Research On Graph-Based Medical Image Segmentation Methods

The interactive graph-based methods have been gaining popularity in the field of image segmentation in recent years.The interactive graph-based methods have shown several merits in clinical practice when compared with the automatic approaches:1)an intuitive and convenient interface is provided for users and a satisfactory result can be obtained within sufficient interaction;2)neither training phase nor prior informa-tion is necessary,thus they are robust to various targets on various images;3)interac-tive approaches can be extended to automatic methods when auto-seeding strategy is involved.Based on former research,we have made our own innovation and improvement on graph-based image segmentation methods.The main work includes two parts:Firstly,a novel Random Walker-based image segmentation method was proposed and applied to multimodal brain tumor MRI(magnetic resonance imaging)images.Se-condly,an iterative semi-supervised spectral clustering-based image segmentation method was proposed and applied to DBT(digital breast tomosynthesis)images.Brain tumors are among the most common and fatal cancers.The MRI is one of the effective imaging modalities for brain tumor diagnosis.Multimodal brain tumor MRI images segmentation is a challenging task and has been gaining popularity in the field of image segmentation.In this text,for multimodal brain tumor MRI images segmentation,we proposed a Random Walkers based image segmentation method,which extended the Random Walkers method to feature space with the intent to cal-culate the probabilities that points in feature space were assigned to the foreground or background according to the labeled points.To make the method more tolerant to er-roneous seeding,we also reformulated the energy function in the unconstraint form.To evaluate the proposed method,a set of multimodal brain tumor MR images is uti-lized.The experimental results demonstrated that proposed method could yield more robust and accurate result than the Graph Cuts and the Random Walkers algorithms for multimodal brain tumor MR image segmentation when seed quantity decreased,seed position deviated,and erroneous seeding were involved.We developed an asymmetry based detecting method to locate the lesions and tumors coarsely to auto-mate the proposed segmentation method.The automated method is tested on all 30 cases for both lesion and tumor segmentation.The DC(Dice’s Coefficient)overlap of lesion segmentation is 0.852 ± 0.107 for all cases,and the DC overlap of tumor seg-mentation is 0.729 ± 0.252 for all cases.A GPU linear algebra toolkit called CULA was utilized in order to solve the large sparse linear equation groups related to the proposed method.For a typical case with resolution of 256 x 256 x 181,it took roughly 5 minutes to complete on our workstation.Computer-aided diagnosis(CAD)systems can use computer technologies to detect abnormalities mammograms.Image segmentation technique plays a key role in computer-aided diagnosis of breast cancer.Aimed at lesion segmentation for breast cancer CAD,a semi-automated semi-supervised spectral clustering based iterative image segmentation method was proposed in this paper.The user interface of Grab-Cut method,which allows incomplete labeling,was introduced into the proposed method,and it could extract the contour by a given bounding box around the object.We constructed a novel type of similarity graphs,which integrated the semi-supervised spectral clustering in feature space and Random Walkers segmenta-tion in image space into a unity framework.The segmentation results could be achieved by iteratively optimizing the energy function.To test the proposed method,60 sets of DBT images from 30 patients with breast lesions(including 15 benign and 15 malignant cases)were utilized,and the average DC overlap is 0.798 ± 0.090.Fur-thermore,the effect of the initial bounding box and the key parameters on results was discussed and analyzed in our research.