Deep learning-based cancer prognosis survival prediction method and device, and storage medium

A deep learning-based cancer prognosis survival prediction method, comprising:

• acquiring sample data, wherein the sample data comprises pathological image data and clinical data of a sample;
• training and evaluating a deep learning classifier model on the basis of the sample data to obtain an optimal classifier model;
• using the optimal classifier model to preprocess the sample data, extracting data features of cancerous regions in the pathological image data and obtaining a first preprocessed data for each sample, converting the clinical data for computer language recognition and obtaining a second preprocessed data for each sample;
• training and evaluating a prediction model by using all the first preprocessed data and all the second preprocessed data to obtain an optimal prediction model;
• performing risk prediction on a new sample on the basis of the optimal classifier model and the optimal prediction model;
• wherein the step of “training and evaluating a deep learning classifier model on the basis of the sample data to obtain an optimal classifier model” comprises:
• training and evaluating the classifier model by a weakly supervised deep learning and with the pathological image data of the sample as input data, to obtain the optimal classifier model, wherein the pathological image data comprises cancerous pathological sections and non-cancerous pathological sections collected in a predefined ratio;
• wherein the step of “extracting data features of cancerous regions in the pathological image data and obtaining a first preprocessed data for each sample” comprises:
• inputting cancerous pathological sections into the optimal classifier model, and selecting the first k image regions with the highest cancer probability from each cancerous pathological section to obtain the first preprocessed data;
• wherein the prediction model comprises two layers, and the process of training and evaluating the prediction model using the first preprocessed data and the second preprocessed data comprises:
• in a first layer of the prediction model, using the optimal classifier model with a fully connected layer removed as a feature extractor, extracting a feature vector from each image region of the first preprocessed data, wherein the feature vector extracted from each image region is a first feature vector;
• compressing the first feature vector using a multilayer perception deep learning algorithm with two hidden layers, obtaining a single feature value with a length of 1 for each first feature vector, and horizontally splicing the single feature value and the second preprocessed data to obtain a second feature vector;
• in a second layer of the prediction model, using an optimization training based on a negative log-likelihood loss function and a concordance index as the performance metric for the prediction model, outputting probability values by an activation function after compressing each second feature vector, wherein the average of all probability values is a prognosis risk indicator for each sample.