Research On Automatic Identification And Annotation Of Digestive Lesions Based On Endoscopic Images

Endoscopic diagnose of digestive diseases is a challenging task in clinic.Artificial diagnose is susceptible to many of negative factors,such as clinicians’ fatigue and insufficient experience,time limitations,huge numbers of endoscopic images,diversity of the appearance of lesions,etc.Therefore,utilizing computer science and technology to develop computer-aided diagnosis(CAD)methods based on endoscopic images is with great significance,because it provides supplements to the clinical diagnose of digestive diseases.It also contributes a lot to the reduction of misdiagnose rate and to the enhancement of diagnostic accuracy and efficiency.This dissertation carried out a deep research on the development of two kinds of CAD methods: automatic identification of lesion images and automatic annotation of lesion regions,the main contents include:(1)Research on the identification of lesion images from gastrointestinal endoscope based on the combination of machine learning and traditional features.In this part,we put forward a novel feature extraction strategy which combines the machine learning and traditional features.For the extraction of machine learning features,a new joint diagonalization algorithm was designed.Unlike the traditional joint diagonalization algorithms,the proposed algorithm does not need iteration,inversion or approximation procedure,therefore it is with higher calculating speed and accuracy.Based on this joint diagonalization algorithm,we modified the traditional machine learning algorithm: Asymmetric Principal Component Analysis(APCA).the modified algorithm preserved the feature extraction performance of APCA but significantly reduced its computing load.For the extraction of traditional features,the color coherence vector algorithm was utilized.Based on the combination of the two kinds of features,a new computer-aided method was proposed to identify the gastrointestinal endoscopic images containing lesions.The clinical data,which consist of 1330 images totally,were used to validate the proposed method.The experimental results showed that,for the identification of early esophageal cancer images,early gastric cancer images and small intestinal bleeding images,the areas under the receiver operator characteristic curves of the proposed method were 0.9471,0.9532 and 0.9776,respectively,higher than those of the compared methods.Therefore,The proposedmethod has a satisfactory performance of identifying lesion images.(2)Research on the annotation of early upper digestive cancers based on two visual saliency levels of endoscopic images.In this part,we modified the traditional annotation frame work used by existing annotation methods,and proposed the two visual saliency levels of endoscopic images to perform a “two-step” annotation.In addition,the traditional segmentation method: SLIC(Simple Linear Iterative Clustering)was improved to a localized version with faster calculating speed.Finally,a comprehensive annotation method was developed,which can automatically annotate early upper digestive cancer lesion regions in gastroscopic images.A total of 871 gastroscopic images were used to validate the proposed method.The experimental results showed that the lesion detection rate and mean Dice similarity coefficients of the proposed method were 97.24 and 75.15%,respectively,higher than the comparison methods.Moreover,the method has a faster running speed,and fewer false-positive outputs.In addition,it improved the low detection rate and bad annotation performances of traditional methods when dealing with small lesions.(3)Research on the annotation of early upper digestive cancers based on depth information and deep learning techniques.In this part,we put forward a novel idea,that is “combining the depth and RGB information of original images to aid the semantic segmentation procedures”.Moreover,the state-of-the-art semantic segmentation network: Deeplabv3+ was modified into a 4-channel form,which further enhanced the annotation performance.In addition,a post-processing step that referenced the clinical requirements was designed and applied,to enhance the visual effects of annotation results.Based on the above improvements,a new annotation methods was developed.Totally 4231 gastroscopic images were used to validate the proposed method.The experiment results showed.that the lesion detection rate and mean Dice similarity coefficient of the proposed method were 97.54% and 74.43%,respectively.Compared with other deep learning-based annotation methods,the proposed method has better annotation performances and fewer false positive outputs.The CAD methods developed in this work showed remarkable performances,and are helpful to the reduction of misdiagnose rate of upper digestive caners and small intestine bleeding.They also offer good potentials to enhance the accuracy and efficiency of clinical endoscopic diagnoses.What is more,the proposed annotation method can effectively aid the endoscopic mucosal resection(ESD)and endoscopicsubmucosal dissection(EMR)in clinic.Therefore,the methods developed in this work are with great clinical application prospects.