Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation Classification Tasks in Multiple Sclerosis.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Artificial intelligence (AI) models often face performance drops after deployment to external datasets. This study evaluated the potential of a novel data augmentation framework based on generative adversarial networks (GAN) that creates synthetic patient image data during model training to improve model generalizability. Model development and external testing were performed for a given classification task, namely the detection of new fluid-attenuated inversion recovery (FLAIR) lesions on MRI during longitudinal follow-up of patients with multiple sclerosis (MS). An internal dataset of 669 patients with MS (n = 3083 examinations) was used to develop an attention-based network, trained both with and without the inclusion of the GAN-based synthetic data augmentation framework. External testing was performed on 134 patients with MS from a different institution, with MR images acquired using different scanners and protocols than images used during training. Models trained using synthetic data augmentation showed a significant performance improvement when applied on external data (AUC 83.6% without synthetic data versus AUC 93.3% with synthetic data augmentation, P = .03), achieving comparable results to the internal test set (AUC 95.5%, P = .53), whereas models without synthetic data augmentation demonstrated a performance drop upon external testing (AUC 93.8% on internal dataset versus AUC 83.6% on external data, P = .03). Data augmentation with synthetic patient data substantially improved performance of AI models on unseen MRI data and may be extended to other clinical conditions or tasks to mitigate domain shift, limit class imbalance, and enhance the robustness of AI applications in medical imaging. ©RSNA, 2024.

Artificial T1-Weighted Postcontrast Brain MRI: A Deep Learning Method for Contrast Signal Extraction.

OBJECTIVES: Reducing gadolinium-based contrast agents to lower costs, the environmental impact of gadolinium-containing wastewater, and patient exposure is still an unresolved issue. Published methods have never been compared. The purpose of this study was to compare the performance of 2 reimplemented state-of-the-art deep learning methods (settings A and B) and a proposed method for contrast signal extraction (setting C) to synthesize artificial T1-weighted full-dose images from corresponding noncontrast and low-dose images. MATERIALS AND METHODS: In this prospective study, 213 participants received magnetic resonance imaging of the brain between August and October 2021 including low-dose (0.02 mmol/kg) and full-dose images (0.1 mmol/kg). Fifty participants were randomly set aside as test set before training (mean age ± SD, 52.6 ± 15.3 years; 30 men). Artificial and true full-dose images were compared using a reader-based study. Two readers noted all false-positive lesions and scored the overall interchangeability in regard to the clinical conclusion. Using a 5-point Likert scale (0 being the worst), they scored the contrast enhancement of each lesion and its conformity to the respective reference in the true image. RESULTS: The average counts of false-positives per participant were 0.33 ± 0.93, 0.07 ± 0.33, and 0.05 ± 0.22 for settings A-C, respectively. Setting C showed a significantly higher proportion of scans scored as fully or mostly interchangeable (70/100) than settings A (40/100, P < 0.001) and B (57/100, P < 0.001), and generated the smallest mean enhancement reduction of scored lesions (-0.50 ± 0.55) compared with the true images (setting A: -1.10 ± 0.98; setting B: -0.91 ± 0.67, both P < 0.001). The average scores of conformity of the lesion were 1.75 ± 1.07, 2.19 ± 1.04, and 2.48 ± 0.91 for settings A-C, respectively, with significant differences among all settings (all P < 0.001). CONCLUSIONS: The proposed method for contrast signal extraction showed significant improvements in synthesizing postcontrast images. A relevant proportion of images showing inadequate interchangeability with the reference remains at this dosage.

External evaluation of a deep learning-based approach for automated brain volumetry in patients with huntington's disease.

A crucial step in the clinical adaptation of an AI-based tool is an external, independent validation. The aim of this study was to investigate brain atrophy in patients with confirmed, progressed Huntington's disease using a certified software for automated volumetry and to compare the results with the manual measurement methods used in clinical practice as well as volume calculations of the caudate nuclei based on manual segmentations. Twenty-two patients were included retrospectively, consisting of eleven patients with Huntington's disease and caudate nucleus atrophy and an age- and sex-matched control group. To quantify caudate head atrophy, the frontal horn width to intercaudate distance ratio and the intercaudate distance to inner table width ratio were obtained. The software mdbrain was used for automated volumetry. Manually measured ratios and automatically measured volumes of the groups were compared using two-sample t-tests. Pearson correlation analyses were performed. The relative difference between automatically and manually determined volumes of the caudate nuclei was calculated. Both ratios were significantly different between the groups. The automatically and manually determined volumes of the caudate nuclei showed a high level of agreement with a mean relative discrepancy of - 2.3 ± 5.5%. The Huntington's disease group showed significantly lower volumes in a variety of supratentorial brain structures. The highest degree of atrophy was shown for the caudate nucleus, putamen, and pallidum (all p < .0001). The caudate nucleus volume and the ratios were found to be strongly correlated in both groups. In conclusion, in patients with progressed Huntington's disease, it was shown that the automatically determined caudate nucleus volume correlates strongly with measured ratios commonly used in clinical practice. Both methods allowed clear differentiation between groups in this collective. The software additionally allows radiologists to more objectively assess the involvement of a variety of brain structures that are less accessible to standard semiquantitative methods.

