Computerized characterization of breast masses on three-dimensional ultrasound volumes

We are developing computer vision techniques for the characterization of breast masses as malignant or benign on radiologic examinations. In this study, we investigated the computerized characterization of breast masses on three-dimensional (3-D) ultrasound (US) volumetric images. We developed 2-D and 3-D active contour models for automated segmentation of the mass volumes. The effect of the initialization method of the active contour on the robustness of the iterative segmentation method was studied by varying the contour used for its initialization. For a given segmentation, texture and morphological features were automatically extracted from the segmented masses and their margins. Stepwise discriminant analysis with the leave-one-out method was used to select effective features for the classification task and to combine these features into a malignancy score. The classification accuracy was evaluated using the area Az under the receiver operating characteristic (ROC) curve, as well as the partial area index Az(0.9), defined as the relative area under the ROC curve above a sensitivity threshold of 0.9. For the purpose of comparison with the computer classifier, four experienced breast radiologists provided malignancy ratings for the 3-D US masses. Our dataset consisted of 3-D US volumes of 102 biopsied masses (46 benign, 56 malignant). The classifiers based on 2-D and 3-D segmentation methods achieved test Az values of 0.87+/-0.03 and 0.92+/-0.03, respectively. The difference in the Az values of the two computer classifiers did not achieve statistical significance. The Az values of the four radiologists ranged between 0.84 and 0.92. The difference between the computer's Az value and that of any of the four radiologists did not achieve statistical significance either. However, the computer's Az(0.9) value was significantly higher than that of three of the four radiologists. Our results indicate that an automated and effective computer classifier can be designed for differentiating malignant and benign breast masses on 3-D US volumes. The accuracy of the classifier designed in this study was similar to that of experienced breast radiologists.     

Characterization of mammographic masses based on level set segmentation with new image features and patient information

Computer-aided diagnosis (CAD) for characterization of mammographic masses as malignant or benign has the potential to assist radiologists in reducing the biopsy rate without increasing false negatives. The purpose of this study was to develop an automated method for mammographic mass segmentation and explore new image based features in combination with patient information in order to improve the performance of mass characterization. The authors' previous CAD system, which used the active contour segmentation, and morphological, textural, and spiculation features, has achieved promising results in mass characterization. The new CAD system is based on the level set method and includes two new types of image features related to the presence of microcalcifications with the mass and abruptness of the mass margin, and patient age. A linear discriminant analysis (LDA) classifier with stepwise feature selection was used to merge the extracted features into a classification score. The classification accuracy was evaluated using the area under the receiver operating characteristic curve. The authors' primary data set consisted of 427 biopsy-proven masses (200 malignant and 227 benign) in 909 regions of interest (ROIs) (451 malignant and 458 benign) from multiple mammographic views. Leave-one-case-out resampling was used for training and testing. The new CAD system based on the level set segmentation and the new mammographic feature space achieved a view-based Az value of 0.83 +/- 0.01. The improvement compared to the previous CAD system was statistically significant (p = 0.02). When patient age was included in the new CAD system, view-based and case-based Az values were 0.85 +/- 0.01 and 0.87 +/- 0.02, respectively. The study also demonstrated the consistency of the newly developed CAD system by evaluating the statistics of the weights of the LDA classifiers in leave-one-case-out classification. Finally, an independent test on the publicly available digital database for screening mammography with 132 benign and 197 malignant ROIs containing masses achieved a view-based Az value of 0.84 +/- 0.02.     

Automated computerized scheme for distinction between benign and malignant solitary pulmonary nodules on chest images

A novel automated computerized scheme has been developed to assist radiologists for their distinction between benign and malignant solitary pulmonary nodules on chest images. Our database consisted of 55 chest radiographs (33 primary lung cancers and 22 benign nodules). In this method, the location of a nodule was indicated first by a radiologist. The difference image with a nodule was produced by use of filters and then represented in a polar coordinate system. The nodule was segmented automatically by analysis of contour lines of the gray-level distribution based on the polar-coordinate representation. Two clinical parameters (age and sex) and 75 image features were determined from the outline, the image, and histogram analysis for inside and outside regions of the segmented nodule. Linear discriminant analysis (LDA) and knowledge about benign and malignant nodules were used to select initial feature combinations. Many combinations for subgroups of 77 features were evaluated as input to artificial neural networks (ANNs). The performance of ANNs with the selected 7 features by use of the round-robin test showed Az = 0.872, which was greater than Az = 0.854 obtained previously with the manual method (P= 0.53). The performance of LDA (Az = 0.886) was slightly improved compared to that of ANNs (P = 0.59) and was greater than that of the manual method (Az = 0.854) reported previously (P = 0.40). The high level of its performance indicates the potential usefulness of this automated computerized scheme in assisting radiologists as a second opinion for distinction between benign and malignant solitary pulmonary nodules on chest images.     

