Faculty Bibliography

BACKGROUND:Our aims were to determine if features derived from texture analysis (TA) can distinguish normal, benign, and malignant tissue on automated breast ultrasound (ABUS); to evaluate whether machine learning (ML) applied to TA can categorise ABUS findings; and to compare ML to the analysis of single texture features for lesion classification. METHODS:This ethically approved retrospective pilot study included 54 women with benign (n = 38) and malignant (n = 32) solid breast lesions who underwent ABUS. After manual region of interest placement along the lesions' margin as well as the surrounding fat and glandular breast tissue, 47 texture features (TFs) were calculated for each category. Statistical analysis (ANOVA) and a support vector machine (SVM) algorithm were applied to the texture feature to evaluate the accuracy in distinguishing (i) lesions versus normal tissue and (ii) benign versus malignant lesions. RESULTS:Skewness and kurtosis were the only TF significantly different among all the four categories (p < 0.000001). In subsets (i) and (ii), a maximum area under the curve of 0.86 (95% confidence interval [CI] 0.82-0.88) for energy and 0.86 (95% CI 0.82-0.89) for entropy were obtained. Using the SVM algorithm, a maximum area under the curve of 0.98 for both subsets was obtained with a maximum accuracy of 94.4% in subset (i) and 90.7% in subset (ii). CONCLUSIONS:TA in combination with ML might represent a useful diagnostic tool in the evaluation of breast imaging findings in ABUS. Applying ML techniques to TFs might be superior compared to the analysis of single TF.

CONTEXT:Recent studies suggested that magnetic resonance imaging (MRI) followed by targeted biopsy ("MRI-stratified pathway") detects more clinically significant prostate cancers (csPCa) than the systematic transrectal ultrasound-guided prostate biopsy (TRUS-Bx) pathway, but controversy persists. Several randomized clinical trials (RCTs) were recently published, enabling generation of higher-level evidence to evaluate this hypothesis. OBJECTIVE:To perform a systematic review and meta-analysis of RCTs comparing the detection rates of csPCa in the MRI-stratified pathway and the systematic TRUS-Bx pathway in patients with a suspicion of prostate cancer (PCa). EVIDENCE ACQUISITION:PubMed, EMBASE, and Cochrane databases were searched up to March 18, 2019. RCTs reporting csPCa detection rates of both pathways in patients with a clinical suspicion of prostate cancer were included. Relative csPCa detection rates of the MRI-stratified pathway were pooled using random-effect model. Study quality was assessed using the Cochrane risk of bias tool for randomized trials. A comparison of detection rates of clinically insignificant PCa (cisPCa) and any PCa was also performed. EVIDENCE SYNTHESIS:Nine RCTs (2908 patients) were included. The MRI-stratified pathway detected more csPCa than the TRUS-Bx pathway (relative detection rate 1.45 [95% confidence interval {CI} 1.09-1.92] for all patients, and 1.42 [95% CI 1.02-1.97] and 1.60 [95% CI 1.01-2.54] for biopsy-naïve and prior negative biopsy patients, respectively). Detection rates were not significantly different between pathways for cisPCa (0.89 [95% CI 0.49-1.62]), but higher in the MRI-stratified pathway for the detection of any PCa (1.39 [95% CI 1.05-1.84]). CONCLUSIONS:The MRI-stratified pathway detected more csPCa than the systematic TRUS-guided biopsy pathway in men with a clinical suspicion of PCa, for both biopsy-naïve patients and those with prior negative biopsy. The detection rate of any PCa was higher in the MRI-stratified pathway, but not significantly different from that of cisPCa. PATIENT SUMMARY:Our meta-analysis of clinical trials shows that the magnetic resonance imaging-stratified pathway detects more clinically significant prostate cancers than the transrectal ultrasound-guided prostate biopsy pathway in men with a suspicion of prostate cancer.

