Yulei Jiang

A study on several Machine-learning methods for classification of Malignant and benign clustered microcalcifications

In this paper, we investigate several state-of-the-art machine-learning methods for automated classification of clustered microcalcifications (MCs). The classifier is part of a computer-aided diagnosis (CADx) scheme that is aimed to assisting radiologists in making more accurate diagnoses of breast cancer on mammograms. The methods we considered were: support vector machine (SVM), kernel Fisher discriminant (KFD), relevance vector machine (RVM), and committee machines (ensemble averaging and AdaBoost), of which most have been developed recently in statistical learning theory. We formulated differentiation of malignant from benign MCs as a supervised learning problem, and applied these learning methods to develop the classification algorithm. As input, these methods used image features automatically extracted from clustered MCs. We tested these methods using a database of 697 clinical mammograms from 386 cases, which included a wide spectrum of difficult-to-classify cases. We analyzed the distribution of the cases in this database using the multidimensional scaling technique, which reveals that in the feature space the malignant cases are not trivially separable from the benign ones. We used receiver operating characteristic (ROC) analysis to evaluate and to compare classification performance by the different methods. In addition, we also investigated how to combine information from multiple-view mammograms of the same case so that the best decision can be made by a classifier. In our experiments, the kernel-based methods (i.e., SVM, KFD, and RVM) yielded the best performance (A/sub z/=0.85, SVM), significantly outperforming a well-established, clinically-proven CADx approach that is based on neural network (A/sub z/=0.80).

Prostate Cancer: Differentiation of Central Gland Cancer from Benign Prostatic Hyperplasia by Using Diffusion-weighted and Dynamic Contrast-enhanced MR Imaging

PURPOSE To analyze the diffusion and perfusion parameters of central gland (CG) prostate cancer, stromal hyperplasia (SH), and glandular hyperplasia (GH) and to determine the role of these parameters in the differentiation of CG cancer from benign CG hyperplasia. MATERIALS AND METHODS In this institutional review board-approved (with waiver of informed consent), HIPAA-compliant study, 38 foci of carcinoma, 38 SH nodules, and 38 GH nodules in the CG were analyzed in 49 patients (26 with CG carcinoma) who underwent preoperative endorectal magnetic resonance (MR) imaging and radical prostatectomy. All carcinomas and hyperplastic foci on MR images were localized on the basis of histopathologic correlation. The apparent diffusion coefficient (ADC), the contrast agent transfer rate between blood and tissue (K(trans)), and extravascular extracellular fractional volume values for all carcinoma, SH, and GH foci were calculated. The mean, standard deviation, 95% confidence interval (CI), and range of each parameter were calculated. Receiver operating characteristic (ROC) and multivariate logistic regression analyses were performed for differentiation of CG cancer from SH and GH foci. RESULTS The average ADCs (× 10(-3) mm(2)/sec) were 1.05 (95% CI: 0.97, 1.11), 1.27 (95% CI: 1.20, 1.33), and 1.73 (95% CI: 1.64, 1.83), respectively, in CG carcinoma, SH foci, and GH foci and differed significantly, yielding areas under the ROC curve (AUCs) of 0.99 and 0.78, respectively, for differentiation of carcinoma from GH and SH. Perfusion parameters were similar in CG carcinomas and SH foci, with K(trans) yielding the greatest AUCs (0.75 and 0.58, respectively). Adding K(trans) to ADC in ROC analysis to differentiate CG carcinoma from SH increased sensitivity from 38% to 57% at 90% specificity without noticeably increasing the AUC (0.79). CONCLUSION ADCs differ significantly between CG carcinoma, SH, and GH, and the use of them can improve the differentiation of CG cancer from SH and GH. Combining K(trans) with ADC can potentially improve the detection of CG cancer. SUPPLEMENTAL MATERIAL http://radiology.rsna.org/lookup/suppl/doi:10.1148/radiol.10100021/-/DC1.