Computerized scheme for determination of the likelihood measure of malignancy for pulmonary nodules on low-dose CT images

An automated computerized scheme has been developed for determination of the likelihood measure of malignancy of pulmonary nodules on low-dose helical CT (LDCT) images. Our database consisted of 76 primary lung cancers (147 slices) and 413 benign nodules (576 slices). With this automated computerized scheme, the location of a nodule was first indicated by a radiologist. The outline of the nodule was segmented automatically by use of a dynamic programming technique. Various objective features on the nodules were determined by use of outline analysis and image analysis, and the likelihood measure of malignancy was determined by use of linear discriminant analysis (LDA). The effect of many different combinations of features and the performance of LDA in distinguishing benign nodules from malignant ones were evaluated by means of receiver operating characteristic (ROC) analysis. The Az value (area under the ROC curve) obtained by the computerized scheme in distinguishing benign nodules from malignant ones was 0.828 when a single slice was employed for each of the nodules. However, the Az value was improved to 0.846 when multiple slices were used for determination of the likelihood measure of malignancy. The Az values obtained by the computerized scheme on LDCT images were significantly greater than the Az value of 0.70, which was obtained from our previous observer studies by radiologists in distinguishing benign nodules from malignant ones on LDCT images. The automated computerized scheme for determination of the likelihood measure of malignancy would be useful in assisting radiologists to distinguish between benign and malignant pulmonary nodules on LDCT images.     

Computerized lesion detection on breast ultrasound

We investigated the use of a radial gradient index (RGI) filtering technique to automatically detect lesions on breast ultrasound. After initial RGI filtering, a sensitivity of 87% at 0.76 false-positive detections per image was obtained on a database of 400 patients (757 images). Next, lesion candidates were segmented from the background by maximizing an average radial gradient (ARD) index for regions grown from the detected points. At an overlap of 0.4 with a radiologist lesion outline, 75% of the lesions were correctly detected. Subsequently, round robin analysis was used to assess the quality of the classification of lesion candidates into actual lesions and false-positives by a Bayesian neural network. The round robin analysis yielded an Az value of 0.84, and an overall performance by case of 94% sensitivity at 0.48 false-positives per image. Use of computerized analysis of breast sonograms may ultimately facilitate the use of sonography in breast cancer screening programs.     

A computerized method for automated identification of erect posteroanterior and supine anteroposterior chest radiographs

A computerized scheme was developed for automated identification of erect posteroanterior (PA) and supine anteroposterior (AP) chest radiographs. The method was based on three features, the tilt angle of the scapula superior border, the tilt angle of the clavicle and the extent of radiolucence in lung fields, to identify the view of a chest radiograph. The three indices A(scapula), A(clavicle) and C(lung) were determined from a chest image for the three features. Linear discriminant analysis was used to classify PA and AP chest images based on the three indices. The performance of the method was evaluated by receiver operating characteristic analysis. The proposed method was evaluated using a database of 600 PA and 600 AP chest radiographs. The discriminant performances Az of A(scapula), A(clavicle) and C(lung) were 0.878 ± 0.010, 0.683 ± 0.015 and 0.962 ± 0.006, respectively. The combination of the three indices obtained an Az value of 0.979 ± 0.004. The results indicate that the combination of the three indices could yield high discriminant performance. The proposed method could provide radiologists with information about the view of chest radiographs for interpretation or could be used as a preprocessing step for analyzing chest images.     

根据磁共振T2加权影像特征预测乳腺癌组织学分级

组织学分级是乳腺癌的重要预后指标。探讨磁共振(MRI)T2加权影像特征与乳腺癌组织学分级的关联性, 可为术前预测乳腺癌组织学分级提供有意义的参考作用。回顾性分析167例术前行MRI检查并经病理诊断为浸润性乳腺癌患者, 其中组织学分级Ⅱ级和Ⅲ级的分别为72例和95例。利用计算机半自动方法, 分割患者乳腺磁共振T2加权影像的病灶区域, 并对其提取包括纹理特征和形态特征的40维影像特征。采用留一法交叉验证方法(LOOCV), 通过统计学方法分析比较影像特征在组织分级Ⅱ级和Ⅲ级组间的差异, 并设计多变量分类预测模型。绘制受试者工作特征曲线(ROC), 并计算对应曲线下的面积(AUC);计算敏感性、特异性、F1-Measure等指标, 对预测模型进行综合评估。对每一维特征进行单变量逻辑回归分析, 在低分级和高分级组间进行统计检验分析(t检验)。形态特征中最优单特征为病灶半径, AUC值为0.742(P<0.05);纹理特征中最优特征为大面积高灰度级增强, AUC值为0.742(P<0.05)。设计多元逻辑回归(MLR)、支持向量机(SVM)、多任务学习(MTL)等3种分类器, 预测乳腺癌组织学分级, 其AUC值分别为0.767±0.036、0.772±0.036和0.771±0.037, 对应特异性分别为0.667、0.653、0.708, 灵敏度分别为0.747、0.737、0.684。研究表明, 乳腺癌的T2加权影像特征在一定程度上反映其组织学分级, 对乳腺癌的预后判断具有潜在价值。     

