Lesion Detection And Segmentation On Medical Images Based On Partial Differential Equations

With the development of imaging technology,information technology,and mathematics,medical imaging has achieved an unprecedented development,which greatly enhances the healthcare and living quality of the human.However,due to the increasing number of medical images,to obtain accurate and fast image analysis is becoming challenging.Compared with the conventionally manual segmentation for targeted anatomical structures or lesion tissues,the segmentation algorithms based on artificial intelligence and mathematical mechanism can provide more stable,more accurate and more repeatable segmentation results.The research focusing on automatic segmentation,has been an attractive topic,and it is also an important direction for developing precision medicine.In clinical applications,precise segmentation of lesions in medical images from multiple modalities can provide reliable information for further precision diagnosis,treatment,surgical planning,and navigation.However,the accurate and fast segmentation of lesions in medical images is still challenge,which is caused by several aspects including image quality and the diversity in image types.Firstly,Biomedical images always be degraded by low contrast,heavy noise,uneven gray distribution,and other factors,which causes the low qualities of medical image.Moreover,due to the characteristics of the imaging theories,the same organ may exhibit different intensities from people to people.In addition,the existence of artifacts may cause blurred edges or even partially missing boundaries.Furthermore,due to individual difference among the human,the images for the same lesion area under the same imaging conditions may exhibit different characteristics such as gray scale,texture,and morphology.In order to solve these problems,in this paper,several strategies are proposed to extract critical information from medical images and eliminate the possible interferences.Therefore accurate segmentation can be implemented in medical images and deal with the situations of low contrast,heavy noises and partial missing boundaries.The main contributions in this dissertation can be summarized as follows:1.Due to the presence of noises,low contrast,complex textures,and inhomogeneities in medical images,robust segmentation for regions of interest is challenging.A modified variational model is proposed to improve segmentation based on intensity and texture features.In this model,information from these two sources are fused by the Dempster-Shafer evidence theory,which thus can provide a compelling evidence for object segmentation.The proposed method can capture the differences of textures and intensities between the object and background,which facilitates the segmentation of lesion areas that have similar intensities to their surroundings.To prove the efficiency of the proposed method,52 renal ultrasound images with low contrast and inhomogeneous lesions are utilized for quantitative evaluation.Compared with the classical CV model,the method using a new distribution metric based on prediction theory,the method using Bayesian theorem to fuse information,the MICO method,and RSF method,the proposed method exhibits the highest mean precision(94.7%)and mean Dice coefficient(92.8%).Moreover,the mean absolute surface distance(MAD)and the symmetric mean absolute surface distance(SMAD)of the proposed method are 0.11 and 0.13,respectively,which are lower than those of other methods.Comparison results demonstrate that the segmentation results of the proposed method are closer to the ground truths in comparison with other methods.Therefore,the proposed method are promising strategy for imaging segmentation and shows the potential in clinical practice.2.Inspired by the isoperimetric inequality in differential geometry,a novel segmentation method in which the isoperimetric constrain is integrated into a level set framework to penalize the ratio of squared perimeter to the enclosed area of an active contour is proposed.The novel model can ensure the compactness of segmenting objects and complete missing and/or blurred parts of their boundarnes simultaneously.As the isoperimetric shape constraint is independent of the explicit expression of shapes and scale-invariant,the proposed method can handle various objects with different scales.In addition,the estimation of shape-related parameters is not necessary.The results showed that the proposed method could segment lesions with blurred and/or partially missing boundaries in ultrasound,Computed Tomography(CT)and Magnetic Resonance(MR)images efficiently.Quantitative evaluation also confirm that the proposed method can provide more accurate segmentation than two well-known level set methods.Therefore,the proposed method shows great potential for the accurate segmentation of lesions and can be applied in diagnoses and surgical planning.3.A variational level set approach to segment lesions with compact shapes in medical images is also proposed.In this study,segmentation of hepatocellular carcinoma which are usually of various shapes,variable intensities,and weak boundaries is investigated.An efficient constraint which is called the isoperimetric constraint to describe the compactness of shapes is applied in this method.In order to ensure the precise segmentation and stable movement of the level set,a distance regularization is also implemented in the proposed variational framework.The proposed method showed positive results in segmenting various hepatocellular carcinoma(HCC)regions in Computed Tomography(CT)images.Comparison results also proved that the proposed method is more accurate than two classical approaches,i.e.the CV model and the BCS model.This proposed method shows their great potential in segmenting the HCC regions in CT images.In experiments,a class of HCC regions,which have complex texture,high level homogeneities,and low contrast with their backgrounds,are founded.Such regions are sensitive to the position of initialization,in segmentation process,the active curve is misled by the high gradient caused by the textures.For this kind regions,a novel quantity to measure the complexity of a region with inhomogeneous intensity is proposed.In order to describe real boundaries of objects,we further design an edge detector which is based on the similarity between object regions and those around them.Imbedding these two measurements of inhomogeneous region into a level set framework,the proposed model is applied to segment HCC regions in CT images with promising results.Additionally,benefitting from the two measurements,segmentation is robust with respect to the initialization.Comparison results also prove that the proposed method is more accurate than two well-known method,the CV model and the BCS model,on segmenting objects with inhomogeneities.4.An automated algorithm of segmentation for multiple types of lesions in ultrasound images is de'veloped.This method can detect and segment lesions automatically,and generate accurate segmentation results for lesion regions.In the detection step,two saliency detection frameworks using global information are adopted to capture the differences between normal and abnormal organs as well as these between lesions and the surrounding normal tissues.In the segmentation step,three types of local information,i.e.image intensity,improved local binary patterns(LBP)features,and an edge indicator,are embedded in a modified level set framework to carry on the segmentation procedure based on the detection results.In this step,the lesion regions can be segmented accurately while the false alarmed regions can be removed.Experiments of the proposed method are implemented in ultrasound images of human kidneys with cyst and carcinoma regions.In these experiments,the lesions can be automatically detected and segmented.The efficiency and accuracy of the proposed segmentation method are validated by quantitative measurements and comparisons with other three segmentation methods.The average precision and dice coefficient of segmenting renal cysts of the proposed method are 95.33%and 90.16%,respectively,and those of renal carcinomas are 94.22%and 91.13%,respectively,which are significantly higher than the results of the other segmentation methods.Therefore,the proposed segmentation method shows the promise in the segmentation of renal lesion in ultrasound images.In addition,since the automatic method utilizes general information of healthy organs/tissues instead of specific features of lesions,this method can be extended to deal with different types of lesions in other organs or to be used in medical images obtained by other modalities.