System For Medical Image Analysis Using Deep Learning And Its Practice In Segmentation Of Gastroscope Video

With the development of deep learning based medical imaging analysis methodologies,deep learning models have exhibited better performance than medical specialists in the assessment of certain clinical examinations,which shows a promising way of intelligent diagnosis for medical images.Meanwhile,three elements which are data,algorithm and calculation respectively play a major role for the further development of this method.In the field of medical image analysis,the emphasis focuses in two aspects that one is to establish a labeled medical image data set,and the other is to research on clinical problem oriented analysis algorithms.Generally,two preparations should be done before medical image analysis.The first one is the acquisition and annotation of large-scale raw medical image data and the second one is the establishment of analytically environment for deep learning algorithm.In view of the above issues,this paper carries out the study of design and implement for medical image analysis system,which includes construction module of medical image data set,imaging annotation set,analysis environment of algorithm.Furthermore,we apply the system to the key anatomical location segmentation in gastroscopic video,The main contents are as follows:1)construction of medical image data set.With analysis of the image format of the gastroscopy and CCTA(Coronary Computed Tomography Angiography),and Combined with the requirements of labeling forms and quality control in medical images,we propose a methodology for construction of Multi-modal medical image dataset.2)implementation of medical image analysis system.Based on the dataset construction method,we design and implement the annotation module for gastroscopic and CCTA image.Meanwhile,on the basis of the annotation module,we develop three medical datasets,which include the gastroscopic image disease annotation set,the gastroscopic image anatomical location set and the CCTA image set.Besides,we accomplish the online deep learning algorithm analysis module in terms of the annotation module and the constructed medical image data set.Accordingly,the medical image analysis system is constituted by the image annotation module,image datasets and the deep learning analysis environment.3)applicating the constructed system to key anatomical location segmentation in gastroscopic video.The proposed medical image analysis system is applied to generate the training dataset and define the deep learning model to study the segmentation algorithm for the key physiological anatomical location in gastroscopic video.The SegNet(semantic segmentation network)is used to train and evaluate the multi-classification task for the lower esophagus,the dentate line,the stomach horn,the pylorus,the descending part of the duodenum and the background.The results show that the GA(global accuracy)of the model is 88.5%while the CAA(classification average accuracy)is 71.2%and the mloU(mean intersection-over-union)is 63.5%.Simultaneously,the testing on the gastroscopy lesion segmentation model also indicates the model can accurately locate the physiological anatomy of the lesion.