Background Recent studies have reported the additive value of combined gallium 68 (⁶⁸Ga)-labeled Glu-urea-Lys (Ahx)-HBED-CC ligand targeting the prostate-specific membrane antigen (PSMA) (hereafter called ⁶⁸Ga-PSMA-11) PET/MRI for the detection and localization of primary prostate cancer compared with multiparametric MRI. Purpose To compare the diagnostic accuracy and interrater agreement of multiparametric MRI and ⁶⁸Ga-PSMA-11 PET/MRI for the detection of extracapsular extension (ECE) and seminal vesicle infiltration (SVI) in patients with prostate cancer. Materials and Methods Retrospective analysis of 40 consecutive men who underwent multiparametric MRI and ⁶⁸Ga-PSMA-11 PET/MRI within 6 months for suspected prostate cancer followed by radical prostatectomy between April 2016 and July 2018. Four readers blinded to clinical and histopathologic findings rated the probability of ECE and SVI at multiparametric MRI and PET/MRI by using a five-point Likert-type scale. The prostatectomy specimen served as the reference standard. Accuracy was assessed with a multireader multicase analysis and by calculating reader-average areas under the receiver operating characteristics curve (AUCs), sensitivity, and specificity for ordinal and dichotomized data in a region-specific and patient-specific approach. Interrater agreement was assessed with the Fleiss multirater κ. Results For multiparametric MRI versus PET/MRI in ECE detection, respectively, AUC, sensitivity, and specificity in the region-specific analysis were 0.67 and 0.75 (P = .07), 28% (21 of 76) and 47% (36 of 76) (P = .09), and 94% (529 of 564) and 90% (509 of 564) (P = .007). For the patient-specific analysis, AUC, sensitivity, and specificity were 0.66 and 0.73 (P = .19), 46% (22 of 48) and 69% (33 of 48) (P = .04), and 75% (84 of 112) and 67% (75 of 112) (P = .19), respectively. For multiparametric MRI versus PET/MRI in SVI detection, respectively, AUC, sensitivity, and specificity of the region-specific analysis were 0.66 and 0.74 (P = .21), 35% (seven of 20) and 50% (10 of 20) (P = .25), and 98% (295 of 300) and 94% (282 of 300) (P < .001). For the patient-specific analysis, AUC, sensitivity, and specificity were 0.65 and 0.79 (P = .25), 35% (seven of 20) and 55% (11 of 20) (P = .20), and 98% (137 of 140) and 94% (131 of 140) (P = .07), respectively. Interrater reliability for multiparametric MRI versus PET/MRI did not differ for ECE (κ, 0.46 vs 0.40; P = .24) and SVI (κ, 0.23 vs 0.33; P = .39). Conclusion Our results suggest that gallium 68 (⁶⁸Ga)-labeled Glu-urea-Lys (Ahx)-HBED-CC ligand targeting the prostate-specific membrane antigen (PSMA) (⁶⁸Ga-PSMA-11) PET/MRI and multiparametric MRI perform similarly for local staging of prostate cancer in patients with intermediate-to-high-risk prostate cancer. The increased sensitivity of ⁶⁸Ga-PSMA-11 PET/MRI for the detection of extracapsular disease comes at the cost of a slightly reduced specificity. © RSNA, 2019.

PURPOSE/OBJECTIVE:To train a CycleGAN on downscaled versions of mammographic data to artificially inject or remove suspicious features, and to determine whether these AI-mediated attacks can be detected by radiologists. MATERIAL AND METHODS/METHODS:From two publicly available datasets, BCDR and INbreast, we selected 680 images with and without lesions as training data. An internal dataset (n = 302 cancers, n = 590 controls) served as test data. We ran two experiments (256 × 256 px and 512 × 408 px) and applied the trained model to the test data. Three radiologists read a set of images (modified and originals) and rated the presence of suspicious lesions on a scale from 1 to 5 and the likelihood of the image being manipulated. The readout was evaluated by multiple reader multiple case receiver operating characteristics (MRMC-ROC) analysis using the area under the curve (AUC). RESULTS:At the lower resolution, the overall performance was not affected by the CycleGAN modifications (AUC 0.70 vs. 0.76, p = 0.67). However, one radiologist exhibited lower detection of cancer (0.85 vs 0.63, p = 0.06). The radiologists could not discriminate between original and modified images (0.55, p = 0.45). At the higher resolution, all radiologists showed significantly lower detection rate of cancer in the modified images (0.80 vs. 0.37, p < 0.001), however, they were able to detect modified images due to better visibility of artifacts (0.94, p < 0.0001). CONCLUSION/CONCLUSIONS:Our proof-of-concept study shows that CycleGAN can implicitly learn suspicious features and artificially inject or remove them in existing images. The applicability of the method is currently limited by the small image size and introduction of artifacts.

