Liver CT image segmentation system and algorithm based on mixed supervised learning

Abstract

The invention discloses a liver CT image segmentation system and algorithm based on mixed supervised learning, the image segmentation system comprises an image preprocessing unit, a feature extraction unit, a word vector segmentation unit and a single-layer convolution classification unit, and the image preprocessing unit is in data connection with the feature extraction unit. The feature extraction unit is in data connection with the word vector segmentation unit and the single-layer convolution classification unit. A multi-task framework is used for carrying out segmentation and classification tasks respectively, high segmentation precision is achieved through a large amount of weak label data and a small amount of strong labels, network parameters shared among multiple tasks are increased as much as possible for the problem that the independent parameter quantity among the multiple tasks is large, and a learnable word vector v is used for achieving the segmentation tasks; the number of independent parameters is reduced, the result consistency among multiple tasks is improved, high performance can be kept for classification and segmentation tasks, and good consistency is embodied.

Description

technical field [0001] The invention belongs to the interdisciplinary field of medicine and computer science, and in particular relates to a liver CT image segmentation system and algorithm based on hybrid supervised learning. Background technique [0002] CT scanning is a routine medical examination method. It uses X-rays and detectors to scan sections around the human body, and reconstructs the detection information into a cross-sectional view, that is, slices, with the help of a computer. Multiple slices can be combined to form a three-dimensional view of the internal organs and tissues of the human body. In abdominal CT images, the pixel-level segmentation of the liver is helpful for the pathological diagnosis of liver diseases, pre-operative planning and post-operative evaluation, etc., and is of great significance for the treatment and research of liver diseases such as hepatitis, cirrhosis, and liver cancer. The pixel-level segmentation of the liver requires pixel-lev...

AI Technical Summary

Problems solved by technology

[0004] The problem with the above-mentioned supervised learning segmentation method is that in the process of making the training data set, pixel-level image annotation requires a lot of time and labor costs, and at the same time requires doctors with rich medical experience to carry out quality control, which makes it difficult to obtain strong labels. Big

However, the improvement of this method is limited, and a large amount of valid data except data with strong labels is wasted

Therefore, in recent years, many studies have proposed semi-supervised learning methods. For unlabeled images, methods such as generative adversarial learning and knowledge distillation are used to generate pseudo-labels to expand the training data set. However, these methods cannot avoid the impact of wrong labels on network learning. It is difficult to further improve the image segmentation accuracy

In this regard, some studies have proposed to use hybrid supervised learning to improve the generalization ability of the network by using weak labels whose labeling cost is much lower than that of strong labels. Weak labels do not require doctors to judge and label the precise location of lesions, and the cost of manual labeling is greatly reduced. And it can be obtained in large batches, and because weak labels have accurate image category information compared to no labels, they can better provide effective supervision and achieve a balance between segmentation accuracy requirements and labeling costs, but existing hybrid supervision methods usually Using a multi-task framework, multiple tasks will cause inconsistent multi-task results due to many independent parameters. The classification task affects the feature representation of the network model to a certain extent, which in turn affects the segmentation accuracy.

Images