Research On X-ray Lateral Cephalometric Landmark Detection Based On Deep Learning

The positioning of X-ray lateral cephalometric landmarks plays a vital role in clinical diagnosis,treatment planning and research.However,at present,the positioning of landmarks in the medical community is manually marked by experienced doctors,which not only makes this work very cumbersome and time-consuming,but also prone to inconsistent labeling by different doctors,Therefore,there is an urgent need for an accurate automatic landmark detection algorithm.Based on deep learning,combined with the transformer model which has achieved great success in natural language processing,Thesis proposes three different algorithms,and experiments are carried out on public medical data sets.The evaluation indexes are greatly improved compared with the previous methods.The specific work and contributions of Thesis are as follows:(1)Proposed an end-to-end depth network FARNet for automatic detection of landmarks.In order to alleviate the problem of limited training data in the medical field,the depth network pre-trained in natural images is used as the backbone network,and several popular backbone network models are compared.At the same time,based on the network,a new loss function is proposed for accurate heat map regression.The loss function mainly focuses on the loss of pixels near the mark point and suppresses the loss of pixels far away from the mark point.The network model has achieved the highest detection accuracy on the three public anatomical landmark detection data sets.Among them,the 2mm successful detection rate of the cephalometric data set is 1.36 % higher than the previous best method.(2)To solve the problem of long semantics,proposed a model TSLDNet based on transformer.The algorithm divides the landmarks detection problem into two stages.In the first stage,a region of interest detection network is used to roughly estimate the center point position of several regions of interest,and then the image block is extracted from the high-resolution image according to the center point position as the input of the second stage.In the second stage,a landmark detection network based on transformer is used to accurately locate the landmarks.Because the first stage is only a rough estimation,the input image resolution is very low,and the input in the second stage is an image block rather than the whole image,which makes the amount of calculation of the whole algorithm very small.The algorithm has been tested on the head shadow data set and achieved high accuracy.The 2mm successful detection rate is only 0.37 % lower than the previous best method,but the amount of parameters and calculation is only 15.8 % and 23.3 %.(3)In order to take advantage of CNN and transformer at the same time,proposed a model Swin-CE based on swin transformer,which combines swin transformer encoder and CNN encoder.CNN encoder can effectively extract local features,while swing transformer encoder is good at capturing global features and learning remote semantic information.At the same time,the model also uses jump connection to integrate the features generated by the two encoders,and finally uses the decoder for the final heat map regression.The algorithm has also been tested on the head shadow data set,and has reached the level close to the highest accuracy.In the 2mm successful detection rate,it is 0.64 %higher than the previous best method.