Ankur M. Doshi

Big Data and the Future of Radiology Informatics

Rapid growth in the amount of data that is electronically recorded as part of routine clinical operations has generated great interest in the use of Big Data methodologies to address clinical and research questions. These methods can efficiently analyze and deliver insights from high-volume, high-variety, and high-growth rate datasets generated across the continuum of care, thereby forgoing the time, cost, and effort of more focused and controlled hypothesis-driven research. By virtue of an existing robust information technology infrastructure and years of archived digital data, radiology departments are particularly well positioned to take advantage of emerging Big Data techniques. In this review, we describe four areas in which Big Data is poised to have an immediate impact on radiology practice, research, and operations. In addition, we provide an overview of the Big Data adoption cycle and describe how academic radiology departments can promote Big Data development.

Three-dimensional MR Cholangiopancreatography in a Breath Hold with Sparsity-based Reconstruction of Highly Undersampled Data

Purpose To develop a three-dimensional breath-hold (BH) magnetic resonance (MR) cholangiopancreatographic protocol with sampling perfection with application-optimized contrast using different flip-angle evolutions (SPACE) acquisition and sparsity-based iterative reconstruction (SPARSE) of prospectively sampled 5% k-space data and to compare the results with conventional respiratory-triggered (RT) acquisition. Materials and Methods This HIPAA-compliant prospective study was institutional review board approved. Twenty-nine patients underwent conventional RT SPACE and BH-accelerated SPACE acquisition with 5% k-space sampling at 3 T. Spatial resolution and other parameters were matched when possible. BH SPACE images were reconstructed by enforcing joint multicoil sparsity in the wavelet domain (SPARSE-SPACE). Two board-certified radiologists independently evaluated BH SPARSE-SPACE and RT SPACE images for image quality parameters in the pancreatic duct and common bile duct by using a five-point scale. The Wilcoxon signed-rank test was used to compare BH SPARSE-SPACE and RT SPACE images. Results Acquisition time for BH SPARSE-SPACE was 20 seconds, which was significantly (P < .001) shorter than that for RT SPACE (mean ± standard deviation, 338.8 sec ± 69.1). Overall image quality scores were higher for BH SPARSE-SPACE than for RT SPACE images for both readers for the proximal, middle, and distal pancreatic duct, but the difference was not statistically significant (P > .05). For reader 1, distal common bile duct scores were significantly higher with BH SPARSE-SPACE acquisition (P = .036). More patients had acceptable or better overall image quality (scores ≥ 3) with BH SPARSE-SPACE than with RT SPACE acquisition, respectively, for the proximal (23 of 29 [79%] vs 22 of 29 [76%]), middle (22 of 29 [76%] vs 18 of 29 [62%]), and distal (20 of 29 [69%] vs 13 of 29 [45%]) pancreatic duct and the proximal (25 of 28 [89%] vs 22 of 28 [79%]) and distal (25 of 28 [89%] vs 24 of 28 [86%]) common bile duct. Conclusion BH SPARSE-SPACE showed similar or superior image quality for the pancreatic and common duct compared with that of RT SPACE despite 17-fold shorter acquisition time. (©) RSNA, 2016.

Use of MRI in Differentiation of Papillary Renal Cell Carcinoma Subtypes: Qualitative and Quantitative Analysis

