Training method and segmentation method and apparatus for a segmentation learning network of a 3D image and medium

Abstract

The present disclosure relates to a training method and a segmentation method and apparatus for a segmentation learning network of a 3D image and a medium. The 3D image comprises a target object and the proportion occupied by the target object is lower than a preset threshold value. The method comprises the following steps that: a segmentation learning network is constructed based on the sequential combination of a plurality of dense blocks, dense connection exists among basic units in the dense blocks, and each basic unit consists of a batch normalization layer, a RELU layer and a rolling layer; a segmented learning network is trained by a processor based on a training dataset of the 3D image using a loss function that focuses more on difficult samples and penalizes negative voxels that are far away from the object of interest. The method can rapidly and accurately segment irregular and small attention objects by using the learning network with a more compact structure and fewer parameters, and the training process of the learning network can solve the unbalance problem of samples and foreground and background and avoid the problem of over-fitting caused by lack of training samples as much as possible.

Description

[0001] cross reference [0002] This application claims the priority of the U.S. Provisional Application No. 62 / 675,765 filed on May 24, 2018, the entire contents of which are incorporated herein by reference. Technical field [0003] The present disclosure generally relates to image processing and analysis. More specifically, the present disclosure relates to a training method of a learning network for 3D image segmentation, a 3D image segmentation method, and a segmentation device, and a non-transitory computer-readable medium having a corresponding program stored thereon. Background technique [0004] Cancer is one of the major diseases facing human beings, and early detection of cancer can greatly increase the survival rate. However, early-stage tumors are usually irregular in shape and account for a small proportion of corresponding medical images. Take nasopharyngeal carcinoma (NPC) as an example, which is one of the most common cancers, accounting for 0.7% of all cancers. ...

AI Technical Summary

Problems solved by technology

The object of interest as the foreground only occupies a small part of the entire image, which will lead to serious foreground and background imbalance problems in the training of the learning network, which will lead to training difficulties

In addition, training samples are scarce due to the tediousness of accurately delineating tumor boundaries, which often leads to the overfitting problem of the learned network

Images