Difficult sample mining method and system, terminal and storage medium

Abstract

The invention relates to a difficult sample mining method and system, a terminal and a storage medium. The method comprises the following steps: collecting medical image data including K label types; constructing a point cloud shape model according to the labels of the medical image data, and selecting a point cloud shape model containing K label types as a template shape; randomly cutting a set number of training samples from the medical image data, inputting the training samples into the segmentation model for training, and updating the reward expected value of the template shape according to the training loss value of the training samples of the current batch after the training of one batch is finished; and mining a difficult sample from the medical image data according to the distribution of the reward expected values on the template shape, and training the segmentation model according to the difficult sample. According to the method, the difficulty degree of the training sample is abstracted into the reward value which can be estimated, so that the difficulty degree of the sample can be evaluated more scientifically, and the difficult sample which is more representative for the segmentation model can be mined more sufficiently.

Description

technical field [0001] The present application belongs to the technical field of medical image processing, and in particular relates to a difficult sample mining method, system, terminal and storage medium. Background technique [0002] In organ segmentation tasks based on three-dimensional medical images such as CT (Computed Tomography, computerized tomography) or MR (Magnetic Resonance, magnetic resonance examination), it is usually necessary to cut out 3D patches from the image (that is, the 3D part in the original 3D medical image) as samples to train the segmentation model. When sampling the original 3D medical image in the form of 3D patch, it will face the problem of sample selection. Commonly used random sampling strategies usually face the following two problems: First, different categories have different sampling probabilities, and categories with a large proportion of voxels will have a higher probability of being sampled, so that the final trained model is more ...

AI Technical Summary

Problems solved by technology

Commonly used random sampling strategies usually face the following two problems: First, different categories have different sampling probabilities, and categories with a large proportion of voxels will have a higher probability of being sampled, so that the final trained model is more accurate to background regions or large organs ( For example, the segmentation effect of the liver is better, but there are often missed detections for the segmentation of small organs (such as the adrenal gland); the second is that insufficient attention is paid to anatomical regions that are difficult to identify, such as low-contrast organ boundaries (such as the pancreas) or between individuals. Anatomical structures with large differences (such as the stomach) make the final training model miss or misdetect the segmentation of these regions

In this scheme, since the size of the original 3D medical image is usually large, there are many 3D patch samples that can be mined from the training set. A sampling probability is maintained and updated for each sample during the training process, which requires high memory and computing resources. , which is also not suitable for the field of medical image segmentation

Images