OBJECTIVE The purpose of this study was to determine whether qualitative and quantitative MRI feature analysis is useful for differentiating type 1 from type 2 papillary renal cell carcinoma (PRCC). MATERIALS AND METHODS This retrospective study included 21 type 1 and 17 type 2 PRCCs evaluated with preoperative MRI. Two radiologists independently evaluated various qualitative features, including signal intensity, heterogeneity, and margin. For the quantitative analysis, a radiology fellow and a medical student independently drew 3D volumes of interest over the entire tumor on T2-weighted HASTE images, apparent diffusion coefficient parametric maps, and nephrographic phase contrast-enhanced MR images to derive first-order texture metrics. Qualitative and quantitative features were compared between the groups. RESULTS For both readers, qualitative features with greater frequency in type 2 PRCC included heterogeneous enhancement, indistinct margin, and T2 heterogeneity (all, p < 0.035). Indistinct margins and heterogeneous enhancement were independent predictors (AUC, 0.822). Quantitative analysis revealed that apparent diffusion coefficient, HASTE, and contrast-enhanced entropy were greater in type 2 PRCC (p < 0.05; AUC, 0.682-0.716). A combined quantitative and qualitative model had an AUC of 0.859. Qualitative features within the model had interreader concordance of 84-95%, and the quantitative data had intraclass coefficients of 0.873-0.961. CONCLUSION Qualitative and quantitative features can help discriminate between type 1 and type 2 PRCC. Quantitative analysis may capture useful information that complements the qualitative appearance while benefiting from high interobserver agreement.

Clinicopathologic Outcomes of Cystic Renal Cell Carcinoma

BACKGROUND The purpose of this study was to describe the clinicopathologic characteristics and oncologic outcomes of patients who underwent nephrectomy for cystic renal masses. PATIENTS AND METHODS Using an institutional review board-approved database, we retrospectively reviewed the clinical, pathologic, radiologic, and oncologic outcome data of patients who received nephrectomy for a complex cystic renal mass. RESULTS Sixty-one patients were identified who received nephrectomy for a complex cystic lesion. Average age was 64 years. Thirty-nine (64%) patients were male. At the time of resection, 1 (1.6%), 3 (4.8%), 53 (86.8%), and 4 (6.5%) had a Bosniak category II, IIF, III, and IV cystic lesion, respectively. Nineteen (31.1%) patients were initially managed expectantly but underwent surgery because of progression of complexity on follow-up. Mean pathologic tumor size was 3.3 cm (range, 0.7-12 cm). Forty-eight (78.6%) of the lesions were found to be malignant. Thirty-seven (77.1%), 5 (10.4%), 4 (8.3%), and 2 (4.1%) were stage T1a, T1b, T2a, and T3a, respectively. Clear cell was the most common histologic subtype (44%), followed by papillary (21.3%), and unclassified RCC (4.9%). With a mean and median follow-up of 48.4 and 43.0 months, respectively, no patients developed a local or metastatic recurrence. All patients were alive at last follow-up. CONCLUSION In our series with moderate follow-up, cystic RCCs do not appear to recur or progress regardless of size, histologic subtype, or grade. These findings suggest the malignant potential of cRCCs is significantly less than solid RCCs. Further investigation is required to determine if cRCCs should be classified and managed independently from solid RCCs.

MRI Features of Renal Cell Carcinoma That Predict Favorable Clinicopathologic Outcomes

OBJECTIVE The purpose of this article is to determine whether MRI features of renal cell carcinoma (RCC), such as enhancing solid component and T1 signal intensity, are associated with clinicopathologic outcomes. MATERIALS AND METHODS This retrospective study included 241 RCCs in 230 patients who underwent preoperative MRI, had pathologic analysis results available, and were monitored for at least 3 months. A radiologist assessed tumor features on MRI, including unenhanced T1 signal relative to renal cortex and the percentage of solid enhancing components. The electronic medical record or follow-up images were reviewed to assess for the development of local recurrence or metastases. Statistical analysis was performed to correlate imaging features at MRI with pathologic and clinical outcome. RESULTS The following tumor features were observed: predominantly cystic morphologic features (defined as solid component≤25%, n=33), solid component greater than 25% (n=208), T1 hypointensity (n=97), and T1 intermediate intensity or hyperintensity (n=144). Local recurrence or metastases were observed in 14 patients. Compared with T1-intermediate or -hyperintense lesions, T1-hypointense RCCs were more likely to be low stage (90.7% vs 74.3%; p=0.001) and low grade (78.9% vs 41.8%; p<0.001) and had a lower rate of recurrence or metastases (3.3% vs 8%; p=0.167). Compared with lesions with greater than 25% solid enhancement, predominantly cystic RCCs were more likely to be lower stage (93.9% vs 78.8%; p=0.053) and lower grade (94.7 vs 56.5%; p<0.001) and to have no incidence of recurrence or metastasis (0% vs 6.9%; p=0.227). RCCs that were both cystic and T1 hypointense (n=14) were lower stage (100% vs 79.6%; p=0.047) and lower grade (92.9% vs 58.1%; p=0.01) and had no recurrence or metastases on follow-up. CONCLUSION Cystic and T1-hypointense RCC show less-aggressive pathologic features and favorable clinical behavior.