Computer-assisted diagnosis of breast cancer using a data-driven Bayesian belief network

This study investigates a simple Bayesian belief network for the diagnosis of breast cancer, and specifically addresses the question of whether integrating image and non-image based features into a single network can yield better performance than hybrid combinations of independent networks. From a dataset of 419 cases, including 92 malignancies, 13 features relating to mammographic findings, physical examinations and patients' clinical histories, were extracted to build three Bayesian belief networks. The scenarios tested included a network incorporating all features and two hybrids which combined the outputs of sub-networks corresponding to the image or non-image features. Average areas (Az) under the corresponding ROC curves were used as measures of performance. The network incorporating only image based features performed better (Az =0.81) than that using nonimage features (Az = 0.71). Both hybrid classifiers yielded better performance (Az =0.85 for averaging and Az = 0.87 for logistic regression), but neither hybrid was as accurate as the network incorporating all features (Az = 0.89). This preliminary study suggests that, like human observers who concurrently consider different types of information, a single classifier that simultaneously evaluates both image and non-image information can achieve better diagnostic performance than the hybrid combinations considered here.     

Improvement of mammographic mass characterization using spiculation meausures and morphological features

We are developing new computer vision techniques for characterization of breast masses on mammograms. We had previously developed a characterization method based on texture features. The goal of the present work was to improve our characterization method by making use of morphological features. Toward this goal, we have developed a fully automated, three-stage segmentation method that includes clustering, active contour, and spiculation detection stages. After segmentation, morphological features describing the shape of the mass were extracted. Texture features were also extracted from a band of pixels surrounding the mass. Stepwise feature selection and linear discriminant analysis were employed in the morphological, texture, and combined feature spaces for classifier design. The classification accuracy was evaluated using the area Az under the receiver operating characteristic curve. A data set containing 249 films from 102 patients was used. When the leave-one-case-out method was applied to partition the data set into trainers and testers, the average test Az for the task of classifying the mass on a single mammographic view was 0.83 +/- 0.02, 0.84 +/- 0.02, and 0.87 +/- 0.02 in the morphological, texture, and combined feature spaces, respectively. The improvement obtained by supplementing texture features with morphological features in classification was statistically significant (p = 0.04). For classifying a mass as malignant or benign, we combined the leave-one-case-out discriminant scores from different views of a mass to obtain a summary score. In this task, the test Az value using the combined feature space was 0.91 +/- 0.02. Our results indicate that combining texture features with morphological features extracted from automatically segmented mass boundaries will be an effective approach for computer-aided characterization of mammographic masses.     

Computerized diagnosis of breast lesions on ultrasound

We present a computer-aided diagnosis (CAD) method for breast lesions on ultrasound that is based on the automatic segmentation of lesions and the automatic extraction of four features related to the lesion shape, margin, texture, and posterior acoustic behavior. Using a database of 400 cases (94 malignant lesions, 124 complex cysts, and 182 benign solid lesions), we investigate the marginal benefit of each feature in our CAD method and the performance of our CAD method in distinguishing malignant lesions from various classes of benign lesions. Finally, independent validation is performed on our CAD method. Eleven independent trials yielded an average Az value of 0.87 in the task of distinguishing malignant from benign lesions.     

Classification of breast masses in ultrasonic B scans using Nakagami and K distributions

Classification of breast masses in greyscale ultrasound images is undertaken using a multiparameter approach. Five parameters reflecting the non-Rayleigh nature of the backscattered echo were used. These parameters, based mostly on the Nakagami and K distributions, were extracted from the envelope of the echoes at the site, boundary, spiculated region and shadow of the mass. They were combined to create a linear discriminant. The performance of this discriminant for the classification of breast masses was studied using a data set consisting of 70 benign and 29 malignant cases. The Az value for the discriminant was 0.96 +/- 0.02, showing great promise in the classification of masses into benign and malignant ones. The discriminant was combined with the level of suspicion values of the radiologist leading to an Az value of 0.97 +/- 0.014. The parameters used here can be calculated with minimal clinical intervention, so the method proposed here may therefore be easily implemented in an automated fashion. These results also support the recent reports suggesting that ultrasound may help as an adjunct to mammography in breast cancer diagnostics to enhance the classification of breast masses.     

