Predicting and Segmenting Colon Cancer Grade Using Deep Learning

Abstract

The invention relates to a method for examining colon cancer tumors using deep learning.

Description

1 Specifications Predicting and Segmenting Colon Cancer Grade Using Deep Learning Technical Area The invention uses a deep learning model to analyze colon 5 digital histopathology images. with a method that can estimate the pixel locations of cancer tumors and the cancer grades It is related. Previous Technique In cancer diseases, early diagnosis and active monitoring of the disease's progression are crucial. This is critical for the treatment phase. Patent document number US11170897B2 digital images 10 It relates to a method that would provide a tissue biopsy using patent US11170897B2. In the method described in the document, the entire slide image is subjected to image magnification, The magnified images are divided into sub-parts. These sub-parts represent the tumor regions. It feeds a deep learning model that helps in detection. The aforementioned invention is based on... Because it uses fragmented parts and cannot perform an analysis for every pixel in the image, the measurement taken is 15. The accuracy level can be improved. The invention, application number US2023267606A1, involves detecting cancer from sample images. It mentions a related deep learning method. The deep learning method in question is two-layered. It is designed in such a way that the aforementioned deep learning method has the number US11170897B2. It feeds the deep learning method by breaking down images into sub-components in a way that is similar to invention. 20 Application number US11954593B2 concerns cancer analysis using an end-to-end learning system. It is related to the estimation method. In this method, even though the slide image is divided into subgroups, the subgroups An abnormality score is obtained by classifying the groups. This Tumor testing can be performed using an abnormality score. All the problems mentioned above require an innovation in the technical field. 25 2 Brief Description and Objectives of the Invention The main purpose of the invention is to develop a method for examining the grade of colon cancer. It is related. Another aim of the invention is to learn images at the pixel level in learning algorithms. The aim is to reduce potential information loss by using visual information in images. Part 5 Because it doesn't just stop at the pixel level, but uses it down to the pixel level, it's better than its other alternatives. The accuracy level has increased compared to the previous comparison. Colon cancer is diagnosed and treated based on its diverse pathophysiologies. Certain difficulties are encountered in this regard. Biopsies taken using invasive methods. By examining them, the possible presence and progression of cancer can be determined. 10 More than 90% of colorectal carcinomas originate from epithelial cells of the colorectal mucosa. These are adenocarcinomas originating from and as a result, CRC, which is an adenocarcinoma, is a type of adenocarcinoma, originating from the 1st, 2nd, and It is divided into 3 main groups, including 3rd degree. Conventional adenocarcinoma, histological tumor It is characterized by glandular formation, which forms the basis of its grading. Histopathological grades can be observed differently by different observers, this 15 This prolongs the process and investigation level, creating a disadvantage in situations requiring urgent investigation. is happening. Another aim of the invention is to develop a method capable of processing histopathological images. Illustrations and descriptions illustrating the invention. Explanations of the figures used to better understand the invention are given below. 20 Figure 1. Different deep learning models and tumor analysis in different biopsy samples. Detailed Description of the Invention The invention uses a deep learning model to analyze an image obtained from a digital scanner. segmentation and grading of tumor regions at the pixel level 25 It is related. Deep learning methods for rapidly examining histopathological conditions, digital pathology. It enables progress in the field. Deep learning methods are a key element in the field of digital pathology. It has made progress. 3 The dataset used to develop the deep learning model used in the invention was surgical data. Samples obtained by removal are fixed using formalin and a machine It is processed with the help of paraffin blocks of 4 microns, taken from the tissue sample. The samples are prepared to a certain thickness. The prepared samples are similar to classical histology practices. They are stained with hematoxylin and eosin compounds. Then the stained samples are digitally analyzed using a 5-bit system. After being digitized by a scanner, it is used to train a deep learning model. The dataset used for Colorectal Cancer Tumor Grading Segmentation has a total of It contains all 103 slide images. Ground-truth is used for these images. The annotations were created by two independent pathologists. The dataset, annotations, 1. Grade 1, Grade 2, and Grade 3 tumor classifications, and normal 10 For the first class, use "Normal-mucosa," and for other areas outside of this, use "Others." It includes pixel-based segmentation masks. In training, the prepared dataset is 70%. The dataset is divided into three parts: 15% for training, 15% for validation, and 15% for testing. To determine the baseline results, leading convolutional neural networks (CNNs) and Transformer-based models have been trained and evaluated. 15 Referring to Figure 1, the results of the method in different examples and different deep learning systems The results are visible. The results show that the transformer-based model called SwinT achieved an average dice score of 63%. that it surpassed other transformer-based models and all CNN-based models, and With the recent success of transformer-based models in the field of computer vision, 20 This shows that it is compatible. The first step of the method mentioned is to scan hematoxylin and a similar device using a digital scanner. A digital image of the tissue sample is obtained using an eosin-stained sample. (The aforementioned...) The image is magnified 40 times, resulting in an image with approximately one hundred thousand pixels by one hundred thousand pixels. is obtained. Then the digital image is reduced in size and divided into parts. These parts are 25 By feeding a deep artificial neural network, it can identify tumors and their grades in the image at the pixel level. is determined. Then, the results obtained from each part are combined to create a complete image. Segmentation and grading results are obtained. The invention preferably targets the region where the tumors are located. This is shown using a heat map. 4 In digital pathology, gigapixel-sized WSI processing presents challenges. The invention solves this problem by using a subsampling method. The aforementioned subsampling method... The sampling rate can be calculated as the ratio of the width and height into which the image is divided. (Depth) The learning method uses a hyperparameter between the values ​​20, 40, 60, and 80. It has been calculated. In addition, ground-truth scaling is also done using the same deep learning network 5. It is scaled by. For parameter selection, DeepLabv3+, UNet, SwinTransformer, SegFormer and ConvNext techniques can be used optionally. The Deeplabv3+ technique has an optimization with a 0.9 momentum and 0.0001 weight reduction. The Stochastic Gradient Descent (SDD) technique was used. The difference between the data in the categories is calculated as 10. The mathematical formula used to balance this is called weighted cross-entropy loss. It has been used. The formula mentioned can be explained as follows. = − w represents the data distribution across categories, i is the index of different classes, t is the minimum accuracy value, and p is the index of different classes. C represents the number of categories, and C represents the number of categories. This is necessary to find the data distribution. The formula can be defined as follows: 15 = ∑∑ nj represents the number of pixels in the j-th image, and nij represents the pixels in the j-th category of the image. In the UNet technique, an SGD optimization is performed using a momentum value of 0.9. is used. The loss incurred during the process is not specifically stated for each category. It is defined as follows: = ∑∑ p represents the estimated probability, and g represents the accuracy class. 20 In the SwinTransformer model, a SuperNet framework is used, with a size of 224x224. Images are selected. When this method is used, the ADAM optimization process has a momentum of 0.9 and 10- It uses 4 intensity reduction settings. The Swin transformer method is also used in SegFormer and ConvNext parameter selection models. Similarly, the ADAM optimization process is used. 5 In a preferred configuration of the invention, a scanner with a resolution of 0.25 µm / pixel. It has been used. The deep learning model in question could preferably be a CNN or a transformer model. Different metrics are also used to measure the performance of a deep learning model used in the invention. It is available. These metrics evaluate images by classifying them into 5 different categories. 10 The formulas used to calculate the aforementioned metrics can be seen below. = + = + = 22 + + , and the pixel counts as true positive, false positive and false positive respectively. Negative numbers represent their values. The average values ​​of metric values ​​are: It can be seen below. = 1 = 1 = 1 6 In a preferred configuration of the invention, the self-attention contained within the transformer models Thanks to this mechanism, the model can be improved.

