solve the problems of slow inversion of cell nucleus size by PLSS technology and lack of uncertainty assessment in size prediction.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for predicting cell nucleus size based on polarization scattering spectra of cells based on Bayesian neural network, comprising the following steps:

[0007] Step 1: Use Mie scattering theory combined with a polarization gate scattering spectroscopy system to generate a single scattering spectrum dataset of biological tissues and divide it into a training set and a test set;

[0008] Step 2, constructing a Bayesian neural network;

[0009] Step 3, using the training set and the test set, training the Bayesian neural network and performing test evaluation to obtain a classification model, wherein the input of the classification model is the single scattering spectrum and the output is the cell nucleus size of the biological tissue;

[0010] Step 4: Measure the single scattering spectrum of the biological tissue, input it into the classification model, and obtain the cell nucleus size.

[0011] In one embodiment, generating the single scattering spectrum of the biological tissue comprises the following steps:

[0012] Step 11: Based on the polarization gate scattering spectrum system and Mie scattering theory, a physical model of single scattering spectrum is established:

[0013]

[0014] Where ΔI is the single scattering spectrum, which is approximately equal to the difference between the parallel polarization component and the perpendicular polarization component, κ is the normalization factor, I0 is the intensity of the incident linearly polarized light, θ is the scattering angle, θ0 is the angular range over which the backscattered light is collected, D is the diameter of the cell nucleus, ΔD is the size range of the cell nucleus diameter, s2 is an element of the scattering amplitude matrix, λ is the wavelength of the incident linearly polarized light, n is the relative refractive index, and f(D) is the cell nucleus size distribution function, which satisfies the Gaussian distribution with a mean of The standard deviation is σ;

[0015] In step 12, the acceptable ranges of D, σ, and n are determined based on the physical properties of the biological tissue, while other parameters are kept unchanged. The physical model in step 11 is then introduced to generate a data set of single scattering spectra.

[0016] In one embodiment, in the dataset, the single scattering spectrum is normalized and used as an input item in the training set, and D, σ, and n are merged into one label, that is, only the size is used as the label of the training set. Then, the dataset is partitioned according to the size with Δμm as the interval, and each area is used as a category. The mean of each area is used as the label of the category, that is, the size is used as the label of the category; the same method is used to generate a test set, wherein the values of D, σ, and n in the training set and the test set are different.

[0017] In one embodiment, the steps of constructing a Bayesian neural network in step 2 are as follows:

[0018] Step 21: Establish a Bayesian deep learning classification network model, where the number of network layers is set to input layer, convolution layer, smoothing layer, Dropout layer and output layer in sequence.

[0019] In step 22, the dimension of the input layer is set to the number of bands of the single scattering spectrum, the nodes of the output layer are set to the number of categories, and the Dropout ratio is set to r.

[0020] In one embodiment, the steps of step 3, training the Bayesian neural network and performing test evaluation, are as follows:

[0021] Step 31, input the training set into the Bayesian neural network;

[0022] Step 32: judge the learning effect based on the accuracy and loss of the validation set, and save the learned network parameters;

[0023] In step 33, the test set is input into the trained Bayesian neural network to obtain the evaluation index. If the evaluation index meets the requirements, the final classification model is obtained; otherwise, the training continues.

[0024] In one embodiment, step 4 includes the following steps:

[0025] Step 41, using the polarization gate scattering spectroscopy system, under the same conditions as those set in step 1, a polarization scattering spectroscopy experiment is performed on the biological tissue to obtain a single scattering spectrum, the single scattering spectrum is normalized, and then input into the classification model to predict T times;

[0026] Step 42: Perform statistical calculations on the T outputs of the classification model and take the average value as the cell nucleus size.

[0027] In one embodiment, the classification model also outputs model uncertainty.

[0028] In one embodiment, the biological tissue is early digestive tract cancer tissue such as esophageal cancer, gastric cancer and colon cancer in the human body.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. Compared with the existing technology, the cell nucleus size prediction method of the present invention, which is based on the combination of polarization scattering spectroscopy technology and Bayesian deep learning technology, uses a Bayesian deep learning network to achieve the inversion of cell nucleus size, which improves the inversion speed and makes it possible to diagnose early cancer in real time using PLSS technology.

