Deep medical image clustering method based on multi-scale structure learning

Abstract

The invention discloses a deep medical image clustering method based on multi-scale structure learning, and the method comprises the steps: combining the representation learning of image data with the global multi-scale structure learning between images; the deep medical image clustering method capable of performing multi-scale structure learning is jointly constructed by a convolutional neural network, an auto-encoder and an image convolutional network, so that the performance of deep image clustering is improved. According to the method, instance-level contrast learning, global clustering structure learning and multi-scale neighborhood structure learning are used for constructing a unified depth image clustering framework for the first time, and neighborhood structure information between images under different scales and representation information of image data are cooperatively learned and updated; the robustness and confidence of the feature information are effectively improved, and the defect that only single-scale neighborhood structure information is used in the prior art is overcome.

Description

technical field [0001] The invention relates to the fields of deep learning and image clustering, in particular to a deep medical image clustering method based on multi-scale structure learning. Background technique [0002] In the daily outpatient clinic of the hospital, a large number of patients take medical images during the examination process every day, and the doctor needs to spend a lot of time analyzing and diagnosing these images. The problem of visual fatigue caused the doctor to make mistakes in judgment and delay the treatment of the patient. For example, when diagnosing ear diseases, taking otoscope images is an important diagnostic link, but there are still few cluster analysis techniques for a large number of otoscope images accumulated in hospitals. Combined with deep image clustering technology, unsupervised classification of a large number of otoscope images can be used for doctors to diagnose ear diseases (such as cholesteatoma, otitis media, upper tympa...

AI Technical Summary

Problems solved by technology

Structural Deep Clustering Network (SDCN) uses a single K-nearest neighbor graph to express single-scale neighborhood information, but on the one hand, it does not use a convolutional network layer and uses vectorized data features as input, which is difficult to apply On the other hand, it only uses single-scale neighborhood information without considering the multi-scale neighborhood information between samples, and it is impossible to carry out collaborative learning of multi-scale neighborhood structures.

[0005] To sum up, the existing deep image clustering methods mainly consider the characteristic information of the image data itself; some methods consider the fixed (single) neighborhood information between data samples, but fail to break through the single-scale neighborhood structure, and do not Ability to use graph neural networks including graph convolutional networks for multi-scale neighborhood structure learning between image data

How to combine the representation learning of the image itself with the neighborhood structure learning between images based on the deep learning framework, and extend the single-scale neighborhood structure information to the multi-scale neighborhood structure learning to achieve higher quality clustering is currently still an issue. no solution

Currently there is no deep clustering technology solution for medical images

Images