An active learning approach to train a deep learning algorithm for tumor segmentation from brain MR images.

PURPOSE: This study focuses on assessing the performance of active learning techniques to train a brain MRI glioma segmentation model. METHODS: The publicly available training dataset provided for the 2021 RSNA-ASNR-MICCAI Brain Tumor Segmentation (BraTS) Challenge was used in this study, consisting of 1251 multi-institutional, multi-parametric MR images. Post-contrast T1, T2, and T2 FLAIR images as well as ground truth manual segmentation were used as input for the model. The data were split into a training set of 1151 cases and testing set of 100 cases, with the testing set remaining constant throughout. Deep convolutional neural network segmentation models were trained using the NiftyNet platform. To test the viability of active learning in training a segmentation model, an initial reference model was trained using all 1151 training cases followed by two additional models using only 575 cases and 100 cases. The resulting predicted segmentations of these two additional models on the remaining training cases were then addended to the training dataset for additional training. RESULTS: It was demonstrated that an active learning approach for manual segmentation can lead to comparable model performance for segmentation of brain gliomas (0.906 reference Dice score vs 0.868 active learning Dice score) while only requiring manual annotation for 28.6% of the data. CONCLUSION: The active learning approach when applied to model training can drastically reduce the time and labor spent on preparation of ground truth training data. CRITICAL RELEVANCE STATEMENT: Active learning concepts were applied to a deep learning-assisted segmentation of brain gliomas from MR images to assess their viability in reducing the required amount of manually annotated ground truth data in model training. KEY POINTS: • This study focuses on assessing the performance of active learning techniques to train a brain MRI glioma segmentation model. • The active learning approach for manual segmentation can lead to comparable model performance for segmentation of brain gliomas. • Active learning when applied to model training can drastically reduce the time and labor spent on preparation of ground truth training data.

Direct Comparison of Quantitative US versus Controlled Attenuation Parameter for Liver Fat Assessment Using MRI Proton Density Fat Fraction as the Reference Standard in Patients Suspected of Having NAFLD.

Background MRI-derived proton density fat fraction (PDFF) is an accurate, reliable, and safe biologic marker for use in the noninvasive diagnosis of hepatic steatosis in patients with nonalcoholic fatty liver disease (NAFLD). Because of the cost and limited availability of MRI, it is necessary to develop an accurate method to diagnose NAFLD with potential point-of-care access. Purpose To compare the diagnostic accuracy of the quantitative US (QUS) fat fraction (FF) estimator with that of the controlled attenuation parameter (CAP) in the diagnosis of NAFLD using contemporaneous MRI-derived PDFF as the reference standard. Materials and Methods Participants with or suspected of having NAFLD were prospectively recruited at the NAFLD Research Center between July 2015 and July 2019. All participants underwent MRI-derived PDFF measurement, transient elastography with CAP measurement, and QUS. QUS FF was derived using computed QUS parameters from the acquired radiofrequency US data using a calibrated reference phantom. The area under the receiver operating characteristic curve (AUC) was calculated to assess the accuracy of QUS FF and CAP in the diagnosis of hepatic steatosis (defined as MRI-derived PDFF ≥ 5%). AUCs were compared using the DeLong test. Results A total of 123 participants were included (mean age, 52 years ± 13 [SD]; 67 [54%] women). Of these participants, 100 (81%) had MRI-derived PDFF of 5% or more. QUS FF had a significantly higher AUC for diagnosis of NAFLD than did CAP (0.92 [95% CI: 0.87, 0.98] vs 0.79 [95% CI: 0.67, 0.90], P = .03). QUS FF had a sensitivity of 98% (98 of 100) and a specificity of 48% (11 of 23). CAP had a sensitivity of 87% (87 of 100) and a specificity of 57% (13 of 23). Conclusion The quantitative US fat fraction estimator is more accurate than the controlled attenuation parameter in the diagnosis of hepatic steatosis in patients with or suspected of having nonalcoholic fatty liver disease. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Ito in this issue.

Comparative diagnostic performance of ultrasound shear wave elastography and magnetic resonance elastography for classifying fibrosis stage in adults with biopsy-proven nonalcoholic fatty liver disease.