PURPOSE/OBJECTIVE:To compare the correlation of transition zone prostate-specific antigen density (TZPSAD) versus whole gland (WG) density (PSAD) with Gleason score. METHODS:In this single-center, retrospective cohort study, men undergoing MRI of the prostate in 2015 and 2016 who had a transperineal template saturation biopsy within 6 months of MRI were included (n = 178; median age 64 y, interquartile range [IQR] 58-68 y; PSA 6.6 ng/ml, 4.6-9.7 ng/ml). The WG and TZ were segmented voxel-wise on T2-weighted transverse planes. The volumes and corresponding PSA-densities were calculated. Correlations with the Gleason score were assessed with Spearman's rho. Optimal thresholds of the PSA densities were computed using the Youden Index of the receiver-operating-characteristics curve. A p-value of ≤ 0.05 was considered statistically significant. RESULTS:) for discrimination between Gleason 3 + 4 and 4 + 3. CONCLUSION/CONCLUSIONS:may help in risk stratification of men with suspected PCa.

PURPOSE/OBJECTIVE:To compare image quality between a 2D T1w turbo spin echo (TSE) sequence and a Compressed SENSE accelerated 3D T1w black blood TSE sequence (equipped with a black blood prepulse for blood signal suppression) in pre- and postcontrast imaging of the pituitary and to assess scan time reductions. METHODS AND MATERIALS/METHODS:For this retrospective study, 56 patients underwent pituitary MR imaging at 3T. 28 patients were scanned with the 2D- and 28 patients with the accelerated 3D sequence. Two board certified neuroradiologists independently evaluated 13 qualitative image features (12 features on postcontrast- and 1 feature on precontrast images).SNR and CNR measurements were obtained. Interreader agreement was assessed with the intraclass correlation coefficient while differences in scores were assessed with exact Wilcoxon rank sum tests. RESULTS:The interreader agreement ranged from fair (visibility of the ophthalmic nerve, ICC = 0.57) to excellent (presence and severity of pulsation artefacts, ICC = 0.97). The Compressed SENSE accelerated 3D sequence outperformed the 2D sequence in terms of "overall image quality" (median: 4 versus 3, p = 0.04) and "presence and severity of pulsation artefacts" (median: 0 versus 1, p < 0.001). There were no significant differences in any other qualitative and quantitative (SNR, CNR) image quality features. Scan time was reduced by 03:53 min (33.1%) by replacing the 2D with the 3D sequence. CONCLUSION/CONCLUSIONS:The Compressed SENSE accelerated 3D T1w black blood TSE sequence is a reliable alternative for the standard 2D sequence in pituitary imaging. The black blood prepulse may aid in suppression of pulsation artefacts.

OBJECTIVES/OBJECTIVE:To evaluate a deep convolutional neural network (dCNN) for detection, highlighting, and classification of ultrasound (US) breast lesions mimicking human decision-making according to the Breast Imaging Reporting and Data System (BI-RADS). METHODS AND MATERIALS/METHODS:One thousand nineteen breast ultrasound images from 582 patients (age 56.3 ± 11.5 years) were linked to the corresponding radiological report. Lesions were categorized into the following classes: no tissue, normal breast tissue, BI-RADS 2 (cysts, lymph nodes), BI-RADS 3 (non-cystic mass), and BI-RADS 4-5 (suspicious). To test the accuracy of the dCNN, one internal dataset (101 images) and one external test dataset (43 images) were evaluated by the dCNN and two independent readers. Radiological reports, histopathological results, and follow-up examinations served as reference. The performances of the dCNN and the humans were quantified in terms of classification accuracies and receiver operating characteristic (ROC) curves. RESULTS:In the internal test dataset, the classification accuracy of the dCNN differentiating BI-RADS 2 from BI-RADS 3-5 lesions was 87.1% (external 93.0%) compared with that of human readers with 79.2 ± 1.9% (external 95.3 ± 2.3%). For the classification of BI-RADS 2-3 versus BI-RADS 4-5, the dCNN reached a classification accuracy of 93.1% (external 95.3%), whereas the classification accuracy of humans yielded 91.6 ± 5.4% (external 94.1 ± 1.2%). The AUC on the internal dataset was 83.8 (external 96.7) for the dCNN and 84.6 ± 2.3 (external 90.9 ± 2.9) for the humans. CONCLUSION/CONCLUSIONS:dCNNs may be used to mimic human decision-making in the evaluation of single US images of breast lesion according to the BI-RADS catalog. The technique reaches high accuracies and may serve for standardization of highly observer-dependent US assessment. KEY POINTS/CONCLUSIONS:• Deep convolutional neural networks could be used to classify US breast lesions. • The implemented dCNN with its sliding window approach reaches high accuracies in the classification of US breast lesions. • Deep convolutional neural networks may serve for standardization in US BI-RADS classification.

