Ethical Considerations and Fairness in the Use of Artificial Intelligence for Neuroradiology.

In this review, concepts of algorithmic bias and fairness are defined qualitatively and mathematically. Illustrative examples are given of what can go wrong when unintended bias or unfairness in algorithmic development occurs. The importance of explainability, accountability, and transparency with respect to artificial intelligence algorithm development and clinical deployment is discussed. These are grounded in the concept of "primum no nocere" (first, do no harm). Steps to mitigate unfairness and bias in task definition, data collection, model definition, training, testing, deployment, and feedback are provided. Discussions on the implementation of fairness criteria that maximize benefit and minimize unfairness and harm to neuroradiology patients will be provided, including suggestions for neuroradiologists to consider as artificial intelligence algorithms gain acceptance into neuroradiology practice and become incorporated into routine clinical workflow.

Repeatability of Automated Image Segmentation with BraTumIA in Patients with Recurrent Glioblastoma.

BACKGROUND AND PURPOSE: Despite high interest in machine-learning algorithms for automated segmentation of MRIs of patients with brain tumors, there are few reports on the variability of segmentation results. The purpose of this study was to obtain benchmark measures of repeatability for a widely accessible software program, BraTumIA (Versions 1.2 and 2.0), which uses a machine-learning algorithm to segment tumor features on contrast-enhanced brain MR imaging. MATERIALS AND METHODS: Automatic segmentation of enhancing tumor, tumor edema, nonenhancing tumor, and necrosis was performed on repeat MR imaging scans obtained approximately 2 days apart in 20 patients with recurrent glioblastoma. Measures of repeatability and spatial overlap, including repeatability and Dice coefficients, are reported. RESULTS: Larger volumes of enhancing tumor were obtained on later compared with earlier scans (mean, 26.3 versus 24.2 mL for BraTumIA 1.2; P < .05; and 24.9 versus 22.9 mL for BraTumIA 2.0, P < .01). In terms of percentage change, repeatability coefficients ranged from 31% to 46% for enhancing tumor and edema components and from 87% to 116% for nonenhancing tumor and necrosis. Dice coefficients were highest (>0.7) for enhancing tumor and edema components, intermediate for necrosis, and lowest for nonenhancing tumor and did not differ between software versions. Enhancing tumor and tumor edema were smaller, and necrotic tumor larger using BraTumIA 2.0 rather than 1.2. CONCLUSIONS: Repeatability and overlap metrics varied by segmentation type, with better performance for segmentations of enhancing tumor and tumor edema compared with other components. Incomplete washout of gadolinium contrast agents could account for increasing enhancing tumor volumes on later scans.

Artificial Intelligence in Neuroradiology: Current Status and Future Directions.

Fueled by new techniques, computational tools, and broader availability of imaging data, artificial intelligence has the potential to transform the practice of neuroradiology. The recent exponential increase in publications related to artificial intelligence and the central focus on artificial intelligence at recent professional and scientific radiology meetings underscores the importance. There is growing momentum behind leveraging artificial intelligence techniques to improve workflow and diagnosis and treatment and to enhance the value of quantitative imaging techniques. This article explores the reasons why neuroradiologists should care about the investments in new artificial intelligence applications, highlights current activities and the roles neuroradiologists are playing, and renders a few predictions regarding the near future of artificial intelligence in neuroradiology.

Quantitative Delta T1 (dT1) as a Replacement for Adjudicated Central Reader Analysis of Contrast-Enhancing Tumor Burden: A Subanalysis of the American College of Radiology Imaging Network 6677/Radiation Therapy Oncology Group 0625 Multicenter Brain Tumor Trial.

BACKGROUND AND PURPOSE: Brain tumor clinical trials requiring solid tumor assessment typically rely on the 2D manual delineation of enhancing tumors by ≥2 expert readers, a time-consuming step with poor interreader agreement. As a solution, we developed quantitative dT1 maps for the delineation of enhancing lesions. This retrospective analysis compares dT1 with 2D manual delineation of enhancing tumors acquired at 2 time points during the post therapeutic surveillance period of the American College of Radiology Imaging Network 6677/Radiation Therapy Oncology Group 0625 (ACRIN 6677/RTOG 0625) clinical trial. MATERIALS AND METHODS: Patients enrolled in ACRIN 6677/RTOG 0625, a multicenter, randomized Phase II trial of bevacizumab in recurrent glioblastoma, underwent standard MR imaging before and after treatment initiation. For 123 patients from 23 institutions, both 2D manual delineation of enhancing tumors and dT1 datasets were evaluable at weeks 8 (n = 74) and 16 (n = 57). Using dT1, we assessed the radiologic response and progression at each time point. Percentage agreement with adjudicated 2D manual delineation of enhancing tumor reads and association between progression status and overall survival were determined. RESULTS: For identification of progression, dT1 and adjudicated 2D manual delineation of enhancing tumor reads were in perfect agreement at week 8, with 73.7% agreement at week 16. Both methods showed significant differences in overall survival at each time point. When nonprogressors were further divided into responders versus nonresponders/nonprogressors, the agreement decreased to 70.3% and 52.6%, yet dT1 showed a significant difference in overall survival at week 8 (P = .01), suggesting that dT1 may provide greater sensitivity for stratifying subpopulations. CONCLUSIONS: This study shows that dT1 can predict early progression comparable with the standard method but offers the potential for substantial time and cost savings for clinical trials.

Repeatability of Standardized and Normalized Relative CBV in Patients with Newly Diagnosed Glioblastoma.