OBJECTIVES: To compare the diagnostic accuracy of US shear wave elastography (SWE) and magnetic resonance elastography (MRE) for classifying fibrosis stage in patients with nonalcoholic fatty liver disease (NAFLD). METHODS: Patients from a prospective single-center cohort with clinical liver biopsy for known or suspected NAFLD underwent contemporaneous SWE and MRE. AUCs for classifying biopsy-determined liver fibrosis stages ≥ 1, ≥ 2, ≥ 3, and = 4, and their respective performance parameters at cutoffs providing ≥ 90% sensitivity or specificity were compared between SWE and MRE. RESULTS: In total, 100 patients (mean age, 51.8 ± 12.9 years; 46% males; mean BMI 31.6 ± 4.7 kg/m2) with fibrosis stage distribution (stage 0/1/2/3/4) of 43, 36, 5, 10, and 6%, respectively, were included. AUCs (and 95% CIs) for SWE and MRE were 0.65 (0.54-0.76) and 0.81 (0.72-0.89), 0.81 (0.71-0.91) and 0.94 (0.89-1.00), 0.85 (0.74-0.96) and 0.95 (0.89-1.00), and 0.91 (0.79-1.00) and 0.92 (0.83-1.00), for detecting fibrosis stage ≥ 1, ≥ 2, ≥ 3, and = 4, respectively. The differences were significant for detecting fibrosis stage ≥ 1 and ≥ 2 (p < 0.01) but not otherwise. At ≥ 90% sensitivity cutoff, MRE yielded higher specificity than SWE at diagnosing fibrosis stage ≥ 1, ≥ 2, and ≥ 3. At ≥ 90% specificity cutoff, MRE yielded higher sensitivity than SWE at diagnosing fibrosis stage ≥ 1 and ≥ 2. CONCLUSIONS: In adults with NAFLD, MRE was more accurate than SWE in diagnosing stage ≥ 1 and ≥ 2 fibrosis, but not stage ≥ 3 or 4 fibrosis. KEY POINTS: • For detecting any fibrosis or mild fibrosis, MR elastography was significantly more accurate than shear wave elastography. • For detecting advanced fibrosis and cirrhosis, MRE and SWE did not differ significantly in accuracy. • For excluding advanced fibrosis and potentially ruling out the need for biopsy, SWE and MRE did not differ significantly in negative predictive value. • Neither SWE nor MRE had sufficiently high positive predictive value to rule in advanced fibrosis.

Liver Fat Assessment in Multiview Sonography Using Transfer Learning With Convolutional Neural Networks.

OBJECTIVES: To develop and evaluate deep learning models devised for liver fat assessment based on ultrasound (US) images acquired from four different liver views: transverse plane (hepatic veins at the confluence with the inferior vena cava, right portal vein, right posterior portal vein) and sagittal plane (liver/kidney). METHODS: US images (four separate views) were acquired from 135 participants with known or suspected nonalcoholic fatty liver disease. Proton density fat fraction (PDFF) values derived from chemical shift-encoded magnetic resonance imaging served as ground truth. Transfer learning with a deep convolutional neural network (CNN) was applied to develop models for diagnosis of fatty liver (PDFF ≥ 5%), diagnosis of advanced steatosis (PDFF ≥ 10%), and PDFF quantification for each liver view separately. In addition, an ensemble model based on all four liver view models was investigated. Diagnostic performance was assessed using the area under the receiver operating characteristics curve (AUC), and quantification was assessed using the Spearman correlation coefficient (SCC). RESULTS: The most accurate single view was the right posterior portal vein, with an SCC of 0.78 for quantifying PDFF and AUC values of 0.90 (PDFF ≥ 5%) and 0.79 (PDFF ≥ 10%). The ensemble of models achieved an SCC of 0.81 and AUCs of 0.91 (PDFF ≥ 5%) and 0.86 (PDFF ≥ 10%). CONCLUSION: Deep learning-based analysis of US images from different liver views can help assess liver fat.

Assessment of Hepatic Steatosis in Nonalcoholic Fatty Liver Disease by Using Quantitative US.

Background Advanced confounder-corrected chemical shift-encoded MRI-derived proton density fat fraction (PDFF) is a leading parameter for fat fraction quantification in nonalcoholic fatty liver disease (NAFLD). Because of the limited availability of this MRI technique, there is a need to develop and validate alternative parameters to assess liver fat. Purpose To assess relationship of quantitative US parameters to MRI PDFF and to develop multivariable quantitative US models to detect hepatic steatosis and quantify hepatic fat. Materials and Methods Adults with known NAFLD or who were suspected of having NAFLD were prospectively recruited between August 2015 and February 2019. Participants underwent quantitative US and chemical shift-encoded MRI liver examinations. Liver biopsies were performed if clinically indicated. The correlation between seven quantitative US parameters and MRI PDFF was evaluated. By using leave-one-out cross validation, two quantitative US multivariable models were evaluated: a classifier to differentiate participants with NAFLD versus participants without NAFLD and a fat fraction estimator. Classifier performance was summarized by area under the receiver operating characteristic curve and area under the precision-recall curve. Fat fraction estimator performance was evaluated by correlation, linearity, and bias. Results Included were 102 participants (mean age, 52 years ± 13 [standard deviation]; 53 women), 78 with NAFLD (MRI PDFF ≥ 5%). A two-variable classifier yielded a cross-validated area under the receiver operating characteristic curve of 0.89 (95% confidence interval: 0.82, 0.96) and an area under the precision-recall curve of 0.96 (95% confidence interval: 0.93, 0.99). The cross-validated fat fraction predicted by a two-variable fat fraction estimator was correlated with MRI PDFF (Spearman ρ = 0.82 [P < .001]; Pearson r = 0.76 [P < .001]). The mean bias was 0.02% (P = .97), and 95% limits of agreement were ±12.0%. The predicted fat fraction was linear with MRI PDFF (R 2 = 0.63; slope, 0.69; intercept, 4.3%) for MRI PDFF of 34% or less. Conclusion A multivariable quantitative US approach yielded excellent correlation with MRI proton density fat fraction for hepatic steatosis assessment in nonalcoholic fatty liver disease. © RSNA, 2020 Online supplemental material is available for this article.