Background Increasing evidence suggests a psychosomatic link between neural systems and the heart. In light of the growing burden of ischemic cardiovascular disease across the globe, a better understanding of heart-brain interactions and their implications for cardiovascular treatment strategies is needed. Thus, we sought to investigate the interaction between myocardial injury and metabolic alterations in central neural areas in patients with suspected or known coronary artery disease. Methods and Results The association between resting metabolic activity in distinct neural structures and cardiac function was analyzed in 302 patients (aged 66.8±10.2 years; 70.9% men) undergoing fluor-18-deoxyglucose positron emission tomography and ^99mTc-tetrofosmin single-photon emission computed tomography myocardial perfusion imaging. There was evidence for reduction of callosal, caudate, and brainstem fluor-18-deoxyglucose uptake in patients with impaired left ventricular ejection fraction (<55% versus ≥55%: P=0.047, P=0.022, and P=0.013, respectively) and/or in the presence of myocardial ischemia (versus normal perfusion: P=0.010, P=0.013, and P=0.016, respectively). In a sex-stratified analysis, these differences were observed in men, but not in women. A first-order interaction term consisting of sex and impaired left ventricular ejection fraction or myocardial ischemia was identified as predictor of metabolic activity in these neural regions (left ventricular ejection fraction: P=0.015 for brainstem; myocardial ischemia: P=0.004, P=0.018, and P=0.003 for callosal, caudate, or brainstem metabolism, respectively). Conclusions Myocardial dysfunction and injury are associated with reduced resting metabolic activity of central neural structures, including the corpus callosum, the caudate nucleus, and the brainstem. These associations differ in women and men, suggesting sex differences in the pathophysiological interplay of the nervous and cardiovascular systems.

BACKGROUND: Inflammation plays a fundamental role in mediating all stages of atherosclerosis. Given the higher prevalence of inflammatory rheumatologic conditions in women and the female propensity towards worse cardiovascular outcomes, refined strategies are needed to better identify the high-risk female cardiovascular phenotype. OBJECTIVES/OBJECTIVE: This article aims to assess sex-specific links between inflammatory processes and the development and progression of ischemic heart disease. PATIENTS AND METHODS/METHODS:Tc-tetrofosmin single-photon emission computed tomography myocardial perfusion imaging (SPECT-MPI). RESULTS: = 0.014). CONCLUSION/CONCLUSIONS: A strong link between bone marrow metabolism and impaired myocardial function and perfusion was observed in women, but not in men. Our data suggest that novel biomarkers of inflammation might help to identify women at risk for ischemic cardiomyopathy and to tailor disease management to the female cardiovascular phenotype.

PURPOSE/OBJECTIVE:To evaluate the diagnostic performance of bone texture analysis (TA) combined with machine learning (ML) algorithms in standard CT scans to identify patients with vertebrae at risk for insufficiency fractures. MATERIALS AND METHODS/METHODS:Standard CT scans of 58 patients with insufficiency fractures of the spine, performed between 2006 and 2013, were analyzed retrospectively. Every included patient had at least two CT scans. Intact vertebrae in a first scan that either fractured ("unstable") or remained intact ("stable") in the consecutive scan were manually segmented on mid-sagittal reformations. TA features for all vertebrae were extracted using open-source software (MaZda). In a paired control study, all vertebrae of the study cohort "cases" and matched controls were classified using ROC analysis of Hounsfield unit (HU) measurements and supervised ML techniques. In a within-subject vertebra comparison, vertebrae of the cases were classified into "unstable" and "stable" using identical techniques. RESULTS:One hundred twenty vertebrae were included. Classification of cases/controls using ROC analysis of HU measurements showed an AUC of 0.83 (95% confidence interval [CI], 0.77-0.88), and ML-based classification showed an AUC of 0.97 (CI, 0.97-0.98). Classification of unstable/stable vertebrae using ROC analysis showed an AUC of 0.52 (CI, 0.42-0.63), and ML-based classification showed an AUC of 0.64 (CI, 0.61-0.67). CONCLUSION/CONCLUSIONS:TA combined with ML allows to identifying patients who will suffer from vertebral insufficiency fractures in standard CT scans with high accuracy. However, identification of single vertebra at risk remains challenging. KEY POINTS/CONCLUSIONS:• Bone texture analysis combined with machine learning allows to identify patients at risk for vertebral body insufficiency fractures on standard CT scans with high accuracy. • Compared to mere Hounsfield unit measurements on CT scans, application of bone texture analysis combined with machine learning improve fracture risk prediction. • This analysis has the potential to identify vertebrae at risk for insufficiency fracture and may thus increase diagnostic value of standard CT scans.