Brain CT medical image processing method based on unsupervised feature matching

Abstract

The invention relates to a brain CT medical image processing method based on unsupervised feature matching, and the method specifically comprises the following steps of: S1, acquiring a large-scale data set where a brain CT image is located, carrying out the extraction of data features according to a pre-training network model on the large-scale data set, and carrying out the dimension and noise reduction processing on the data features; S2, carrying out distance calculation on the processed brain CT image, completing feature matching of data features, and carrying out calculation on an unlabeled brain CT image through a depth active learning algorithm of uncertainty sampling to obtain an uncertainty calculation result; and S3, according to a feature matching result of the data features, performing resampling on an uncertainty calculation result, and performing calculation through an improved depth active learning algorithm of uncertainty sampling to obtain a brain CT medical image including rich value information. Compared with the prior art, the method has the advantages of improving the stability and efficiency of the image processing result of the brain CT image in the face of large-scale unlabeled data, and the like.

Description

technical field [0001] The invention relates to the technical field of medical image processing, in particular to a brain CT medical image processing method based on unsupervised feature matching. Background technique [0002] Active learning aims to sample the most valuable data for network training from large-scale unlabeled data, so as to improve the learning efficiency and quality of the entire deep learning network model as much as possible under the premise of limited labeling costs. However, traditional active learning algorithms, especially deep active learning algorithms based on uncertainty sampling, are prone to data bias when faced with large-scale unlabeled data, that is, the sampled data cannot well represent the original unlabeled data. . In special cases, the performance of traditional deep active learning algorithms based on inaccurate sampling is even lower than that of random sampling when they face large-scale data. Although there are currently a lot of...

AI Technical Summary

Problems solved by technology

Although there are currently a lot of literature, work, etc. aimed at alleviating the data bias problem that uncertainty sampling is prone to produce, such methods usually face very high computing power costs and need to design complex training losses, so such methods face different Inability to show versatility when tasking

[0003] In the current field of medical image processing, active learning algorithms are widely used, but there are still problems with the learning efficiency and quality of the learning network model submitted above, especially for the processing of brain CT images. The number of images is large, but the difference between each image is small, and the accuracy of the image itself is not high, so when the learning efficiency and quality of the learning network model are not high, the final processed brain needs to be put into use. Poor accuracy of CT images

Images