Gobert N. Lee

Segmentation of liver region with tumorous tissues

Segmentation of an abnormal liver region based on CT or MR images is a crucial step in surgical planning. However, precisely carrying out this step remains a challenge due to either connectivities of the liver to other organs or the shape, internal texture, and homogeneity of liver that maybe extensively affected in case of liver diseases. Here, we propose a non-density based method for extracting the liver region containing tumor tissues by edge detection processing. False extracted regions are eliminated by a shape analysis method and thresholding processing. If the multi-phased images are available then the overall outcome of segmentation can be improved by subtracting two phase images, and the connectivities can be further eliminated by referring to the intensity on another phase image. Within an edge liver map, tumor candidates are identified by their different gray values relative to the liver. After elimination of the small and nonspherical over-extracted regions, the final liver region integrates the tumor region with the liver tissue. In our experiment, 40 cases of MDCT images were used and the result showed that our fully automatic method for the segmentation of liver region is effective and robust despite the presence of hepatic tumors within the liver.

K-means Clustering for Classifying Unlabelled MRI Data

Texture analysis of the liver for the diagnosis of cirrhosis is usually region-of-interest (ROI) based. Integrity of the label of ROI data may be a problem due to sampling. This paper investigates the use of K- means clustering, an unsupervised classifier which does not depend on the label of the data, for classification. Moreover, a procedure for generating a ROC curve for k-means clustering is also described in this paper. Using a MRI database of 44 patients with 16 cirrhotic and 28 non-cirrhotic liver cases, k-means clustering achieves an area under the ROC curve (AUC) index of 0.704. This is comparable to the performance of a linear discriminant analysis (LDA) and an artificial neural network (ANN) with the former attains a resubstitution and an average leave-one- case-out AUC of 0.781 and 0.779, respectively, and the latter attains a testing AUC of 0.801.

Automated analysis of breast parenchymal patterns in whole breast ultrasound images: preliminary experience

PurposeA computerized classification scheme to recognize breast parenchymal patterns in whole breast ultrasound (US) images was developed. A preliminary evaluation of the system performance was performed.MethodsBreast parenchymal patterns were classified into three categories: mottled pattern (MP), intermediate pattern (IP), and atrophic pattern (AP). Each classification was defined as proposed by an experienced physician. A total of 281 image features were extracted from a volume of interest which was automatically segmented. Canonical discriminant analysis with stepwise feature selection was employed for the classification of the parenchymal patterns.ResultsThe classification scheme accuracy was computed to be 83.3% (10/12 cases) in MP cases, 91.7% (22/24 cases) in IP cases, 92.9% (13/14 cases) in AP cases, and 90.0% (45/50 cases) in all the cases.ConclusionsThe feasibility of an automated ultrasonography classifier for parenchymal patterns was demonstrated with promising results in whole breast US images.

Multi-organ segmentation of CT images using statistical region merging

Segmentation is one of the key steps in the process of developing anatomical models for calculation of safe medical dose of radiation for children. This study explores the potential of the Statistical Region Merging segmentation technique for tissue segmentation in CT images. An analytical criterion allowing for an automatic tuning of the method is developed. The experiments are performed using a data set of 54 images from one patient, demonstrating the validity of the proposed criterion. The results are evaluated using the Jaccard index and a measure of border error with tolerance which addresses, application-dependant, acceptable error. The outcome shows that the technique has a great potential to become a method of choice for segmentation of CT images with an overall average boundary precison, for six representative tissues, equal to 0.937.

Automated segmentation of mammary gland regions in non-contrast X-ray CT images

The identification of mammary gland regions is a necessary processing step during the anatomical structure recognition of human body and can be expected to provide useful information for breast tumor diagnosis. This paper proposes a fully automated scheme for segmenting the mammary gland regions in non-contrast torso CT images. This scheme calculates the probability of each voxel belonging to the mammary gland or chest muscle in CT images as the reference of the segmentation, and decides the mammary gland regions based on CT number automatically. The probability is estimated from the location of the mammary glands and chest muscles in CT images. The location is investigated from a knowledge base that stores pre-recognized anatomical structures using a number of different CT scans. We applied this scheme to 66 patient cases (female, age: 20-80) and evaluated the accuracy by using the Jaccard similarity coefficient (JSC) between the segmented results and two gold standards that were generated manually by 2 medical experts independently for each CT case. The result showed that the mean value of the JSC score was 0.83 with the standard deviation of 0.09 for 66 CT cases. The proposed scheme was applied to investigate the breast density distributions in normal mammary gland regions so as to demonstrate the effect and usefulness of the proposed scheme.