Quantitative Analysis of Multiparametric Prostate MR Images: Differentiation between Prostate Cancer and Normal Tissue and Correlation with Gleason Score—A Computer-aided Diagnosis Development Study

PURPOSE To evaluate the potential utility of a number of parameters obtained at T2-weighted, diffusion-weighted, and dynamic contrast material-enhanced multiparametric magnetic resonance (MR) imaging for computer-aided diagnosis (CAD) of prostate cancer and assessment of cancer aggressiveness. MATERIALS AND METHODS In this institutional review board-approved HIPAA-compliant study, multiparametric MR images were acquired with an endorectal coil in 48 patients with prostate cancer (median age, 62.5 years; age range, 44-73 years) who subsequently underwent prostatectomy. A radiologist and a pathologist identified 104 regions of interest (ROIs) (61 cancer ROIs, 43 normal ROIs) based on correlation of histologic and MR findings. The 10th percentile and average apparent diffusion coefficient (ADC) values, T2-weighted signal intensity histogram skewness, and Tofts K(trans) were analyzed, both individually and combined, via linear discriminant analysis, with receiver operating characteristic curve analysis with area under the curve (AUC) as figure of merit, to distinguish cancer foci from normal foci. Spearman rank-order correlation (ρ) was calculated between cancer foci Gleason score (GS) and image features. RESULTS AUC (maximum likelihood estimate ± standard error) values in the differentiation of prostate cancer from normal foci of 10th percentile ADC, average ADC, T2-weighted skewness, and K(trans) were 0.92 ± 0.03, 0.89 ± 0.03, 0.86 ± 0.04, and 0.69 ± 0.04, respectively. The combination of 10th percentile ADC, average ADC, and T2-weighted skewness yielded an AUC value for the same task of 0.95 ± 0.02. GS correlated moderately with 10th percentile ADC (ρ = -0.34, P = .008), average ADC (ρ = -0.30, P = .02), and K(trans) (ρ = 0.38, P = .004). CONCLUSION The combination of 10th percentile ADC, average ADC, and T2-weighted skewness with CAD is promising in the differentiation of prostate cancer from normal tissue. ADC image features and K(trans) moderately correlate with GS.

Cell Distance Mapping Identifies Functional T Follicular Helper Cells in Inflamed Human Renal Tissue

Visualizing and quantifying the spatial relationships between T and B cells identifies adaptive immune cell networks in human inflammation. Putting Human Inflammation on the Map B cells cannot fight infection by antigen stimulation alone—they need help from T cells. In mice, two-photon electron microscopy has demonstrated that T follicular helper (TFH) cells are critical for providing B cell help in germinal centers. However, it has remained unclear whether—and if so, how—TFH cells provide B cell help in humans. Now, Liarski et al. report that cell distance mapping (CDM) can be used to demonstrate cognate TFH-mediated B cell help in the context of human inflammation. CDM is a computational tool that quantifies spatial relationships between different cell types in tissue. The authors used CDM to measure the internuclear distances between TFH and B cells in inflamed human tissues. They were able to discriminate between noncognate and cognate interactions, which are required for providing help. They also characterized cognate-competent TFH cells and found that they expressed Bcl-6 and IL-21. This technique should be generalizable to diverse antigen presentation and immune cell interactions and, if so, should enhance our knowledge of the immune system in situ. T follicular helper (TFH) cells are critical for B cell activation in germinal centers and are often observed in human inflamed tissue. However, it is difficult to know if they contribute in situ to inflammation. Expressed markers define TFH subsets associated with distinct functions in vitro. However, such markers may not reflect in situ function. The delivery of T cell help to B cells requires direct cognate recognition. We hypothesized that by visualizing and quantifying such interactions, we could directly assess TFH cell competency in situ. Therefore, we developed computational tools to quantify spatial relationships between different cell subtypes in tissue [cell distance mapping (CDM)]. Analysis of inflamed human tissues indicated that measurement of internuclear distances between TFH and B cells could be used to discriminate between apparent cognate and noncognate interactions. Furthermore, only cognate-competent TFH cell populations expressed high levels of Bcl-6 and interleukin-21. These data suggest that CDM can be used to identify adaptive immune cell networks driving in situ inflammation. Such knowledge should help identify diseases, and disease subsets, that may benefit from therapeutic targeting of specific T cell–antigen-presenting cell interactions.

Seminal Vesicle Invasion in Prostate Cancer: Evaluation by Using Multiparametric Endorectal MR Imaging

