Research On Breast Cancer Detection And Lung Cancer Subtype Classification Based On Neural Network

Cancer poses a huge threat to human health,and the impact of its high mortality rate has received widespread attention from society.According to cancer statistics,lung cancer is the first cancer among men,while breast cancer is the first cancer among women.This paper uses the public data set of breast cancer and the public data set of lung cancer for research.In terms of breast cancer dataset,this paper uses the supervised learning method to determine whether each patient’s pathological image has a dichotomy problem of breast cancer cancer or non breast cancer.In terms of lung cancer data set,since lung adenocarcinoma and lung squamous cell carcinoma are two branches of lung cancer,this paper uses the method of weak supervised learning to judge whether each pathological image of lung cancer diagnosed is specifically the dichotomy between lung adenocarcinoma and lung squamous cell carcinoma.In this paper,the dichotomy of pathological images is realized from two aspects: supervised learning and weak supervised learning.For breast cancer tumor detection based on supervised learning,preprocessing is carried out first,and low pixel small cut blocks are cut from the slide.Extract useful information through mask technology.Randomly select points from areas with training value in the slides,and randomly select 100 points from each slide to generate 100 low pixel small blocks.This article uses Efficient Net-B0 as a feature extractor to establish a binary prediction model for low pixel small blocks,enabling CNN to recognize the probability of small blocks being cancer or non cancer.Then,random forest is introduced into the slide classifier by generating the probabilistic thermogram,and the features in the probabilistic thermogram are extracted through machine learning.The classifier that classifies the whole slide mainly uses feature extraction to extract morphological features from the thermogram as the basis to judge whether each slide has a breast cancer tumor,so as to realize the prediction of slide pathological images.Based on the research of lung cancer subtype classification with weak supervised learning,the preprocessing also generates low pixel blocks to facilitate subsequent learning and training.Generate probability heat maps by using attention based learning.Subsequently,transfer learning and pre trained convolutional neural network encoder Res Net50 were used to extract image features of slides.Convert each small block into1024 feature vectors,which is a low dimensional feature embedding set.Then 1024 feature information is input to the classifier of the whole slide based on random forest,and the slide classifier is responsible for the second classification of lung cancer subtypes of the whole pathological picture.The deep learning framework based on cluster constrained attention multi instance learning classifies low pixel small block level features into high pixel slide level representations using attention based pooling functions.During the training at the high pixel level,the model checks all the slices in the organizational area of the entire slide and ranks them,assigning an attention score to each small block.These attention scores are converted into quantile scores with scores ranging from 0 to 1(where 1indicates the most concerned and 0 indicates the least concerned).By having two parallel attention branches,they together calculate two unique slide level representations,each of which is determined by a different set of highly focused areas in the image viewed by the network.Then,the classification layer checks each class specific slide representation to obtain the final probability score prediction of the entire image,achieving the binary classification task for lung adenocarcinoma and lung squamous cell carcinoma.