Classification of Benign and Malignant Masses in Ultrasound Breast Image Based on Geometric and Echo Features

The aim of this paper is to study the use of geometric and echo features in classifying masses in ultrasound images as benign or malignant. While mammography is very effective in detecting masses and other lesions, breast ultrasound is a valuable adjunct in distinguishing solid and fluid-filled masses where the former is mostly malignant and the latter benign. Six features including two geometric features and four echo features derived from the segmented mass and its neighboring regions are employed in this study. They are the compactness and orientation of the mass, two intensity ratios of the mass and its neighboring regions, homogeneity, and depth-to-width ratio of the mass. Linear discriminant analysis and receiver operating characteristic (ROC) analysis are employed for classification and performance evaluation. The area under the ROC curve (AUC) has a value of 0.940 using all breast masses for training and testing and 0.923 using the leave-one-mass-out cross-validation method. Clinically significance of the results will be evaluated using a larger dataset.

State-of-the-Art of computer-aided detection/diagnosis (CAD)

This paper summarizes the presentations given in the special ICMB2010 session on state-of-the-art of computer-aided detection/diagnosis (CAD). The topics are concerned with the latest development of technologies and applications in CAD, which include brain MR images, fundus photographs, dynamic chest radiography, chest CT images, whole breast ultrasonography, CT colonography and torso FDG-PET scans.

Classifying masses as benign or malignant based on co-occurrence matrix textures: a comparison study of different gray level quantizations

In this paper, co-occurrence matrix based texture features are used to classify masses as benign or malignant. As (digitized) mammograms have high depth resolution (4096 gray levels in this study) and the size of a co-occurrence matrix depends on Q, the number of gray levels used for image intensity (depth) quantization, computation using co-occurrence matrices derived from mammograms can be expensive. Re-quantization using a lower value of Q is routinely performed but the effect of such procedure has not been sufficiently investigated. This paper investigates the effect of re-quantization using different Q. Four feature pools are formed with features measured on co-occurrence matrices with Q ∈{400}, Q ∈{100}, Q ∈{50} and Q ∈{400, 100, 50}. Classification results are obtained from each pool separately with the use of a genetic algorithm and the Fishers linear discriminant classifier. For Q ∈{400, 100, 50}, the best feature subsets selected by the genetic algorithm and of size k=6,7,8 have a leave-one-out area under the receiver operating characteristic (ROC) curve of 0.92, 0.93 and 0.94, respectively. Pairwise comparisons of the area index show that the differences in classification results for Q ∈{400, 100, 50} and Q ∈{50} are significant (p<0.06) for all k while that for Q ∈{400, 100, 50} and Q ∈{400} or Q ∈{100} are not significant.

Spatial Shape Constrained Fuzzy C-Means (FCM) Clustering for Nucleus Segmentation in Pap Smear Images

Precise segmentation of Pap smear cell nucleus is crucial for early diagnosis of cervical cancer. This task is particularly challenging because of cell overlapping, inconsistent staining, poor contrast and other imaging artifacts. In this study, a novel method is proposed to segment cell nucleus from overlapping Pap smear cell images. The proposed technique introduces a circular shape function (CSF) to increase the robustness of Pap cell nucleus segmentation using fuzzy c-means clustering. CSF imposes a shape constrain over the formed clusters, while improves the boundary of the nucleus. The shape function helps to differentiate the pixels having similar intensity value but located in different spatial regions. The method is evaluated using Overlapping Cervical Cytology Image Segmentation Challenge - ISBI 2014 dataset and compared with the traditional FCM clustering and recently published state-of-the-art methods. Both qualitative and quantitative measures indicate that the new technique performs favorably with others.

Full-body CT segmentation using 3D extension of two graph-based methods: a feasibility study

The paper studies the feasibility of using 3D extensions of two state-of-the-art segmentation techniques, the Statistical Region Merging (SRM) method and the Efficient Graph-based Segmentation (EGS) technique, for automatic anatomy segmentation on clinical 3D CT images. The proposed methods are tested on a dataset of 55 images. The test is for segmentation of eight representative tissues (lungs, stomach, liver, heart, kidneys, spleen, bones and the spinal cord) which are vital for accurate calculation of radiation doses. The results are evaluated using the Dice index, the Hausdorff distance and the Ht index, a measure of border error with tolerance t pixels addressing the uncertainty in the ground truth. The outcome shows that the 3D-SRM method outperforms 3D-EGS and has a great potential to become the method of choice for segmentation of full-body CT images. Using 3D-SRM, the average Dice index, the Hausdorff distance across the 8 tissues, and the H2 were0.89, 12.5 mm and0.93, respectively.