[0031] 2. Compared with traditional deep learning networks, the present invention uses Bayesian deep learning networks to provide uncertainty information for the network's predictions; based on the uncertainty information, difficult-to-predict cases can be effectively distinguished.

[0032] 3. The present invention cleverly converts the size prediction problem into a classification problem and improves the inversion accuracy by utilizing the advantages of convolutional neural networks in classification problems.

[0033] 4. The present invention utilizes the probability posterior distribution output by the deep learning classification network to improve the utilization rate of model uncertainty; in the size prediction based on the regression model, since the model uncertainty is the standard deviation of the size predicted by the network, the model uncertainty is insensitive to the prediction error; while in the size prediction based on the classification model, since the model uncertainty is the standard deviation of the probability predicted by the network, it is very sensitive to the prediction error; therefore, the size inversion method based on the deep learning classification model improves the effectiveness of the model uncertainty. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the present invention.

[0035] Figure 2 This is a schematic diagram of the Bayesian neural classification model framework. DETAILED DESCRIPTION

[0036] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0037] Based on the above, existing PLSS techniques use least squares fitting to invert cell size or diagnose pathological conditions based on the statistical characteristics of multiple single scattering spectra. The former method requires a significant amount of time to iterate during the inversion process, making it impossible to provide real-time results in practical applications. While the latter method offers rapid diagnostic speed, it lacks information on cell nucleus size, significantly reducing the value of PLSS technology in early cancer detection. In this paper, we combine polarized scattering spectroscopy with Bayesian deep learning techniques to propose a method for predicting cell nucleus size.

[0038] like Figure 1 As shown, the cell nucleus size prediction method of the present invention includes two aspects: generating a training set based on the physical model of single scattering spectra, constructing a Bayesian network, and training the model. After collecting actual single scattering spectra, they are normalized and input into the trained model to predict the cell nucleus size and model uncertainty. Ultimately, a diagnosis result can be provided based on the cell nucleus size and uncertainty information. The method specifically includes the following steps:

[0039] The first step is to generate a single scattering spectrum training set.

[0040] Based on the Mie scattering theory and the polarization gate scattering spectrum system, a single scattering spectrum dataset of biological tissue is generated and divided into a training set and a test set. In the field of gastrointestinal diseases, it is difficult to collect a large number of real spectra of gastrointestinal samples. A promising method is to use a physical model to generate a single scattering spectrum dataset. In order to train the Bayesian network, the present invention generates a spectral dataset based on the analysis model in Formula 1. In order to make the generated dataset more scientific and closer to the real spectrum, the parameters used to generate the spectrum should fully cover the parameters that may appear in actual experiments. The specific steps are as follows:

[0041] (1) Based on the polarization gate scattering spectrum system and Mie scattering theory, a physical model of single scattering spectrum is established:

[0042]

[0043] Where ΔI is the single scattering spectrum, which is approximately equal to the difference between the parallel polarization component and the perpendicular polarization component. κ is the normalization factor, I0 is the intensity of the incident linear polarization light, θ is the scattering angle, θ0 is the angle range of the backscattered light collected, D is the diameter of the cell nucleus, ΔD is the size range of the cell nucleus diameter, s2 is the element of the scattering amplitude matrix, λ is the wavelength of the incident linear polarization light, n is the relative refractive index, and f(D) is the cell nucleus size distribution function. In the present invention, the cell nucleus size distribution satisfies the Gaussian distribution with a mean of The standard deviation is σ.

[0044] In one embodiment, the wavelength of the incident light is in the range of 480-680 nm, and the angle range of the backscattered light collected is 1°.

[0045] (2) In this model, D, σ, and n are free parameters, while other parameters such as the intensity of the incident linearly polarized light and the scattering angle are constant. Based on the physical properties of biological tissue, the acceptable ranges of D, σ, and n are determined and substituted into the physical model of formula (1) to generate a data set of single scattering spectra.

[0046] In the generated dataset, the normalized single scattering spectra are used as input in the training set, and the free parameters D, σ, and n are used as outputs or labels for the training set. In this case, each spectrum has three parameters, i.e., three labels.