Claims

7 REQUESTS 1. A deep learning method to determine tumor grade of cells in colon cancer. It is a method, and its characteristic is; • Obtaining a digital image of the tumor tissue sample, 5 • Applying a magnification factor to the aforementioned digital image, • Reducing the size of the digital image, • Breaking down a digital image into its constituent parts at the pixel level, • Feeding the components into the deep artificial neural network model, • Deep artificial neural network modeling of tumors at the pixel level and 10 examining their degrees, • In order to obtain segmentation and grading results, the aforementioned tumors and Combining the results obtained for the degrees and the entire image It is obtained as a result of segmentation and grading.

2. This is a method according to Claim 1, the characteristic of which is that tumors are identified when images are combined. It is the representation of regions using a heat map.

3. This is a method according to Claim 1, and its characteristic is that it uses the DeepLabv3+ model in parameter selection. It is the use of. 20 4. A method according to claim 3, whose characteristic is data imbalance between categories. The formula for its removal: = − w represents the data distribution in categories, i is the index of different classes, and t is 25. Weighted, where p represents the minimum accuracy value, p represents the probabilities, and C represents the number of categories. This involves using Cross Entropy Loss. 8 5. It is a method according to claim 1, and its characteristic is that it uses UNet, SegFormer, in parameter selection. This involves using either the SwinTransformer or ConvNext model.

6. It is a method according to claim 1, and its characteristic is that the correctness of the method... = + = + = 22 + + It is checked using formulas.

7. It is a method according to Claim 1, and its characteristic is that it models the aforementioned Deep Artificial Neural Network. The annotations refer to Grade 1, Grade 2, and Grade 3 tumor classes, and 10 for the normal class. Pixel-based segmentation for "Normal-mucosa" and "Others" for other regions. It is trained with a dataset that includes masks.

8. It is a method according to claim 7, and its characteristic is that it is a deep learning method mentioned. It is the CNN method. 15 9. It is a method according to claim 1, and its characteristic is that the aforementioned deep learning method is a It is a transformer model.

Images