Human brain clearance imaging: Pathways taken by magnetic resonance imaging contrast agents after administration in cerebrospinal fluid and blood.

Over the last decade, it has become evident that cerebrospinal fluid (CSF) plays a pivotal role in brain solute clearance through perivascular pathways and interactions between the brain and meningeal lymphatic vessels. Whereas most of this fundamental knowledge was gained from rodent models, human brain clearance imaging has provided important insights into the human system and highlighted the existence of important interspecies differences. Current gold standard techniques for human brain clearance imaging involve the injection of gadolinium-based contrast agents and monitoring their distribution and clearance over a period from a few hours up to 2 days. With both intrathecal and intravenous injections being used, which each have their own specific routes of distribution and thus clearance of contrast agent, a clear understanding of the kinetics associated with both approaches, and especially the differences between them, is needed to properly interpret the results. Because it is known that intrathecally injected contrast agent reaches the blood, albeit in small concentrations, and that similarly some of the intravenously injected agent can be detected in CSF, both pathways are connected and will, in theory, reach the same compartments. However, because of clear differences in relative enhancement patterns, both injection approaches will result in varying sensitivities for assessment of different subparts of the brain clearance system. In this opinion review article, the "EU Joint Programme - Neurodegenerative Disease Research (JPND)" consortium on human brain clearance imaging provides an overview of contrast agent pharmacokinetics in vivo following intrathecal and intravenous injections and what typical concentrations and concentration-time curves should be expected. This can be the basis for optimizing and interpreting contrast-enhanced MRI for brain clearance imaging. Furthermore, this can shed light on how molecules may exchange between blood, brain, and CSF.

Machine Learning-Based Perivascular Space Volumetry in Alzheimer Disease.

OBJECTIVES: Impaired perivascular clearance has been suggested as a contributing factor to the pathogenesis of Alzheimer disease (AD). However, it remains unresolved when the anatomy of the perivascular space (PVS) is altered during AD progression. Therefore, this study investigates the association between PVS volume and AD progression in cognitively unimpaired (CU) individuals, both with and without subjective cognitive decline (SCD), and in those clinically diagnosed with mild cognitive impairment (MCI) or mild AD. MATERIALS AND METHODS: A convolutional neural network was trained using manually corrected, filter-based segmentations (n = 1000) to automatically segment the PVS in the centrum semiovale from interpolated, coronal T2-weighted magnetic resonance imaging scans (n = 894). These scans were sourced from the national German Center for Neurodegenerative Diseases Longitudinal Cognitive Impairment and Dementia Study. Convolutional neural network-based segmentations and those performed by a human rater were compared in terms of segmentation volume, identified PVS clusters, as well as Dice score. The comparison revealed good segmentation quality (Pearson correlation coefficient r = 0.70 with P < 0.0001 for PVS volume, detection rate in cluster analysis = 84.3%, and Dice score = 59.0%). Subsequent multivariate linear regression analysis, adjusted for participants' age, was performed to correlate PVS volume with clinical diagnoses, disease progression, cerebrospinal fluid biomarkers, lifestyle factors, and cognitive function. Cognitive function was assessed using the Mini-Mental State Examination, the Comprehensive Neuropsychological Test Battery, and the Cognitive Subscale of the 13-Item Alzheimer's Disease Assessment Scale. RESULTS: Multivariate analysis, adjusted for age, revealed that participants with AD and MCI, but not those with SCD, had significantly higher PVS volumes compared with CU participants without SCD ( P = 0.001 for each group). Furthermore, CU participants who developed incident MCI within 4.5 years after the baseline assessment showed significantly higher PVS volumes at baseline compared with those who did not progress to MCI ( P = 0.03). Cognitive function was negatively correlated with PVS volume across all participant groups ( P ≤ 0.005 for each). No significant correlation was found between PVS volume and any of the following parameters: cerebrospinal fluid biomarkers, sleep quality, body mass index, nicotine consumption, or alcohol abuse. CONCLUSIONS: The very early changes of PVS volume may suggest that alterations in PVS function are involved in the pathophysiology of AD. Overall, the volumetric assessment of centrum semiovale PVS represents a very early imaging biomarker for AD.

Differentiation of pseudoprogression and real progression in glioblastoma using ADC parametric response maps.