Automatic identification and classification of characteristic kinetic curves of breast lesions on DCE-MRI

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) of the breast is being used increasingly in the detection and diagnosis of breast cancer as a complementary modality to mammography and sonography. Although the potential diagnostic value of kinetic curves in DCE-MRI is established, the method for generating kinetic curves is not standardized. The inherent reason that curve identification is needed is that the uptake of contrast agent in a breast lesion is often heterogeneous, especially in malignant lesions. It is accepted that manual region of interest selection in 4D breast magnetic resonance (MR) images to generate the kinetic curve is a time-consuming process and suffers from significant inter- and intraobserver variability. We investigated and developed a fuzzy c-means (FCM) clustering-based technique for automatically identifying characteristic kinetic curves from breast lesions in DCE-MRI of the breast. Dynamic contrast-enhanced MR images were obtained using a T1-weighted 3D spoiled gradient echo sequence with Gd-DTPA dose of 0.2 mmol/kg and temporal resolution of 69 s. FCM clustering was applied to automatically partition the signal-time curves in a segmented 3D breast lesion into a number of classes (i.e., prototypic curves). The prototypic curve with the highest initial enhancement was selected as the representative characteristic kinetic curve (CKC) of the lesion. Four features were then extracted from each characteristic kinetic curve to depict the maximum contrast enhancement, time to peak, uptake rate, and washout rate of the lesion kinetics. The performance of the kinetic features in the task of distinguishing between benign and malignant lesions was assessed by receiver operating characteristic analysis. With a database of 121 breast lesions (77 malignant and 44 benign cases), the classification performance of the FCM-identified CKCs was found to be better than that from the curves obtained by averaging over the entire lesion and similar to kinetic curves generated from regions drawn within the lesion by a radiologist experienced in breast MRI.     

DCE-MRI及DWI影像特征对乳腺癌病理组织学分级及Ki-67表达的预测研究

联合动态增强磁共振成像(DCE-MRI)以及弥散加权成像(DWI)的影像特征,通过建立模型,分别对乳腺癌的组织学分级以及Ki-67的表达进行预测。对144例未经过任何手术或化疗的乳腺浸润性导管癌患者的数据进行回顾性分析,这些患者均采用3T 扫描仪进行术前乳腺 MRI 检查,从中获取DCE-MRI以及DWI影像,并从 DWI 中计算得到表观扩散系数 (ADC)。对不同参数磁共振影像进行肿瘤分割,并分别从整个肿瘤区域中提取纹理特征、统计特征、形态特征等。采用无监督判别特征选择方法(UDFS)和Fisher Score算法进行特征选择,将分类模型分别应用于DCE-MRI及DWI图像数据,将得到的不同分类器进行多分类器模型融合,最终得到多参数图像的联合预测结果。为了评估所建立模型的分类性能, 通过留一法交叉验证 (LOOCV) 的方法计算ROC曲线下的面积(AUC)。对于分级任务,DCE-MRI的第二增强序列达到0.780的最优AUC,(特异度为0.647,灵敏度为0.934);对于Ki-67预测任务,DWI序列达到0.756的最优AUC(特异度为0.806,灵敏度为0.695)。经过融合,分级的预测结果提高到AUC为0.808(特异度为0.706,灵敏度为0.895),Ki-67的预测结果提高到AUC为0.783(特异度为0.778,灵敏度为0.722)。结果表明,相比采用单一参数的磁共振图像数据,DCE-MRI和DWI的影像特征联合可以提高分类器的性能。     

Classification of breast masses via nonlinear transformation of features based on a kernel matrix

