Method for fine segmentation of prostate and its internal lesion area based on large pathological section

Abstract

The application discloses a fine segmentation method for prostate and internal lesion areas based on large pathological sections, and specific steps are as follows: step 1, sequentially performing fixation, paraffin embedding, continuous transverse sectioning and HE staining operations on the whole tissue; step 2, extracting an HE staining image; step 3, sequentially scanning the pathological sections in step 2 into digital pathology; step 4, processing the digital pathology; step 5, performing image registration and three-dimensional image reconstruction on analysis results of multiple pathological sections processed in step 4; step 6, training a magnetic resonance multi-modal sequence segmentation model; step 7, extracting features of normal prostate tissue and lesion tissue in three sequences; step 8, acquiring high-risk factor information and quantifying features; step 9, constructing a sample feature matrix; and step 10, predicting the malignancy degree of prostate cancer. The method can realize more accurate benign and malignant evaluation of prostate lesions and prediction of the malignancy degree of prostate cancer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a fine segmentation method for prostate and its internal lesion area based on large pathological sections. BACKGROUND

[0002] Prostate cancer accounts for 7% of newly diagnosed cancers in men worldwide. There are 1.2 million new cases diagnosed each year, and more than 350,000 prostate cancer-related deaths, making the disease one of the leading causes of male cancer-related deaths. Digital rectal examination (DRE), serum prostate-specific antigen (PSA) level determination, and MRI are standard diagnostic tools for detecting prostate cancer. In the absence of prostate cancer, the results of DRE, PSA, and prostate MRI may be abnormal, and in the presence of prostate cancer, the results of DRE, PSA, and prostate MRI may be normal. There is no absolute linear relationship between the three examinations and prostate cancer.

[0003] In addition, the reading of prostate cancer MRI is still relatively complex and requires a lot of time, and there is a huge market for artificial intelligence diagnosis. At present, there are various methods to realize prostate MRI image segmentation, but it is still impossible to accurately identify the tumor area. The increase in the number of prostate biopsies and the shortage of urological pathologists worldwide have brought huge work pressure to pathology departments. In addition, the heterogeneity of prostate cancer can lead to over-treatment and under-treatment of prostate cancer. In order to alleviate these problems, it is necessary to develop an artificial intelligence (AI) system with clinically acceptable accuracy for prostate MRI artificial intelligence diagnosis, localization, and tumor grading.

[0004] Previous reports of prostate magnetic resonance image artificial intelligence identification only focus on imaging data. In fact, in clinical work, radiologists need to read the film in combination with clinical information. Therefore, it is necessary to combine clinical multi-modal data (psa, DRE, whether there is a tumor history, BMI) on the basis of magnetic resonance to train artificial intelligence, and finally to make a preliminary judgment and grading of prostate cancer by artificial intelligence for specific patients. According to the current prostate puncture indications (① rectal examination finds suspicious nodules of prostate, any PSA value; ② TRUS or MRI finds suspicious lesions, any PSA value; ③ PSA>10ng / ml, any f / tPSA and PSAD value; ④ PSA 4~10ng / ml, abnormal f / tPSA value and / or PSAD value), there are still a large number of patients who are negative after puncture, and these patients can avoid puncture; in addition, puncture is for patients with prostate cancer, and a part of the patients is clinically insignificant cancer, and these patients generally choose active observation and do not need special treatment, so there is a possibility of over-medical treatment for prostate puncture biopsy. Therefore, it is necessary to combine large section pathology, clinical multi-modal data and prostate magnetic resonance to re-determine the puncture indications according to the artificial intelligence method, and guide whether to perform prostate puncture biopsy.

[0005] In the nuclear magnetic resonance image, there is no obvious boundary between the normal prostate tissue and the cancer tissue, which leads to the inability to realize the pixel-level segmentation of the cancer tissue. If the malignant degree of prostate lesions is evaluated before operation, the usual method is to evaluate the entire prostate image or the approximate region of the suspected lesion of the prostate, but due to the inaccuracy of the selected image region, accurate benign and malignant evaluation cannot be achieved. The existing method is mostly the segmentation of the internal and external regions of the entire prostate, and the present application proposes a fine segmentation method based on large pathological sections of prostate and internal lesion regions. SUMMARY

[0006] The purpose of the present application is to provide a fine segmentation method based on large pathological sections of prostate and internal lesion regions.