PURPOSE To retrospectively evaluate the diagnostic performance of multiparametric endorectal magnetic resonance (MR) imaging, including T2-weighted, diffusion-weighted (DW), and dynamic contrast material-enhanced (DCE) MR techniques, for the diagnosis of seminal vesicle invasion (SVI) and to determine the incremental value of DW MR and DCE MR images. MATERIALS AND METHODS This retrospective HIPAA-compliant study was approved by the institutional review board, with a waiver of informed consent. The study included 131 patients (mean age, 68 years; range, 43-75 years) who underwent endorectal MR imaging before radical prostatectomy between January 2007 and April 2010. Two radiologists (A: experienced, B: less experienced) estimated the likelihood of SVI by using a five-point ordinal scale in three image-viewing settings: T2-weighted images alone; T2-weighted and DW MR images; and T2-weighted, DW MR, and DCE MR images. Sensitivity, specificity, positive predictive value, negative predictive value, and area under the receiver operating characteristic curve (AUC) were calculated. Confidence intervals estimated with bootstrapping and the McNemar test or Fisher exact test were used to compare sensitivity, specificity, positive predictive value, and negative predictive value. RESULTS Of the 131 patients, 23 (17.6%) had SVI identified after surgery. Review of T2-weighted MR images alone resulted in high specificity (93.1% and 93.6%, for radiologists A and B, respectively) and high negative predictive value (94.8% and 94.0%) but moderate sensitivity (59% and 52%) and positive predictive value (52% and 50%). Review of T2-weighted and DW MR images significantly improved specificity (96.6% [P = .02] and 98.3% [P = .003]) and positive predictive value (70% [P < .05] and 79% [P < .05]) without significantly improving AUC. Additional review of DCE MR images did not yield further incremental improvement. CONCLUSION Additional review of DW MR images improves specificity and positive predictive value in SVI detection compared with reviewing T2-weighted images alone. Addition of DCE MR images to this combination, however, does not provide incremental value for diagnosis of SVI.

Validation of Quantitative Analysis of Multiparametric Prostate MR Images for Prostate Cancer Detection and Aggressiveness Assessment: A Cross-Imager Study

PURPOSE To validate three previously identified quantitative image features across multiparametric magnetic resonance (MR) images acquired with imagers made by two different manufacturers to differentiate prostate cancer (PC) from normal prostatic tissue and to assess cancer aggressiveness. MATERIALS AND METHODS This study was HIPAA-compliant and approved by the institutional review board. Preoperative 1.5-T multiparametric endorectal MR images of 119 PC patients (dataset A, 71 patients; dataset B, 48 patients) were analyzed, and 265 PC and normal peripheral zone regions of interests (ROIs) were identified through histologic and MR consensus review. The 10th percentile average apparent diffusion coefficient (ADC) value, average ADC value, and skewness of T2-weighted signal-intensity histogram were evaluated with area under the receiver operating characteristic curve (AUC). The image features were combined with a linear discriminant analysis classifier and evaluated both on the image dataset of each type of imager alone (leave-one-patient-out evaluation) and across the datasets (training on one dataset, testing on the other). Spearman correlation coefficient was calculated between the image features and ROI-specific Gleason scores. RESULTS AUC values of the image features combined were 0.95 ± 0.02 (standard error) and 0.88 ± 0.03 on dataset B and dataset A alone, respectively, and 0.96 ± 0.02 and 0.89 ± 0.03 when training on dataset A and testing on dataset B and vice versa, respectively. Spearman correlation coefficients between Gleason scores and the ADC features were between -0.27 and -0.34. CONCLUSION Consistently across images from datasets A and B, the 10th percentile ADC value, average ADC value, and T2-weighted skewness can distinguish PC from normal-tissue ROIs, and ADC features correlate moderately with ROI-specific Gleason scores.

Components-of-Variance Models for Random-Effects ROC Analysis: The Case of Unequal Variance Structures Across Modalities

RATIONALE AND OBJECTIVES Several of the authors have previously published an analysis of multiple sources of uncertainty in the receiver operating characteristic (ROC) assessment and comparison of diagnostic modalities. The analysis assumed that the components of variance were the same for the modalities under comparison. The purpose of the present work is to obtain a generalization that does not require that assumption. MATERIALS AND METHODS The generalization is achieved by splitting three of the six components of variance in the previous model into modality-dependent contributions. Two distinct formulations of this approach can be obtained from alternative choices of the three components to be split; however, a one-to-one relationship exists between the magnitudes of the components estimated from these two formulations. RESULTS The method is applied to a study of multiple readers, with and without the aid of a computer-assist modality. performing the task of discriminating between benign and malignant clusters of microcalcifications. Analysis according to the first method of splitting shows large decreases in the reader and reader-by-case components of variance when the computer assist is used by the readers. Analysis in terms of the alternative splitting shows large decreases in the corresponding modality-interaction components. CONCLUSION A solution to the problem of multivariate ROC analysis without the assumption of equal variance structure across modalities has been provided. Alternative formulations lead to consistent results related by a one-to-one mapping. A surprising result is that estimates of confidence intervals and numbers of cases and readers required for a specified confidence interval remain the same in the more general model as in the restricted model.

Image feature analysis and computer-aided diagnosis in mammography: reduction of false-positive clustered microcalcifications using local edge-gradient analysis.