We propose methods to perform a certain nonlinear transformation of features based on a kernel matrix, before the classification step, aiming to improve the discriminating power of the comparatively weak edge-sharpness and texture features of breast masses in mammograms, and seek better incorporation of features representing different radiological characteristics than shape features only. Kernel principal component analysis (KPCA) is applied to improve the discriminating power of each single feature in an expanded feature space and the discriminating capability of different feature combinations in other transformed, more informative, lower-dimensional feature spaces. A kernel partial least squares (KPLS) method is developed to derive score vectors for a shape feature set, and an edge-sharpness and texture feature set, respectively, with minimal covariance between each other, to help in achieving improved diagnosis using multiple radiological characteristics of breast masses. Fisher's linear discriminant analysis (FLDA) is employed to evaluate the classification capability of the transformed features. The methods were tested with a set of 57 regions in mammograms, of which 20 are related to malignant tumors and 37 to benign masses, represented using five shape features, three edge-sharpness features, and 14 texture features. The classification performance of the edge-sharpness and texture features, via KPCA transformation, was significantly improved from 0.75 to 0.85 in terms of the area under the receiver operating characteristics curve (Az). The classification performance of all of the shape, edge-sharpness, and texture features, via KPLS transformation, was improved from 0.95 to 1.0 in Az value.     

Optimized approach to decision fusion of heterogeneous data for breast cancer diagnosis

As more diagnostic testing options become available to physicians, it becomes more difficult to combine various types of medical information together in order to optimize the overall diagnosis. To improve diagnostic performance, here we introduce an approach to optimize a decision-fusion technique to combine heterogeneous information, such as from different modalities, feature categories, or institutions. For classifier comparison we used two performance metrics: The receiving operator characteristic (ROC) area under the curve [area under the ROC curve (AUC)] and the normalized partial area under the curve (pAUC). This study used four classifiers: Linear discriminant analysis (LDA), artificial neural network (ANN), and two variants of our decision-fusion technique, AUC-optimized (DF-A) and pAUC-optimized (DF-P) decision fusion. We applied each of these classifiers with 100-fold cross-validation to two heterogeneous breast cancer data sets: One of mass lesion features and a much more challenging one of microcalcification lesion features. For the calcification data set, DF-A outperformed the other classifiers in terms of AUC (p < 0.02) and achieved AUC=0.85 +/- 0.01. The DF-P surpassed the other classifiers in terms of pAUC (p < 0.01) and reached pAUC=0.38 +/- 0.02. For the mass data set, DF-A outperformed both the ANN and the LDA (p < 0.04) and achieved AUC=0.94 +/- 0.01. Although for this data set there were no statistically significant differences among the classifiers' pAUC values (pAUC=0.57 +/- 0.07 to 0.67 +/- 0.05, p > 0.10), the DF-P did significantly improve specificity versus the LDA at both 98% and 100% sensitivity (p < 0.04). In conclusion, decision fusion directly optimized clinically significant performance measures, such as AUC and pAUC, and sometimes outperformed two well-known machine-learning techniques when applied to two different breast cancer data sets.     

Analysis of temporal changes of mammographic features: computer-aided classification of malignant and benign breast masses

A new classification scheme was developed to classify mammographic masses as malignant and benign by using interval change information. The masses on both the current and the prior mammograms were automatically segmented using an active contour method. From each mass, 20 run length statistics (RLS) texture features, 3 speculation features, and 12 morphological features were extracted. Additionally, 20 difference RLS features were obtained by subtracting the prior RLS features from the corresponding current RLS features. The feature space consisted of the current RLS features, the difference RLS features, the current and prior speculation features, and the current and prior mass sizes. Stepwise feature selection and linear discriminant analysis classification were used to select and merge the most useful features. A leave-one-case-out resampling scheme was used to train and test the classifier using 140 temporal image pairs (85 malignant, 55 benign) obtained from 57 biopsy-proven masses (33 malignant, 24 benign) in 56 patients. An average of 10 features were selected from the 56 training subsets: 4 difference RLS features, 4 RLS features, and 1 speculation feature from the current image, and 1 speculation feature from the prior, were most often chosen. The classifier achieved an average training Az of 0.92 and a test Az of 0.88. For comparison, a classifier was trained and tested using features extracted from the 120 current single images. This classifier achieved an average training Az of 0.90 and a test Az of 0.82. The information on the prior image significantly (p = 0.015) improved the accuracy for classification of the masses.     

Application of the iris filter for automatic detection of pulmonary nodules on computed tomography images

We have developed a computer-aided diagnosis (CAD) system to detect pulmonary nodules on thin-slice helical computed tomography (CT) images. We have also investigated the capability of an iris filter to discriminate between nodules and false-positive findings. Suspicious regions were characterized with features based on the iris filter output, gray level and morphological features, extracted from the CT images. Functions calculated by linear discriminant analysis (LDA) were used to reduce the number of false-positives.The system was evaluated on CT scans containing 77 pulmonary nodules. The system was trained and evaluated using two completely independent data sets. Results for a test set, evaluated with free-response receiver operating characteristic (FROC) analysis, yielded a sensitivity of 80% at 7.7 false-positives per scan.