[0007] To solve the above technical problems, the present application adopts the following technical solutions:

[0008] The present application provides a fine segmentation method based on large pathological sections of prostate and internal lesion regions, characterized in that the specific steps are as follows:

[0009] Step 1. Obtain the complete prostate tissue after prostatectomy, and sequentially fix, paraffin-embed, continuously transverse section and HE stain the entire tissue;

[0010] Step 2. Slice the whole prostate with a thickness of 5um and perform HE staining; every 100 slices, extract one HE staining picture as a 0.5mm-thick tissue feature picture; the extraction interval is 0.5mm;

[0011] Step 3. Scan the pathological slices in step 2 into digital pathology in sequence through a digital pathology whole slice scanner, and the number of scanning slices is 20-50;

[0012] Step 4. Process the above digital pathology through a pre-trained histopathological analysis framework, and identify the lesion area and normal prostate tissue in HE tissue staining;

[0013] Step 5. The analysis results of multiple pathological slices processed in step 4 are subjected to image registration and three-dimensional image reconstruction through a pixel point feature extraction method SIFT and edge features LoG operator and Robert operator; this step can obtain two three-dimensional images: a three-dimensional pathological HE staining picture of the whole prostate and a three-dimensional tissue benign and malignant area segmentation picture of the whole prostate;

[0014] Step 6. Use multiple nuclear magnetic resonance three-dimensional sequence pictures to train a nuclear magnetic resonance multi-modal sequence segmentation model; the model takes 3D U-Net as a basic framework, increases multi-modal interaction during up-sampling, so that the model can simultaneously refer to the image features of T2WI, DWI (ADC) and dynamic enhancement images in the arterial phase in the three sequences, four groups of images;

[0015] Step 7. Through an image feature extraction method of 100 radiomics features designed by artificial design, the features of the normal prostate tissue and the lesion tissue in the four groups of images in the three sequences are extracted, that is, 100*3*2=600 image features are obtained;

[0016] Step 8. Obtain the patient's digital rectal examination, serum prostate-specific antigen, whether there is a tumor history, BMI and perform feature quantization;

[0017] Step 9. Combine the features extracted in steps 7 and 8 to construct a sample feature matrix;

[0018] Step 10. Load the feature values in step 9 using an integrated learning method such as gradient boosting tree and perform machine learning classification algorithm training to predict the malignancy degree of prostate cancer.

[0019] The beneficial effects of the present application are that the method can achieve more accurate benign and malignant evaluation of prostate lesions and prediction of the malignancy degree of prostate cancer. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative effort based on these drawings.

[0021] Figure 1 The flowchart of the method for fine segmentation of prostate and its internal lesion area based on large pathological sections is provided in the present application. DETAILED DESCRIPTION

[0022] The technical solutions of the present application will be described clearly and completely in the embodiments of the present application. Obviously, the described embodiments only represent some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present application.

[0023] As shown in the method for fine segmentation of prostate and its internal lesion area based on large pathological sections, the specific steps are as follows: Figure 1

[0024] Step 1. Obtain the complete prostate tissue after prostatectomy, and sequentially perform fixation, paraffin embedding, continuous transverse sectioning (10 μm thick), and HE staining on the entire tissue.

[0025] Step 2. Slice the entire prostate with a thickness of 5 μm and perform HE staining; every 100 slices, extract one HE stained image (extraction interval 0.5 mm) as the 0.5 mm thick tissue feature map.

[0026] Step 3. Scan the pathological sections in step 2 into digital pathology (WSI) sequentially by a digital pathology whole section scanner (Whole Slice Image, WSI scanner), with a scanning number of 20-50.

[0027] Step 4. Process the above digital pathology by a pre-trained histopathological analysis framework (such as the open-source weakly supervised learning CLAM pathological section analysis framework), to identify the lesion area and normal prostate tissue in HE tissue staining (Patches method).

