Medical image processing method and device, electronic equipment and medium

Abstract

The embodiment of the invention discloses a medical image processing method and device, electronic equipment and a medium, and the method comprises the steps: inputting a to-be-detected original image into a feature extraction model, generating a feature map sequence of multiple stages, and selecting a feature map sequence of one stage as a to-be-detected feature map sequence; inputting the to-be-detected feature map sequence into the classifier, generating a prediction probability value of each category, and determining a target category value; generating a corresponding mask input image based on the to-be-detected feature image and the to-be-detected original image of each dimension; based on the mask input image of each dimension and the target category value, obtaining the weight of the to-be-detected feature image of each dimension for the to-be-detected original image; and generating a target visualization result according to the to-be-detected feature images of all dimensions and the weight of each to-be-detected feature image for the to-be-detected original image. The accuracy of early cancerization lesion area prediction is improved, and a lesion explanation basis can be provided for each prediction stage.

Description

technical field [0001] The present application relates to the technical field of artificial intelligence, and in particular to a medical image processing method, device, electronic equipment and media. Background technique [0002] Due to heredity, gene mutation, long-term eating habits, work and rest habits, etc., the human body may develop malignant cancers, such as gastric cancer and midgut cancer. These cancerous tissues can be divided into early stage, advanced stage and late stage according to the lesion stage, and different stages correspond to different treatment methods. Among them, early detection of cancer has the least difficulty in treatment, and the patient has the greatest chance of recovery. In the prior art, medical imaging images are usually used to detect early cancers, but since early cancer cells and non-pathological cells have very inconspicuous features in shape and color, the diagnosis of early cancers is a fine image detection problem, and it is dif...

AI Technical Summary

Problems solved by technology

But at least it has the following disadvantages: First, in order to obtain the weight corresponding to the feature after global average pooling, the classifier of the model must use the fully connected layer in the neural network, which limits the scalability and versatility of the model

Second, this method only considers the influence of the feature map before global average pooling on the classification results, but the difference between a large number of lesions and normal cells is very subtle, and the proportion of lesions is low, so the prediction accuracy of lesions is low.

Images