Inter-platform reproducibility of ultrasonic attenuation and backscatter coefficients in assessing NAFLD.

OBJECTIVES: To assess inter-platform reproducibility of ultrasonic attenuation coefficient (AC) and backscatter coefficient (BSC) estimates in adults with known/suspected nonalcoholic fatty liver disease (NAFLD). METHODS: This HIPAA-compliant prospective study was approved by an institutional review board; informed consent was obtained. Participants with known/suspected NAFLD were recruited and underwent same-day liver examinations with clinical ultrasound scanner platforms from two manufacturers. Each participant was scanned by the same trained sonographer who performed multiple data acquisitions in the right liver lobe using a lateral intercostal approach. Each data acquisition recorded a B-mode image and the underlying radio frequency (RF) data. AC and BSC were calculated using the reference phantom method. Inter-platform reproducibility was evaluated for AC and log-transformed BSC (logBSC = 10log10BSC) by intraclass correlation coefficient (ICC), Pearson's correlation, Bland-Altman analysis with computation of limits of agreement (LOAs), and within-subject coefficient of variation (wCV; applicable to AC). RESULTS: Sixty-four participants were enrolled. Mean AC values measured using the two platforms were 0.90 ± 0.13 and 0.94 ± 0.15 dB/cm/MHz while mean logBSC values were - 30.6 ± 5.0 and - 27.9 ± 5.6 dB, respectively. Inter-platform ICC was 0.77 for AC and 0.70 for log-transformed BSC in terms of absolute agreement. Pearson's correlation coefficient was 0.81 for AC and 0.80 for logBSC. Ninety-five percent LOAs were - 0.21 to 0.13 dB/cm/MHz for AC, and - 9.48 to 3.98 dB for logBSC. The wCV was 7% for AC. CONCLUSIONS: Hepatic AC and BSC are reproducible across two different ultrasound platforms in adults with known or suspected NAFLD. KEY POINTS: • Ultrasonic attenuation coefficient and backscatter coefficient are reproducible between two different ultrasound platforms in adults with NAFLD. • This inter-platform reproducibility may qualify quantitative ultrasound biomarkers for generalized clinical application in patients with suspected/known NAFLD.

Inter-sonographer reproducibility of quantitative ultrasound outcomes and shear wave speed measured in the right lobe of the liver in adults with known or suspected non-alcoholic fatty liver disease.

OBJECTIVES: To assess inter-sonographer reproducibility of ultrasound attenuation coefficient (AC), backscatter coefficient (BSC) and shear wave speed (SWS) in adults with known/suspected non-alcoholic fatty liver disease (NAFLD). METHODS: The institutional review board approved this HIPAA-compliant prospective study; informed consent was obtained. Participants with known/suspected NAFLD were recruited and underwent same-day liver examinations with a clinical scanner. Each participant was scanned by two of the six trained sonographers. Each sonographer performed multiple data acquisitions in the right liver lobe using a lateral intercostal approach. A data acquisition was a single operator button press that recorded a B-mode image, radio-frequency data, and the SWS value. AC and BSC were calculated from the radio-frequency data using the reference phantom method. SWS was calculated automatically using product software. Intraclass correlation coefficient (ICC) and within-subject coefficient of variation (wCV) were calculated for applicable metrics. RESULTS: Sixty-one participants were recruited. Inter-sonographer ICC was 0.86 (95% confidence interval: 0.77-0.92) for AC and 0.87 (0.78-0.92) for log-transformed BSC (logBSC = 10log10BSC) using one acquisition per sonographer. ICC was 0.88 (0.80-0.93) for both AC and logBSC averaging 5 acquisitions. ICC for SWS was 0.57 (0.29-0.74) using one acquisition per sonographer, and 0.84 (0.66-0.93) using 10 acquisitions. The wCV was ~7% for AC, and 19-43% for SWS, depending on number of acquisitions. CONCLUSIONS: Hepatic AC, BSC and SWS measures on a clinical scanner have good inter-sonographer reproducibility in adults with known or suspected NAFLD. Multiple acquisitions are required for SWS but not AC or BSC to achieve good inter-sonographer reproducibility. KEY POINTS: • AC, BSC and SWS measurements are reproducible in adults with NAFLD. • Inter-sonographer reproducibility of SWS measurement improves with more acquisitions being averaged. • Multiple acquisitions are required for SWS but not AC or BSC.