[0047] In one embodiment, to ensure sufficient spectral diversity and non-overlapping, two free parameters can be fixed while the other parameter is gradually varied. The subjects are the colon, with an average diameter of 5.5-12.5 μm, with a step size of 0.1 μm; a relative refractive index of 1.02-1.05, with a step size of 0.001; and a standard deviation of 1.3-1.6 μm, with a step size of 0.01. Ultimately, 63,000 spectra are generated. All of these parameters can be flexibly adjusted based on the subject being examined. The free parameters serve as the output, or labels, of the training set.

[0048] (3) Convert the three-label input dataset in (2) into a classification problem.

[0049] In one embodiment, the three labels D, σ, and n are first merged into one label, that is, only the size is used as the label of the training set. Then, the data set is partitioned according to size, with Δμm as the interval, for example, Δ=0.2. Each zone is regarded as a class, and the mean value of each zone is used as the label of the class, that is, the size is used as the label of the class. Finally, the size prediction problem is transformed into a classification problem for 35 classes (the size range of 5.5μm-12.5μm can be divided into 35 zones with an interval of 0.2μm, and the middle size of each zone is used as the class of the zone, that is, the label). The size interval and the number of classes here are selected based on the measurement accuracy of the system and the difficulty of network learning, and can be flexibly adjusted according to actual conditions.

[0050] The same method can be used to generate a test set. Obviously, the values of D, σ, and n are different in the training set and the test set.

[0051] The second step is to build a Bayesian neural network, that is, a deep learning classification network, and set the network type, number of network layers, and number of nodes in each layer. The specific steps are as follows:

[0052] (1) Establish a Bayesian deep learning classification network model, specifically a convolutional neural classification network.

[0053] The common network layers of a convolutional neural network are input layer, convolution layer, smoothing layer, dropout layer and output layer. Figure 2 As shown in (a), the number of network layers is set to input layer, four convolutional layers, one pooling layer, one convolutional layer, one smoothing layer (Global average pooling), one Dropout layer and output layer.

[0054] (2) Set the dimension of the input layer to the number of bands of the single scattering spectrum, the nodes of the output layer to the number of categories, and the Dropout ratio to r.

[0055] Specifically, in one embodiment, the input layer feature of the network is set to 101, that is, the number of spectrum bands, the number of convolution kernels of each convolution layer is set to 32, 64, 128, 256, and 256 from left to right, the convolution kernel size is set to 11, the Dropout ratio r is set to 0.1, and the number of output layer categories is 35.

[0056] The third step is to train the Bayesian neural network. The specific steps are as follows:

[0057] (1) Input the training set obtained by the present invention into the Bayesian neural network.

[0058] (2) The learning effect is judged based on the accuracy and loss of the validation set, and the learned network parameters are saved. In the present invention, 10% to 30% of the data in the training set is used as the validation set.

[0059] The fourth step is to construct a single scattering spectrum test set to evaluate the classification model. The specific steps are as follows:

[0060] (1) Based on the aforementioned physical model of single scattering spectroscopy, a test set is generated using free parameters D, σ, and n that are different from those of the training set.

[0061] In one embodiment, in order to make the test set and the training set non-overlapping, two free parameters are fixed and the other parameter is gradually changed, where the average diameter is 5.55-12.55 μm with a step size of 0.1 μm, the relative refractive index is 1.02-1.05 with a step size of 0.001, and the standard deviation is 1.3-1.6 μm with a step size of 0.01.

[0062] (2) The test set is input into the trained Bayesian neural network to obtain evaluation metrics such as accuracy and confusion matrix. If the evaluation metrics meet the requirements, the final classification model is obtained; otherwise, training continues. The input of the classification model is the single scattering spectrum, and the output is the cell nucleus size of the biological tissue.

[0063] The fifth step is to obtain the inversion results of the cell nucleus size. The specific steps are as follows:

[0064] (1) The single scattering spectrum of biological tissue is measured using a polarization gate scattering spectroscopy system, which is first normalized and then input into the obtained classification model. Dropout is turned on and predictions are made T times.

[0065] Obviously, in the measurement, the intensity of the incident polarized light, the scattering angle and other parameters should be consistent with the parameter settings of the generation environment in the first step. Polarization scattering spectrum experiments are performed on biological tissues to obtain single scattering spectra.

