Transformer-based structuring of free-text radiology report databases.

OBJECTIVES: To provide insights for on-site development of transformer-based structuring of free-text report databases by investigating different labeling and pre-training strategies. METHODS: A total of 93,368 German chest X-ray reports from 20,912 intensive care unit (ICU) patients were included. Two labeling strategies were investigated to tag six findings of the attending radiologist. First, a system based on human-defined rules was applied for annotation of all reports (termed "silver labels"). Second, 18,000 reports were manually annotated in 197 h (termed "gold labels") of which 10% were used for testing. An on-site pre-trained model (Tmlm) using masked-language modeling (MLM) was compared to a public, medically pre-trained model (Tmed). Both models were fine-tuned on silver labels only, gold labels only, and first with silver and then gold labels (hybrid training) for text classification, using varying numbers (N: 500, 1000, 2000, 3500, 7000, 14,580) of gold labels. Macro-averaged F1-scores (MAF1) in percent were calculated with 95% confidence intervals (CI). RESULTS: Tmlm,gold (95.5 [94.5-96.3]) showed significantly higher MAF1 than Tmed,silver (75.0 [73.4-76.5]) and Tmlm,silver (75.2 [73.6-76.7]), but not significantly higher MAF1 than Tmed,gold (94.7 [93.6-95.6]), Tmed,hybrid (94.9 [93.9-95.8]), and Tmlm,hybrid (95.2 [94.3-96.0]). When using 7000 or less gold-labeled reports, Tmlm,gold (N: 7000, 94.7 [93.5-95.7]) showed significantly higher MAF1 than Tmed,gold (N: 7000, 91.5 [90.0-92.8]). With at least 2000 gold-labeled reports, utilizing silver labels did not lead to significant improvement of Tmlm,hybrid (N: 2000, 91.8 [90.4-93.2]) over Tmlm,gold (N: 2000, 91.4 [89.9-92.8]). CONCLUSIONS: Custom pre-training of transformers and fine-tuning on manual annotations promises to be an efficient strategy to unlock report databases for data-driven medicine. KEY POINTS: • On-site development of natural language processing methods that retrospectively unlock free-text databases of radiology clinics for data-driven medicine is of great interest. • For clinics seeking to develop methods on-site for retrospective structuring of a report database of a certain department, it remains unclear which of previously proposed strategies for labeling reports and pre-training models is the most appropriate in context of, e.g., available annotator time. • Using a custom pre-trained transformer model, along with a little annotation effort, promises to be an efficient way to retrospectively structure radiological databases, even if not millions of reports are available for pre-training.

Correction to: Is liver lesion characterisation by simplified IVIM DWI also feasible at 3.0 T?

The original version of this article, published on 08 April 2019, unfortunately contained a mistake. The following correction has therefore been made in the original: The caption of Fig. 2 is wrong. The corrected version is given below.

Is liver lesion characterisation by simplified IVIM DWI also feasible at 3.0 T?

OBJECTIVE: To evaluate simplified intravoxel incoherent motion (IVIM) diffusion-weighted imaging (DWI) for liver lesion characterisation at 3.0 T and to compare it with 1.5 T. METHODS: 3.0-T DWI data from a respiratory-gated MRI sequence with b = 0, 50, 250, and 800 s/mm2 were analysed in 116 lesions (78 patients) and 27 healthy livers. Apparent diffusion coefficient ADC = ADC(0,800) and IVIM-based parameters D1' = ADC(50,800), D2' = ADC(250,800), f1' = f(0,50,800), f2' = f(0,250,800), D*' = D*(0,50,250,800), ADClow = ADC(0,50), and ADCdiff = ADClow-D2' were calculated voxel-wise and analysed on per-patient basis. Results were compared with those of 173 lesions (110 patients) and 40 healthy livers at 1.5 T. RESULTS: Focal nodular hyperplasias were best discriminated from all other lesions by f1' and haemangiomas by D1' with an area under the curve (AUC) of 0.993 and 1.000, respectively. For discrimination between malignant and benign lesions, ADC was best suited (AUC of 0.968). The combination of D1' and f1' correctly identified more lesions as malignant or benign than the ADC (99.1% vs 88.8%). Discriminatory power for differentiating malignant from benign lesions tended to be higher at 3.0 T than at 1.5 T. CONCLUSION: Simplified IVIM is suitable for lesion characterisation at 3.0 T with a trend of superior diagnostic accuracy for discriminating malignant from benign lesions compared with 1.5 T. KEY POINTS: • Simplified IVIM is also suitable for liver lesion characterisation at 3.0 T. • Excellent accuracy was reached for discriminating malignant from benign lesions. • The acquisition of only three b-values (0, 50, 800 s/mm 2 ) is required.

Accurate IVIM model-based liver lesion characterisation can be achieved with only three b-value DWI.

