PCR prediction method based on multi-type medical data

Abstract

The invention discloses a pCR prediction method based on multi-type medical data. The method includes the following steps that: clinical data, CT diagnosis reports and enteroscope images are obtainedfrom a medical department; an SVM is trained by using the clinical data, a BERT model is obtained by means of transfer learning training with the CT diagnosis reports, and a Faster-RCNN model is obtained by means of transfer learning training on the basis of the enteroscope images; and the clinical data, CT diagnosis report and enteroscope image of a patient to be predicted are inputted into the three trained models, predicted pCR probabilities p1, p2 and p3 are obtained, the predicted pCR probabilities p1, p2 and p3 are fused, a final predicted pCR probability p is obtained, and if p is greater than a set threshold T, it is predicted that the patient has pCR. According to the method, the Faster-RCNN network is used; a tumor ROI can be automatically generated; a whole process does not needmanual intervention; and prediction efficiency is improved. The neural network is used for representation learning; manual setting and feature selection are not needed; and therefore, the accuracy and efficiency of prediction are improved. PCR prediction is carried out by combining the clinical data and the CT diagnosis report of the patient, so that prediction accuracy is improved.

Description

technical field [0001] The invention relates to the fields of machine learning, computer vision and natural language processing, in particular to a pCR prediction method based on multi-type medical data. Background technique [0002] The standard treatment for advanced mid-low rectal cancer is to give patients neoadjuvant chemoradiotherapy first, followed by radical surgery, that is, radical resection of the lesion area. Doctors conducted pathological examinations on the surgically excised lesions and found that about 10% to 20% of the patients were cured after chemotherapy, that is to say, unnecessary operations were performed. If patients with pCR (pathological complete remission) after neoadjuvant therapy can be found before surgery, so that they can avoid unnecessary surgery and adopt a "wait&see" strategy, it will be of great significance to patients. However, at present, there is no uniform standard in medicine for the pCR determination of rectal cancer patients after...

AI Technical Summary

Problems solved by technology

[0007] 1. To obtain the ROI of the tumor, manual segmentation is required. The requirement is to include the complete tumor area and exclude the intestinal tract. Therefore, the labeling personnel need to have high medical professionalism. The labeling process is time-consuming and laborious, which reduces the prediction speed of the entire model. Moreover, the accuracy of segmentation will affect the subsequent prediction effect;

[0008] 2. The features extracted by this technology from the ROI of MRI images are all artificial features, such as statistical features, voxel grayscale features, and wavelet features. In other words, it is difficult to represent the core and most prominent features of MRI images, which will affect the subsequent prediction effect;

[0009] 3. The scheme relies too much on the features of MRI images. The pCR prediction model finally established uses 30 image features retained after feature selection, and only uses one clinical feature, that is, the lesion diameter after neoadjuvant treatment, without mining and use more clinical features, which may degrade subsequent predictive performance

Images