PURPOSE: The purpose of this study was to investigate whether a voxel-wise analysis of apparent diffusion coefficient (ADC) values may differentiate between progressive disease (PD) and pseudoprogression (PsP) in patients with high-grade glioma using the parametric response map, a newly introduced postprocessing tool. METHODS: Twenty-eight patients with proven PD and seven patients with PsP were identified in this retrospective feasibility study. For all patients ADC baseline and follow-up maps on four subsequent MRIs were available. ADC maps were coregistered on contrast enhanced T1-weighted follow-up images. Subsequently, enhancement in the follow-up contrast enhanced T1-weighted image was manually delineated and a reference region of interest (ROI) was drawn in the contralateral white matter. Both ROIs were transferred to the ADC images. Relative ADC (rADC) (baseline)/reference ROI values and rADC (follow up)/reference ROI values were calculated for each voxel within the ROI. The corresponding voxels of rADC (follow up) and rADC (baseline) were subtracted and the percentage of all voxels within the ROI that exceeded the threshold of 0.25 was quantified. RESULTS: rADC voxels showed a decrease of 59.2% (1st quartile (Q1) 36.7; 3rd quartile (Q3) 78.6) above 0.25 in patients with PD and 18.6% (Q1 3.04; Q3 26.5) in patients with PsP (p = 0.005). Receiver operating characteristic curve analysis showed the optimal decreasing rADC cut-off value for identifying PD of > 27.05% (area under the curve 0.844±0.065, sensitivity 0.86, specificity 0.86, p = 0.014). CONCLUSION: This feasibility study shows that the assessment of rADC using parametric response maps might be a promising approach to contribute to the differentiation between PD and PsP. Further research in larger patient cohorts is necessary to finally determine its clinical utility.

Prognostic value of combined visualization of MR diffusion and perfusion maps in glioblastoma.

We analyzed whether the combined visualization of decreased apparent diffusion coefficient (ADC) values and increased cerebral blood volume (CBV) in perfusion imaging can identify prognosis-related growth patterns in patients with newly diagnosed glioblastoma. Sixty-five consecutive patients were examined with diffusion and dynamic susceptibility-weighted contrast-enhanced perfusion weighted MRI. ADC and CBV maps were co-registered on the T1-w image and a region of interest (ROI) was manually delineated encompassing the enhancing lesion. Within this ROI pixels with ADC values the 70th percentile (CBVmax) and the intersection of pixels with ADCmin and CBVmax were automatically calculated and visualized. Initially, all tumors with a mean intersection greater than the upper quartile of the normally distributed mean intersection of all patients were subsumed to the first growth pattern termed big intersection (BI). Subsequently, the remaining tumors' growth patterns were categorized depending on the qualitative representation of ADCmin, CBVmax and their intersection. Log-rank test exposed a significantly longer overall survival of BI (n = 16) compared to non-BI group (n = 49) (p = 0.0057). Thirty-one, four and 14 patients of the non-BI group were classified as predominant ADC-, CBV- and mixed growth group, respectively. In a multivariate Cox regression model, the BI-, CBV- and mixed groups had significantly lower adjusted hazard ratios (p-value, α(Bonferroni) < 0.006) when compared to the reference group ADC: 0.29 (0.0027), 0.11 (0.038) and 0.33 (0.0059). Our study provides evidence that the combination of diffusion and perfusion imaging allows visualization of different glioblastoma growth patterns that are associated with prognosis. A possible biological hypothesis for this finding could be the interpretation of the ADCmin fraction as the invasion-front of tumor cells while the CBVmax fraction might represent the vascular rich tumor border that is "trailing behind" the invasion-front in the ADC group.

Association of overall survival in patients with newly diagnosed glioblastoma with contrast-enhanced perfusion MRI: Comparison of intraindividually matched T1 - and T2 (*) -based bolus techniques.

BACKGROUND: To compare intraindividual dynamic susceptibility contrast (DSC) and dynamic contrast enhanced (DCE) MR perfusion parameters and determine the association of DCE parameters with overall survival (OS) with the established predictive DSC parameter cerebral blood volume (CBV) in patients with newly diagnosed glioblastoma. METHODS: Perfusion data were analyzed retrospectively, and included scans performed preoperatively at 3.0 Tesla in 37 patients (25 males, 12 females, 39-83 years, median 65) later diagnosed with glioblastoma. All patients received standard treatment consisting of surgery and radiochemotherapy. Images were spatially coregistered and maximum region of interest-based DCE and DSC parameter measurements compared and thresholds identified using multivariate linear regression, Pearson's correlation coefficients and using receiver operating characteristic analysis. Survival analysis was performed using Kaplan-Meier curves. RESULTS: While both, elevated volume transfer constant (K(trans) ) (>0.29 min(-1) ; P = 0.041) and CBV (>23.7 mL/100 mL; P < 0.001) were significantly associated with OS, elevated CBV was associated with worse OS compared with elevated K(trans) . K(trans) was significantly correlated with the leakage correction factor K2 but not with CBV. CONCLUSION: The combined use of DSC and DCE MR perfusion may provide additional information of prognostic value for glioblastoma patient survival prediction. As K(trans) was not tightly coupled to CBV, both parameters may reflect different stages in the pathogenetic sequence of glioblastoma growth.