[0028] ​Step 5. The results of the analysis of the multiple pathological sections processed in step 4 are subjected to image registration and three-dimensional image reconstruction by pixel feature extraction methods SIFT and edge features LoG operator and Robert operator. This step can obtain two three-dimensional images: a three-dimensional stereoscopic pathological HE staining image of the whole prostate and a three-dimensional benign and malignant region segmentation image of the whole prostate.

[0029] Step 6. A magnetic resonance multi-modal sequence segmentation model is trained using multiple magnetic resonance (T1WI, T2WI, DWI) three-dimensional sequence images. The model uses 3D U-Net as the basic framework and increases the interaction between multiple modalities during up-sampling, so that the model can refer to the image features of the three sequences during segmentation. Unlike previous methods, the calculation of the deep learning loss function no longer refers to the artificial annotation results, but uses the three-dimensional pathological results in step 5 as the gold standard to divide the pixel points in the magnetic resonance images into three labels: background, normal prostate and lesion prostate. During model training, the gold standard of pathological tissue sections is used to guide the segmentation of normal prostate tissue and lesion tissue without introducing artificial annotation errors.

[0030] Step 7. Image feature extraction is performed on the normal prostate tissue and lesion tissue in the three sequences using 100 artificially designed radiomics features (including morphological features, statistical features, and texture features).

[0031] Step 8. Obtain patient digital rectal examination (DRE), serum prostate specific antigen (PSA), tumor history, and BMI (prostate cancer risk factors) information and perform feature quantification.

[0032] Step 9. The features extracted in steps 7 and 8 are combined to form a sample feature matrix.

[0033] Step 10. Load the feature values in step 9 and perform machine learning classification algorithm training using integrated learning methods such as gradient boosting trees (such as XGBoost or LightBoost) to predict the degree of prostate malignancy (pathological detection results). The prediction is that the prostate is benign or inflammatory, and no biopsy is needed; for low-risk patients, the guideline recommends active observation or no biopsy; for clinically significant prostate cancer, further treatment is recommended, and biopsy is recommended.

[0034] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for fine segmentation of prostate and its internal lesion regions based on large pathology slides, characterized in that, The specific steps are as follows: Step 1. Obtain the complete prostate tissue after radical prostatectomy, and sequentially perform fixation, paraffin embedding, continuous transverse sectioning, and HE staining on the entire tissue; Step 2. Slice the entire prostate with a thickness of 5 μm and perform HE staining; every 100 slices, extract one HE stained image as a 0.5 mm thick tissue feature map; the extraction interval is 0.5 mm; Step 3. Scan the pathological sections in step 2 into digital pathology in sequence through a digital pathology whole slide scanner, with a scanning number of 20-50; Step 4. Process the above digital pathology through a pre-trained histopathological analysis framework to identify the lesion area and normal prostate tissue in HE tissue staining; Step 5. Process the multiple pathological section analysis results in step 4 through the pixel point feature extraction method SIFT and the edge feature LoG operator and Robert operator for image registration and three-dimensional image reconstruction; this step can obtain two three-dimensional images: a three-dimensional pathological HE staining image of the whole prostate and a three-dimensional tissue benign and malignant area segmentation image of the whole prostate; Step 6. Use multiple three-dimensional sequence images of nuclear magnetic resonance to train a nuclear magnetic resonance multi-modal sequence segmentation model; this model uses 3D U-Net as the basic framework, and increases the interaction between multiple modalities during up-sampling, so that the model can simultaneously refer to the image features of the three sequences: T2WI, DWI-ADC, and dynamic enhancement image in the arterial phase image; Step 7. Use the image feature extraction method of 100 radiomics features designed by artificial design to extract features from the normal and lesion tissues of the four images in the three sequences, i.e. 100x3x2=600 image features; Step 8. Obtain the patient's digital rectal examination, serum prostate-specific antigen, history of tumor, and BMI and perform feature quantification; Step 9. Combine the features extracted in steps 7 and 8 to construct a sample feature matrix; Step 10. Load the feature values in step 9 using the gradient boosting tree method and perform machine learning classification algorithm training to predict the malignancy of prostate cancer.

Images