Retrospective Assessment of Histogram-Based Diffusion Metrics for Differentiating Benign and Malignant Endometrial Lesions.

Objective Our study aimed to retrospectively evaluate the utility of volumetric histogram-based diffusion metrics in differentiating benign from malignant endometrial abnormalities. Methods A total of 54 patients underwent pelvic magnetic resonance imaging with diffusion-weighted imaging before endometrial tissue diagnosis. Two radiologists placed volumes of interest on the apparent diffusion coefficient (ADC) map encompassing the entire endometrium and focal endometrial lesions. The mean ADC, percentile ADC values, kurtosis, skewness, and entropy of ADC were compared between benign and malignant abnormalities. Results In premenopausal patients, significant independent predictors of malignancy were whole-endometrium analysis for R1, 10th to 25th ADC percentile (P = 0.012); whole-endometrium analysis for R2, mean ADC (P = 0.001) and skewness (P = 0.004); focal lesion analysis for R1, skewness (P = 0.045); focal lesion analysis for R2, 10th to 25th ADC percentile (P ⩽ 0.0001). The area under the curve for malignancy was 90.0% to 97.3% and 76.1% to 77.3% for the more and less experienced radiologists, respectively. In postmenopausal patients, the only significant difference was kurtosis using whole-endometrium analysis for R1 (P = 0.042). Conclusions Volumetric ADC histogram metrics may help radiologists assess the risk of malignancy in endometrial abnormalities on magnetic resonance imaging in premenopausal patients.

High Spatiotemporal Resolution Dynamic Contrast-Enhanced MR Enterography in Crohn Disease Terminal Ileitis Using Continuous Golden-Angle Radial Sampling, Compressed Sensing, and Parallel Imaging

OBJECTIVE The purpose of this article was to assess the feasibility of golden-angle radial acquisition with compress sensing reconstruction (Golden-angle RAdial Sparse Parallel [GRASP]) for acquiring high temporal resolution data for pharmacokinetic modeling while maintaining high image quality in patients with Crohn disease terminal ileitis. MATERIALS AND METHODS Fourteen patients with biopsy-proven Crohn terminal ileitis were scanned using both contrast-enhanced GRASP and Cartesian breath-hold (volume-interpolated breath-hold examination [VIBE]) acquisitions. GRASP data were reconstructed with 2.4-second temporal resolution and fitted to the generalized kinetic model using an individualized arterial input function to derive the volume transfer coefficient (K(trans)) and interstitial volume (v(e)). Reconstructions, including data from the entire GRASP acquisition and Cartesian VIBE acquisitions, were rated for image quality, artifact, and detection of typical Crohn ileitis features. RESULTS Inflamed loops of ileum had significantly higher K(trans) (3.36 ± 2.49 vs 0.86 ± 0.49 min(-1), p < 0.005) and v(e) (0.53 ± 0.15 vs 0.20 ± 0.11, p < 0.005) compared with normal bowel loops. There were no significant differences between GRASP and Cartesian VIBE for overall image quality (p = 0.180) or detection of Crohn ileitis features, although streak artifact was worse with the GRASP acquisition (p = 0.001). CONCLUSION High temporal resolution data for pharmacokinetic modeling and high spatial resolution data for morphologic image analysis can be achieved in the same acquisition using GRASP.

