| During the researches on medical imaging,insufficient medical image data and poor data quality,which brings great difficulty to diagnosis with computer vision,is a huge obstacle.Besides,high-quality medical annotation on images consumes a lot of energy and time for professional.Therefore,this thesis intends to study the medical image augmentation method based on the ocular surface images that were annotated by the doctor,so as to generate medical image that is in line with the actual situation,and owns high resolution and includes diversity diseases through computer vision algorithms.Moreover,these data can be used to build a good deep learning model in computer vision field,as well alleviate the annotation demand for professional doctors.This thesis proposes a generation method that supplements sparse data set in the nature by using generating image component to combine.Specifically,the generative method consists of the following three processes.(1)The common ocular surface diseases were generated by the VQVAE-2 method based on deep learning.The model consists of two parts: auto encoder and the autoregressive model,which are mainly used to generate key lesions in three common ocular surface diseases(keratitis,cataracts and pterygiums)and normal eye images.(2)The generated images were fed into the semantic segmentation model(UNet)to obtain the semantic components in the image.These semantic components not only can be used as the weak annotation for training the segmentation model,but also can be used in the subsequent generation of image components.The main purpose of this step is using the segmentation model to segment the pupil and cornea in the ocular surface images.(3)Combine the segmentation components with another slit lamp ocular surface image.Shape mapping and poisson fusion were used during this step.Image divergence smoothing was performed through the fusion method to offset for the discontinuity of the combined area,finally high-quality medical images can be obtain.Since the combined region is extracted based on the semantic segmentation model,the semantic information has been existed,so this region can also be used as the mask label for the semantic segmentation model.In the course of the experiments in this thesis,it verified that the data generation method based on the coordination of auto encoder and autoregressive model,VQVAE-2,can generate high-quality image data.At the same time,the image combination method can further expand the data set and preserve the characteristics and semantics information of the data meanwhile.The cooperation of these two methods is competent as a data set augmentation strategy for solving the problem brought by insufficient data when training deep learning models.UNet with an improved pixel weighting method can achieve DSC of 0.978 and 0.946 for pupil and cornea respectively,and supplemented annotation information to the generated data.The component based image combination method(poisson fusion)can better combine the segmented images and reduce the inconsistency on the component boundary.It is verified that the classification model trained with the generated data together with real data can achieve the performance close to that trained with the real in terms of precision and recall.Moreover,they are better than the model trained with a little amount of samples.The image combination generation method proposed in this paper can complete the task of generating high-quality and clear medical images so that more medical images can be produced for research.These generated ocular surface images can be used as additional training data to improve the performance of the model.At the same time,the method used in this thesis can be used as a reference in other fields of other types of images to help solve the problem of insufficient training data. |
PDF Full Text Request