Research on robust medical image segmentation method based on time adaptive neural network

Abstract

In medical image segmentation, the significant performance of a convolutional neural network (CNN) decreases when a training image and a test image do not match in terms of acquiring details. In order to solve the problem, a segmentated CNN is designed into connection of two sub-networks: a relatively shallow image normalization CNN, and a deep CNN for segmenting a normalization image. The two sub-networks are trained using the training data set consisting of annotated images from specific scanners and protocol settings. During testing, an image normalization sub-network is adopted for each test image, and implicit priori on a predicted segmentation label is used as guidance. An independently trained de-noising autoencoder (DAE) is employed in order to model such implicit priori on reasonable anatomical segmentation tags. The proposed idea is verified on a multi-center magnetic resonance imaging data set of three anatomical organs of the brain, the heart and the prostate. The proposed test time adaptation always provides performance improvements, demonstrating the prospect and versatility of the method. This design can be used in any split network to increase robustness to imaging scanners and protocol variations.

Description

technical field [0001] The invention belongs to the technical field of medical image processing, and in particular relates to a method for robust medical image segmentation based on a time-adaptive neural network. Background technique [0002] Medical image segmentation is an important prerequisite for a variety of clinical analyses. Techniques based on convolutional neural networks (CNNs) have taken the lead in recent years. These methods achieve excellent performance in several challenges. In fact, for some anatomies and imaging modalities, the performance of these methods has been comparable to inter-expert variance. However, one of the key issues hindering the large-scale adoption of these methods in practical applications is the lack of robustness to changes in imaging protocols and scanners between training images and test images. In this invention, the success of segmental CNNs is built upon by increasing their robustness to input variations. [0003] Although CNN...

AI Technical Summary

Problems solved by technology

The solution of training a separate CNN for each new scanner and protocol is impractical due to the difficulty of repeatedly assembling large training datasets

Images