Spine CT image recognition method based on multi-dimensional residual network

Abstract

The invention discloses a spine CT image recognition method based on a multi-dimensional residual network. The method comprises the following steps: establishing a pathological spine recognition model based on a multi-dimensional residual neural network; setting labels for the training samples, carrying out dimension raising so that the dimensions of the training samples match the multi-dimensional residual neural network; training a sick spine recognition model by using the processed training sample; then raising the dimension of a spine CT image to be identified, so that the dimension of the spine CT image matches the multi-dimensional residual neural network; and outputting identification features by the sick spine identification model. Spatial information of the CT medical image can be well reserved, and the image features can be more comprehensively and accurately recognized.

Description

technical field [0001] The invention relates to a spine CT image recognition method, in particular to a spine CT image recognition method based on a multidimensional residual network. Background technique [0002] At present, in the medical field, deep learning is mainly used in three aspects: medical image recognition, medical image segmentation and computer-aided disease diagnosis, reducing the repetitive and heavy workload of medical staff and providing convenience for smart medical care. In 2016, YunliangCai et al. proposed a multimodal vertebral recognition framework using Transformational Deep Convolutional Network (TDCN). TDCN automatically extracts modalities and performs adaptive, high-discrimination, constituting invariant features for recognition. Using a TDCN-based recognition system can simultaneously identify the position, label the pose of the vertebral structures in MR and CT. The system has been successfully tested on multimodal datasets for lumbar and who...

AI Technical Summary

Problems solved by technology

But in fact, medical images, especially CT images, are generally three-dimensional images stored in DICOM, Analyze, and NIfTI formats. In this case, using 2DCNN is undoubtedly to reduce the dimensionality of medical images, ignoring the space to a certain extent. information, there is a certain bias in identifying pathological spines

On the other hand, as the number of training layers increases, deep learning will generally become more and more difficult. Some networks may also degenerate when they start to converge, causing the accuracy to quickly reach saturation. higher rate phenomenon

[0004] The existing technology basically uses two-dimensional convolutional neural network for the pathological recognition of spine CT, which will lose spatial information to a large extent and reduce the accuracy rate.

Moreover, the existing data sets related to the spine are too small, and there is no way to train in large batches, which leads to a lower error tolerance rate.

Images