[0066] In the present invention, the biological tissue can be early digestive tract cancer tissue such as esophageal cancer, gastric cancer and colon cancer, or other human tissue. The tissue is first obtained from the human body and then tested in vitro.

[0067] (2) Perform statistical calculations on the T-time output of the classification model to obtain the average size of the cell nucleus and the model uncertainty.

[0068] Model uncertainty is the variance of the model predictions; however, in deep learning, model uncertainty is defined as follows:

[0069] Model uncertainty, also known as epistemic uncertainty, states that a model's own estimates of input data may be inaccurate due to factors such as poor training or insufficient training data, and are unrelated to any single piece of data. Therefore, epistemic uncertainty measures the uncertainty of the model parameters estimated by the training process itself. This uncertainty can be mitigated or even resolved through targeted adjustments (such as increasing training data).

[0070] The principle and process of obtaining the average size and model uncertainty of the present invention are as follows:

[0071] In a general Bayesian deep learning network, for a new input x*, the predicted posterior probability distribution p(y*|x*,X,Y) is:

[0072] p(y*|x * ,X,Y)=∫p(y * |x * ,w)p(w|X,Y)dw, (2)

[0073] Where p(y*|x*,w) is the probability of y* given the input x* and weight w, and p(w|X,Y) is the probability of weight w given the dataset (X,Y). To capture the uncertainty of the model, a simple variational distribution q(w) can be used to approximate p(w|X,Y). However, it is generally difficult to obtain p(w|X,Y) analytically. By performing Monte Carlo integration over T samples (the number of random forward propagations), Equation (2) can be approximated as:

[0074]

[0075] An approximation of the posterior probability distribution can be calculated by sampling multiple predictions with dropout turned on during the test phase. This allows for efficient Bayesian inference to be performed directly using existing networks. For a Bayesian classification network, the probability that a case belongs to a particular category is a distribution that describes all possible predictions given the network weights w and the input x*. The width of the predicted posterior distribution can reflect the uncertainty of the model. By performing T random forward passes on the Dropout network to obtain several random outputs, the final average μ of T predictions for an input can be expressed as:

[0076]

[0077] The variance of the T random prediction results will be considered as the uncertainty of the model prediction:

[0078]

[0079] like Figure 2 As shown in (b), in the Bayesian CNN classification network, the network outputs T probability values for each category (D1-DN). The mean value μ and the standard deviation σ of the probability are calculated according to equations (4) and (5). The category corresponding to the maximum μ is the final prediction of the network, that is, the size of the network prediction, and the standard deviation σ is the model uncertainty of the network.

[0080] In summary, the present invention ultimately predicts the size of the cell nucleus and further determines the model uncertainty. The significance of obtaining this result lies in its application in diagnosing early cancer. The specific steps of this application are as follows:

[0081] (1) Tissues are classified into two categories based on their average size: larger tissues are diagnosed as early-stage cancer, and smaller tissues are diagnosed as normal;

[0082] (2) Based on the size of the model uncertainty σ and the posterior distribution of the classification model prediction probability, the classification results of the tissue are screened, and tissues with large model uncertainty and relatively dispersed probability distribution are separated as further detection objects, and secondary diagnosis is performed using other diagnostic methods.

[0083] In summary, this invention combines polarized light scattering spectroscopy with Bayesian deep learning to predict cell nuclear size. By measuring the single-scattering spectrum of diseased tissue and using a trained Bayesian neural network to predict the input single-scattering spectrum, the tissue's cell nuclear size can be obtained. This Bayesian deep learning-based method achieves rapid inversion of cell nuclear size, improving the inversion speed and reducing the uncertainty of the prediction.

[0084] To verify the effectiveness of the present invention, the present invention utilizes a commonly used method in the field, that is, polystyrene microspheres close in size to the cell nucleus are suspended in a solution to simulate biological tissue, and then the size of the microspheres is predicted using the method of the present invention. As shown in Table 1, the predicted results of the polystyrene microspheres are compared.

[0085] Table 1. Prediction results of polystyrene microsphere diameter

[0086]

[0087] It can be seen that the prediction accuracy of the present invention is comparable to that of the existing least squares inversion method, and both are very close to the actual size, but the prediction time is much shorter than that of the existing method, and when the prediction accuracy drops, an uncertainty of corresponding size is provided to give a prompt.

