Medical image small lesion segmentation method

Abstract

The invention discloses a medical image small lesion segmentation method, which comprises the steps that a segmentation network is formed by attention modules of a rough segmentation first stage, a refined second stage and a segmentation error region, and five-fold cross validation is used in the training of the first stage; during cross validation, each sample of the training set is contained inthe validation set, and all the samples have an opportunity to be regarded as data of the validation set and are tested on a model trained on the corresponding training set; a prediction result of thefirst stage is compared with a real segmentation result to obtain a difference which reflects a difficult-to-predict part of the model, and mismatching information is taken as supervision of the second stage; and in the second stage, the information enhancement features are input into a DA module, and an attention mechanism is used, so that the segmentation precision of the network is enhanced.

Description

technical field [0001] The invention belongs to the field of semantic segmentation of computer vision, and relates to a method for segmenting small medical image lesions. Background technique [0002] Semantic segmentation algorithm: The semantic segmentation algorithm is aimed at the classification of each pixel on the image, which is a pixel-level problem. Therefore, a label needs to be attached to each pixel of the image in the training set. Expressed by a formula: from the label space L={l 1 , l 2 , l 3 ,...,l k} represents a set of random variables X={x 1 ,x 2 ,x 3 ,...,x N}. Each label l represents a different class or object, e.g., airplane, car, traffic sign, etc. This labeled space has k possible states, which are usually extended to k+1, with l and 0 as the background or empty class. x represents the pixel of the image, and the number of pixels is N. The currently widely used semantic segmentation networks are all improved based on FCN. The FCN network ...

AI Technical Summary

Problems solved by technology

[0013] In order to solve the technical problem of insufficient 3D small target segmentation accuracy in the existing semantic segmentation technology, the present invention proposes a 3D convolutional network DANet combined with a distraction mechanism, which can better segment lesion regions

Images