OBJECTIVE: The objective of this study was to evaluate a simplified intravoxel incoherent motion (IVIM) approach of diffusion-weighted imaging (DWI) with four b-values for liver lesion characterisation at 1.5 T. METHODS: DWI data from a respiratory-gated MRI sequence with b = 0, 50, 250, 800 s/mm2 were retrospectively analysed in 173 lesions and 40 healthy livers. The apparent diffusion coefficient ADC = ADC(0,800) and IVIM-based parameters D1' = ADC(50,800), D2' =ADC(250,800), f1', f2', D*', ADClow = ADC(0,50), and ADCdiff=ADClow-D2' were calculated voxel-wise without fitting procedures. Differences between lesion groups were investigated. RESULTS: Focal nodular hyperplasias were best discriminated from all other lesions by f1' with an area under the curve (AUC) of 0.989. Haemangiomas were best discriminated by D1' (AUC of 0.994). For discrimination between malignant and benign lesions, ADC(0,800) and D1' were best suited (AUC of 0.915 and 0.858, respectively). Discriminatory power was further increased by using a combination of D1' and f1'. CONCLUSION: IVIM parameters D and f approximated from three b-values provided more discriminatory power between liver lesions than ADC determined from two b-values. The use of b = 0, 50, 800 s/mm2 was superior to that of b = 0, 250, 800 s/mm2. The acquisition of four instead of three b-values has no further benefit for lesion characterisation. KEY POINTS: • Diffusion and perfusion characteristics are assessable with only three b-values. • Association of b = 0, 50, 800 s/mm2is superior to b = 0, 250, 800 s/mm2. • A fourth acquired b-value has no benefit for differential diagnosis. • For liver lesion characterisation, simplified IVIM analysis is superior to ADC determination. • Simplified IVIM approach guarantees numerically stable, voxel-wise results and short acquisition times.

Intravoxel incoherent motion model-based liver lesion characterisation from three b-value diffusion-weighted MRI.

OBJECTIVE: To evaluate intravoxel incoherent motion (IVIM) model-based liver lesion characterisation from three b-value diffusion-weighted imaging (DWI). METHODS: The 1.5-T DWI data from a respiratory gated spin-echo echo-planar magnetic resonance imaging sequence (b = 0, 50, 800 s/mm(2)) were retrospectively analysed in 38 patients with different liver lesions. Conventional apparent diffusion coefficient ADC = ADC(0,800) as well as IVIM-based parameters D' = ADC(50,800), ADC_low = ADC(0,50), and f' were calculated voxel-wise. Sixty-one regions of interest in hepatocellular carcinomas (HCCs, n = 24), haemangiomas (HEMs, n = 11), focal nodular hyperplasias (FNHs, n = 11), and healthy liver tissue (REFs, n = 15) were analysed. Group differences were investigated using Student's t-test and receiver-operating characteristic (ROC) analysis. RESULTS: Mean values ± standard deviations of ADC, D', ADC_low (in 10(-5) mm(2)/s), and f' (in %) for REFs/FNHs/HEMs/HCCs were 130 ± 11/143 ± 27/168 ± 16/113 ± 25, 104 ± 12/123 ± 25/162 ± 18/102 ± 23, 518 ± 66/437 ± 97/268 ± 69/283 ± 120, and 18 ± 3/14 ± 4/6 ± 3/9 ± 5, respectively. Differences between lesions and REFs were more significant for IVIM-based parameters than for conventional ADC. ROC analysis showed the best discriminability between HCCs and FNHs for ADC_low and f' and between HEMs and FNHs or HCCs for D'. CONCLUSION: Three instead of two b-value DWI enables a numerically stable and voxel-wise IVIM-based analysis for improved liver lesion characterisation with tolerable acquisition time. KEY POINTS: • Quantitative analysis of diffusion-weighted MRI helps liver lesion characterisation. • Analysis of intravoxel incoherent motion is superior to apparent diffusion coefficient determination. • Only three b-values enable separation of diffusion and microcirculation effects. • The method presented is numerically stable, with voxel-wise results and short acquisition times.

Quantitative MRI of the liver: Evaluation of extracellular volume fraction and other quantitative parameters in comparison to MR elastography for the assessment of hepatopathy.