To improve the performance of a computerized scheme for detection of clustered microcalcifications in digitized mammograms, causes of detected false-positive microcalcification signals were analyzed. The false positives were grouped into four categories, namely, microcalcification like noise patterns, artifacts, linear patterns, and others. In an edge-gradient analysis, local edge-gradient values at signal-perimeter pixels of detected microcalcification signals were determined to eliminate false positives that look like subtle microcalcifications or are due to artifacts. In a linear-pattern analysis, the degree of linearity for linear patterns was determined from local gradient values from a set of linear templates oriented in 16 different directions. Threshold values for the edge-gradient analysis and the linear-pattern analysis were determined using a training database of 39 mammograms. It was possible to eliminate 59% and 25%, respectively, of 91 detected false-positive clusters with loss of only 3% of true-positive clusters. The combination of the two methods further improved the scheme in eliminating a total of 73% of the false-positive clusters with loss of 3% of true-positive clusters. Using these thresholds, the two methods were evaluated on another database of 50 mammograms. 62%, 31%, and 80% of the false-positive clusters were eliminated with loss of 3% of true-positive clusters or less, in the edge-gradient analysis, the linear-pattern analysis, and the combination of the two methods, respectively. The edge-gradient analysis and the linear-pattern analysis can reduce the false-positive detection rate, while maintaining a high level of the sensitivity.

Anniversary Paper: Evaluation of medical imaging systems

Medical imaging used to be primarily within the domain of radiology, but with the advent of virtual pathology slides and telemedicine, imaging technology is expanding in the healthcare enterprise. As new imaging technologies are developed, they must be evaluated to assess the impact and benefit on patient care. The authors review the hierarchical model of the efficacy of diagnostic imaging systems by Fryback and Thornbury [Med. Decis. Making 11, 88-94 (1991)] as a guiding principle for system evaluation. Evaluation of medical imaging systems encompasses everything from the hardware and software used to acquire, store, and transmit images to the presentation of images to the interpreting clinician. Evaluation of medical imaging systems can take many forms, from the purely technical (e.g., patient dose measurement) to the increasingly complex (e.g., determining whether a new imaging method saves lives and benefits society). Evaluation methodologies cover a broad range, from receiver operating characteristic (ROC) techniques that measure diagnostic accuracy to timing studies that measure image-interpretation workflow efficiency. The authors review briefly the history of the development of evaluation methodologies and review ROC methodology as well as other types of evaluation methods. They discuss unique challenges in system evaluation that face the imaging community today and opportunities for future advances.

Dynamic Contrast-enhanced MR Imaging Curve-type Analysis: Is It Helpful in the Differentiation of Prostate Cancer from Healthy Peripheral Zone?

PURPOSE To evaluate the performance and interobserver agreement of qualitative dynamic contrast material enhanced magnetic resonance (MR) imaging curve analysis as described in the Prostate Imaging Reporting and Data System (PI-RADS) for the differentiation of prostate cancer (PCa) from healthy prostatic tissue in the peripheral zone (PZ). MATERIALS AND METHODS This Health Insurance Portability and Accountability Act-compliant institutional review board-approved retrospective analysis included 120 consecutive pretreatment dynamic contrast-enhanced (DCE) MR imaging PCa examinations. Regions of interest (ROIs) were placed in 251 spots, including 95 (37.8%) in healthy PZ tissue and 156 (62.2%) in PCa, by using detailed histologic-multiparametric MR correlation review. Three radiologists reviewed the DCE time curves and assessed qualitative curve types as described in PI-RADS: type 1 (progressive), type 2 (plateau), or type 3 (washout). Receiver operating characteristic curve analysis was used to assess accuracy in differentiating PCa from healthy tissue on the basis of curve type, and κ was calculated to assess interobserver agreement. RESULTS Receiver operating characteristic curves were similar for all observers, but mean areas under the receiver operating characteristic curve were poor (0.58 ± 0.04 [standard deviation] to 0.63 ± 0.04). No differences in accuracy were seen for varying DCE time resolution and imaging length. Observer agreement in assessment of type 3 versus types 1 or 2 curves was substantial (0.66 < κ < 0.79), better for PCa ROIs than for healthy-tissue ROIs. The agreement between type 1 and type 2 curves was moderate to substantial (0.49 < κ < 0.78). CONCLUSION Qualitative DCE MR imaging time-curve-type analysis performs poorly for differentiation of PCa from healthy prostatic tissue. Interobserver agreement is excellent in assessment of type 3 curves but only moderate for type 1 and 2 curves.