Expandable convolutional neural network training method and CT image segmentation model construction method

Abstract

The invention provides an expandable convolutional neural network training method and a CT image segmentation model construction method. The expandable convolutional neural network training method comprises the steps of performing down-sampling on an original image to obtain training samples of different scales, performing convolution calculation on the training samples in an expandable convolutional neural network according to the sizes from small to large, and training expandable convolution kernel parameters; and after training of the training sample of each size is completed, expanding the expandable convolution kernel, and inheriting a result obtained by previous training to carry out fine training on the expandable convolution kernel. By adopting the training mode that the large-size data training process depends on the small-size data training result, a cascade relation of multi-step training is formed, the training speed of the network model can be remarkably increased, pre-training data does not need to be additionally prepared, the training process is suitable for various network models of different dimensions, and the segmentation precision of the network model can be improved on the premise of not using the pre-training model.

Description

technical field [0001] The invention relates to the field of machine learning, in particular to an expandable convolutional neural network training method and a CT image segmentation model building method. Background technique [0002] CT image analysis plays a vital role in the diagnosis and treatment of clinical medicine today. With the improvement of CT photography technology and corresponding algorithm performance, clinicians rely more and more on CT image analysis. However, CT image diagnosis is highly professional, and even simple image interpretation requires professional radiologists to spend a long time completing CT image analysis, delineation, and final report. Manual operation makes it difficult to further reduce the economic cost and time cost of CT diagnosis. In the process of CT imaging diagnosis, target delineation is the basic basis of imaging diagnosis and an important part of the diagnosis report. With the improvement of the performance of computer visi...

AI Technical Summary

Problems solved by technology

However, the expansion of network size will also lead to difficulties in network training.

Large-scale network models without proper training methods are very prone to overfitting

It is shown to have a very good segmentation effect on the training set, but is prone to serious mis-segmentation on the test set

Images