BACKGROUND: Chronic liver diseases pose a major health problem worldwide, while common tests for diagnosis and monitoring of diffuse hepatopathy have considerable limitations. Preliminary data on the quantification of hepatic extracellular volume fraction (ECV) with magnetic resonance imaging (MRI) for non-invasive assessment of liver fibrosis are encouraging, with ECV having the potential to overcome several of these constraints. PURPOSE: To clinically evaluate ECV provided by quantitative MRI for assessing the severity of liver disease. MATERIALS AND METHODS: In this prospective study, multiparametric liver MRI, including T1 mapping and magnetic resonance elastography (MRE), was performed in subjects with and without hepatopathy between November 2018 and October 2019. T1, T2, T2*, proton density fat fraction and stiffness were extracted from parametric maps by regions of interest and ECV was calculated from T1 relaxometries. Serum markers of liver disease were obtained by clinical database research. For correlation analysis, Spearman rank correlation was used. ROC analysis of serum markers and quantitative MRI data for discrimination of liver cirrhosis was performed with MRE as reference standard. RESULTS: 109 participants were enrolled (50.7 ± 16.1 years, 61 men). ECV, T1 and MRE correlated significantly with almost all serum markers of liver disease, with ECV showing the strongest associations (up to r = 0.67 with MELD, p < 0.01). ECV and T1 correlated with MRE (0.75 and 0.73, p < 0.01 each). ECV (AUC 0.89, cutoff 32.2%, sensitivity 85%, specificity 87%) and T1 mapping (AUC 0.85, cutoff 592.5 ms, sensitivity 83%, specificity 75%) featured good performances in detection of liver cirrhosis with only ECV performing significantly superior to model of end stage liver disease (MELD), AST/ALT ratio and international normalized ratio (p < 0.01, respectively). CONCLUSION: Quantification of hepatic extracellular volume fraction with MRI is suitable for estimating the severity of liver disease when using MRE as the standard of reference. It represents a promising tool for non-invasive assessment of liver fibrosis and cirrhosis.

Comparison of magnetic resonance feature tracking with harmonic phase imaging analysis (CSPAMM) for assessment of global and regional diastolic function.

AIMS: Complex post-processing is required for strain-derived assessment of diastolic dysfunction (DD) using CMR-tagging (TAG). Feature-tracking (FT), allows for rapid systolic strain assessment using conventional steady-state free precession (SSFP)-Cine sequences. Aim of this study was to investigate whether FT may be employed for the clinically applicable quantification of DD. METHODS AND RESULTS: 40 individuals (20 patients with DD I-III°, 20 controls) were investigated. CSPAMM and SSFP-Cine sequences were acquired in identical short-axis locations. Global and regional early diastolic strain rate (EDSR), peak diastolic strain rate (PDSR), twist, untwist and torsion were calculated from tagged and SSFP-Cine datasets. DD indices were compared, intra- as well inter-observer variability assessed. RESULTS: for global EDSR correlated strongly (r=0.94), revealed good agreement and no significant differences between both methods. Correlation for regional EDSR was lower, results differed significantly in the anterior wall (p<0.05). Correlation for PDSR was moderate (r=0.63), results in the healthy control group differed significantly (p<0.05). FT derived rotational indices correlated poorly with TAG (twist: r=0.28; untwist: r=0.02; torsion: r=0.26), subgroup analysis revealed significant differences (p<0.05). Intra- and inter-observer variability for FT derived global EDSR and PDSR were comparable to TAG, but significantly higher for regional EDSR and rotational indices. CONCLUSION: FT derived global EDSR allows for rapid clinical determination of diastolic dysfunction, revealing good agreement with TAG and low intra- as well as interobserver variability. However, TAG analysis not only yields higher accuracy and reproducibility of global- and regional diastolic strain, but also delivers reliable information about diastolic rotational and untwisting dynamics.

Characterization of the failing murine heart in a desmin knock-out model using a clinical 3 T MRI scanner.

The purpose of this study was to establish an MRI protocol on a clinical scanner for assessment of left (LV) and right (RV) ventricular myocardial function of the murine heart, and to apply this protocol for the first in vivo assessment of myocardial function in a mouse model of cardiomyopathy (Desmin-/-). MRI was performed on a clinical 3 T whole body MRI system using a dedicated solenoid receive-only coil. Contiguous short axis slices were acquired covering the entire heart using a spoiled cine gradient echo sequence (TR 9-12 ms, TE 3-4 ms, α 25°, 1.0 × 0.23 × 0.23 mm³). Global LV- and RV-myocardial functional parameters such as end-diastolic ventricular volume, ejection fraction (EF), LV mass and cardiac output (CO) of Desmin-/- mice and age-matched controls were determined. Global myocardial functional data of healthy controls (n = 4) were in very good agreement with previously reported data. The transgenic mice (n = 8) revealed a significantly reduced LV- and RV-EF as well as CO. Body weight-normalized LV- and RV-end-diastolic volumes and LV mass were significantly increased. In addition desmin deficient mice exhibited segmental wall thinning and akinesia, suggesting myocardial necrosis. This study demonstrates that clinical 3 T MRI-systems may reliably be used for non-invasive assessment of LV- and RV-myocardial function in normal and in genetically engineered mice with cardiomyopathies. In addition, this proof of principle study presents first in vivo MRI data of the cardiac phenotype of desmin knock-out mice.