Claims

1. A method for predicting cell nucleus size from polarized scattering spectra based on Bayesian neural networks, characterized in that: The steps include: Step 1: Generate a single scattering spectrum dataset of biological tissue using Mie scattering theory combined with a polarization gate scattering spectrum system, and divide it into a training set and a test set; generating the single scattering spectrum of biological tissue includes the following steps: Step 11: Based on the polarization gate scattering spectrum system and Mie scattering theory, a physical model of single scattering spectrum is established: Where ΔI is the single scattering spectrum, κ is the normalization factor, I0 is the intensity of the incident polarized light, θ is the scattering angle, θ0 is the angular range of the backscattered light collected, D is the diameter of the cell nucleus, ΔD is the size range of the cell nucleus diameter, s2 is the element of the scattering amplitude matrix, λ is the wavelength of the incident polarized light, n is the relative refractive index, and f(D) is the cell nucleus size distribution function. The cell nucleus size distribution satisfies the Gaussian distribution with a mean of The standard deviation is σ; Step 12: Determine the acceptable ranges of D, σ, and n based on the physical properties of the biological tissue, keep other parameters unchanged, and bring them into the physical model in step 11 to generate a single scattering spectrum dataset. Step 2: Construct a Bayesian neural classification network; Step 3, using the training set and the test set, training the Bayesian neural network and performing test evaluation to obtain a classification model, wherein the input of the classification model is the single scattering spectrum and the output is the cell nucleus size of the biological tissue; Step 4: Measure the single scattering spectrum of the biological tissue and input it into the classification model to obtain the cell nucleus size.

2. The method for predicting cell nucleus size based on polarization scattering spectrum using Bayesian neural network according to claim 1, characterized in that: In the dataset, the single scattering spectrum is normalized and used as the input item in the training set. D, σ and n are merged into one label, that is, only the size is used as the label of the training set. Then, the dataset is partitioned according to the size with Δμm as the interval, and each area is used as a category. The mean of each area is used as the label of the category, that is, the size is used as the label of the category. The same method is used to generate a test set, wherein the values of D, σ and n in the training set and the test set are different.

3. The method for predicting cell nucleus size based on polarization scattering spectrum using Bayesian neural network according to claim 1, characterized in that: The steps of constructing the Bayesian neural classification network in step 2 are as follows: Step 21, establish a Bayesian deep learning classification network model, where the network layers are set to input layer, convolution layer, smoothing layer, Dropout layer and output layer in sequence; In step 22, the dimension of the input layer is set to the number of bands of the single scattering spectrum, the nodes of the output layer are set to the number of categories, and the Dropout ratio is set to r.

4. The method for predicting cell nucleus size based on polarization scattering spectrum using Bayesian neural network according to claim 1 or 3, characterized in that: The steps of step 3, training the Bayesian neural classification network and performing test evaluation, are as follows: Step 31, inputting the training set into the Bayesian neural classification network; Step 32: judge the learning effect based on the accuracy and loss of the validation set, and save the learned network parameters; In step 33, the test set is input into the trained Bayesian neural classification network to obtain the evaluation index. If the evaluation index meets the requirements, the final classification model is obtained; otherwise, the training continues.

5. The method for predicting cell nucleus size based on polarization scattering spectrum based on Bayesian neural network according to claim 1, characterized in that: The step 4 comprises the following steps: Step 41, using the polarization gate scattering spectroscopy system, under the same conditions as those set in step 1, a polarization scattering spectroscopy experiment is performed on the biological tissue to obtain a single scattering spectrum, denoising and normalizing the single scattering spectrum, and then inputting the single scattering spectrum into the classification model to predict T times; Step 42: Perform statistical calculations on the T outputs of the classification model and take the average value as the cell nucleus size.

6. The method for predicting cell nucleus size based on polarization scattering spectrum using Bayesian neural network according to claim 1, characterized in that: The classification model also outputs model uncertainty.

7. The method for predicting cell nucleus size based on polarization scattering spectrum based on Bayesian neural network according to claim 1, characterized in that: The biological tissue is early cancer tissue of the digestive tract.

Images