Prediction Of KRAS Gene Status Based On MRI In Rectal Cancer

Rectal cancer,the third largest cancer in the world,poses a huge hidden danger to human life.The treatment of rectal cancer is divided into two clinically based on whether there is KRAS gene mutation: whether to use targeted drugs for treatment.This is because studies have confirmed that patients with rectal cancer with mutations in the KRAS gene are highly resistant to targeted drug treatments.Therefore,it is very necessary to be able to accurately determine the patient’s KRAS gene status.Traditional genetic testing methods mainly include biopsy and pathological testing of postoperative samples,which are regarded as the gold standard for genetic mutation status testing.However,the sampling process of biopsy and postoperative samples can cause harm to the patient’s body,and the biopsy is random sampling,which may cause unstable results due to different sampling locations.With the development of deep learning in recent years,predicting the KRAS gene status of patients with rectal cancer based on artificial intelligence and image processing has become a possible research direction.This article is mainly based on rectal cancer MRI data and the latest deep learning methods to explore ways to improve the accuracy of using medical images to predict the KRAS gene status of rectal cancer.In response to the above-mentioned problems,this article carried out some studies based on deep learning methods to predict the mutation status of KRAS gene based on T2-weighted MRI data of patients with rectal cancer:(1)Convolutional neural network based on attention mechanism predicts the mutation status of KRAS geneCompared with the manual features extracted by image genomics,the depth features extracted by convolutional neural networks can search for a larger feature space,and it is easier to find hidden features that people have not discovered.We try to combine deep learning to establish the correlation between gene mutations and MRI data of patients with rectal cancer,and ultimately improve the model’s ability to predict the mutation status of KRAS gene.In Chapter 3,this paper extracts a multi-branch convolutional neural network based on cross-channel attention,which can reduce the probability of model overfitting and improve the model’s prediction of KRAS gene mutations by learning the common features between different image modalities.Performance.In addition,we propose an inter-branch loss for the multi-branch model,which is used to constrain the neural networks of different branches to pay attention to the common image features in the data and avoid the model from overfitting.In the end,the model proposed in this paper achieves an accuracy of 88.92% on the MRI data of rectal cancer collected by the cooperative hospital,and has a greater advantage than other models in terms of performance.(2)Prediction of KRAS gene mutation status based on shape statistical model and generative confrontation networkIt is very important to make full use of the annotation data of the imaging physician.In order to make full use of the labeled data,in Chapter 4,this paper proposes a generative confrontation network based on the shape statistical model to generate new self-labeled rectal cancer lesion data.The shape statistical model is mainly to establish a general model of the rectal cancer lesion contour that we have extracted,so that more new regular lesion contours can be generated according to the model,and then effective rectal cancer lesion data can be generated.However,the lesion data generated at this time is limited by the instability of the model generation against network training and the excessive amount of model parameters,and the size of the lesion data is only 64×64.Therefore,this paper further proposes a multi-stage generative confrontation network to enlarge the lesion data to the size of the actual lesion image.Subsequently,we use the UNet segmentation network to segment and train the data generated by the generation confrontation network,and save the model weight of the network as the pre-training model of the classification network.In the end,our proposed segmentation promotion classification model can effectively improve the prediction accuracy of KRAS gene mutation in rectal cancer to 90.11%,which shows that the method proposed in this paper is effective.