Dynamic contrast-enhanced MRI of the prostate: An intraindividual assessment of the effect of temporal resolution on qualitative detection and quantitative analysis of histopathologically proven prostate cancer.

To assess the effects of temporal resolution (RT) in dynamic contrast‐enhanced magnetic resonance imaging (DCE‐MRI) on qualitative tumor detection and quantitative pharmacokinetic parameters in prostate cancer.

Use of a Machine-learning Method for Predicting Highly Cited Articles Within General Radiology Journals.

RATIONALE AND OBJECTIVES This study aimed to assess the performance of a text classification machine-learning model in predicting highly cited articles within the recent radiological literature and to identify the models most influential article features. MATERIALS AND METHODS We downloaded from PubMed the title, abstract, and medical subject heading terms for 10,065 articles published in 25 general radiology journals in 2012 and 2013. Three machine-learning models were applied to predict the top 10% of included articles in terms of the number of citations to the article in 2014 (reflecting the 2-year time window in conventional impact factor calculations). The model having the highest area under the curve was selected to derive a list of article features (words) predicting high citation volume, which was iteratively reduced to identify the smallest possible core feature list maintaining predictive power. Overall themes were qualitatively assigned to the core features. RESULTS The regularized logistic regression (Bayesian binary regression) model had highest performance, achieving an area under the curve of 0.814 in predicting articles in the top 10% of citation volume. We reduced the initial 14,083 features to 210 features that maintain predictivity. These features corresponded with topics relating to various imaging techniques (eg, diffusion-weighted magnetic resonance imaging, hyperpolarized magnetic resonance imaging, dual-energy computed tomography, computed tomography reconstruction algorithms, tomosynthesis, elastography, and computer-aided diagnosis), particular pathologies (prostate cancer; thyroid nodules; hepatic adenoma, hepatocellular carcinoma, non-alcoholic fatty liver disease), and other topics (radiation dose, electroporation, education, general oncology, gadolinium, statistics). CONCLUSIONS Machine learning can be successfully applied to create specific feature-based models for predicting articles likely to achieve high influence within the radiological literature.

Multi-institutional analysis of CT and MRI reports evaluating indeterminate renal masses: comparison to a national survey investigating desired report elements

PurposeTo determine the need for a standardized renal mass reporting template by analyzing reports of indeterminate renal masses and comparing their contents to stated preferences of radiologists and urologists.MethodsThe host IRB waived regulatory oversight for this multi-institutional HIPAA-compliant quality improvement effort. CT and MRI reports created to characterize an indeterminate renal mass were analyzed from 6 community (median: 17 reports/site) and 6 academic (median: 23 reports/site) United States practices. Report contents were compared to a published national survey of stated preferences by academic radiologists and urologists from 9 institutions. Descriptive statistics and Chi-square tests were calculated.ResultsOf 319 reports, 85% (271; 192 CT, 79 MRI) reported a possibly malignant mass (236 solid, 35 cystic). Some essential elements were commonly described: size (99% [269/271]), mass type (solid vs. cystic; 99% [268/271]), enhancement (presence vs. absence; 92% [248/271]). Other essential elements had incomplete penetrance: the presence or absence of fat in solid masses (14% [34/236]), size comparisons when available (79% [111/140]), Bosniak classification for cystic masses (54% [19/35]). Preferred but non-essential elements generally were described in less than half of reports. Nephrometry scores usually were not included for local therapy candidates (12% [30/257]). Academic practices were significantly more likely than community practices to include mass characterization details, probability of malignancy, and staging. Community practices were significantly more likely to include management recommendations.ConclusionsRenal mass reporting elements considered essential or preferred often are omitted in radiology reports. Variation exists across radiologists and practice settings. A standardized template may mitigate these inconsistencies.