BACKGROUND AND PURPOSE: For more widespread clinical use advanced imaging methods such as relative cerebral blood volume must be both accurate and repeatable. The aim of this study was to determine the repeatability of relative CBV measurements in newly diagnosed glioblastoma multiforme by using several of the most commonly published estimation techniques. MATERIALS AND METHODS: The relative CBV estimates were calculated from dynamic susceptibility contrast MR imaging in double-baseline examinations for 33 patients with treatment-naïve and pathologically proved glioblastoma multiforme (men = 20; mean age = 55 years). Normalized and standardized relative CBV were calculated by using 6 common postprocessing methods. The repeatability of both normalized and standardized relative CBV, in both tumor and contralateral brain, was examined for each method with metrics of repeatability, including the repeatability coefficient and within-subject coefficient of variation. The minimum sample size required to detect a parameter change of 10% or 20% was also determined for both normalized relative CBV and standardized relative CBV for each estimation method. RESULTS: When ordered by the repeatability coefficient, methods using postprocessing leakage correction and ΔR2*(t) techniques offered superior repeatability. Across processing techniques, the standardized relative CBV repeatability in normal-appearing brain was comparable with that in tumor (P = .31), yet inferior in tumor for normalized relative CBV (P = .03). On the basis of the within-subject coefficient of variation, tumor standardized relative CBV estimates were less variable (13%-20%) than normalized relative CBV estimates (24%-67%). The minimum number of participants needed to detect a change of 10% or 20% is 118-643 or 30-161 for normalized relative CBV and 109-215 or 28-54 for standardized relative CBV. CONCLUSIONS: The ΔR2* estimation methods that incorporate leakage correction offer the best repeatability for relative CBV, with standardized relative CBV being less variable and requiring fewer participants to detect a change compared with normalized relative CBV.

Repeatability of quantitative metrics derived from MR diffusion tractography in paediatric patients with epilepsy.

OBJECTIVE: To quantify the test-retest repeatability of mean diffusivity (MD) and fractional anisotropy (FA) derived from diffusion tensor imaging (DTI) tractography in a cohort of paediatric patients with localization-related epilepsy. METHODS: 30 patients underwent 2 DTI acquisitions [repetition time/echo time (ms), 7000/90; flip, 90°; b-value, 1000 s mm(-2); voxel (mm), 2 × 2 × 2]. Two observers used Diffusion Toolkit and TrackVis ( www.trackvis.org ) to segment and analyse the following tracts: corpus callosum, corticospinal tracts, arcuate fasciculi, inferior longitudinal fasciculi and inferior fronto-occipital fasciculi. Mean MD and mean FA were calculated for each tract. Each observer independently analysed one of the DTI data sets for every patient. RESULTS: Segmentation identified all tracts in all subjects, except the arcuate fasciculus. There was a highly consistent relationship between repeated observations of MD (r = 0.993; p < 0.0001) and FA (r = 0.990; p < 0.0001). For each tract, coefficients of variation ranged from 0.9% to 2.1% for MD and from 1.5% to 2.8% for FA. The 95% confidence limits (CLs) for change ranged from 2.8% to 6% for MD and from 4.3% to 8.6% for FA. For the arcuate fasciculus, Cohen's κ for agreement between the observers (identifiable vs not identifiable) was 1.0. CONCLUSION: We quantified the repeatability of two commonly utilized scalar metrics derived from DTI tractography. For an individual patient, changes greater than the repeatability coefficient or 95% CLs for change are unlikely to be related to variability in their measurement. ADVANCES IN KNOWLEDGE: Reproducibility of these metrics will aid in the design of future studies and might one day be used to guide management in patients with epilepsy.

A change in the apparent diffusion coefficient after treatment with bevacizumab is associated with decreased survival in patients with recurrent glioblastoma multiforme.

OBJECTIVES: The aim of this study was to determine the prognostic significance of changes in parameters derived from diffusion tensor imaging (DTI) that occur in response to treatment with bevacizumab and irinotecan in patients with recurrent glioblastoma multiforme. METHODS: 15 patients with recurrent glioblastoma multiforme underwent serial 1.5 T MRI. Axial single-shot echo planar DTI was obtained on scans performed 3 days and 1 day prior to and 6 weeks after initiation of therapy with bevacizumab and irinotecan. Apparent diffusion coefficient (ADC) and fractional anisotropy (FA) maps were registered to whole brain contrast-enhanced three-dimensional (3D) spoiled gradient recalled and 3D fluid attenuation inversion recovery (FLAIR) image volumes. Anatomic image volumes were segmented to isolate regions of interest defined by tumour-related enhancement (TRE) and FLAIR signal abnormality (FSA). Mean ADC and mean FA were calculated for each region. A Bland-Altman repeatability coefficient was also calculated for each parameter based on the two pre-treatment studies. A patient was considered to have a change in FA or ADC after therapy if the difference between the pre- and post-treatment values was greater than the repeatability coefficient for that parameter. Survival was compared using a Cox proportional hazard model. RESULTS: DTI detected a change in ADC within FSA after therapy in nine patients (five in whom ADC was increased; four in whom it was decreased). Patients with a change in ADC within FSA had significantly shorter overall survival (p=0.032) and progression free survival (p=0.046) than those with no change. CONCLUSION: In patients with recurrent glioblastoma multiforme treated with bevacizumab and irinotecan, a change in ADC after therapy in FSA is associated with decreased survival.

Prognostic significance of parameters derived from co-registered 18F-fluorodeoxyglucose PET and contrast-enhanced MRI in patients with high-grade glioma.

OBJECTIVE: The aim of this study was to determine the prognostic significance of the volume and intensity of abnormal (18)F-fluorodeoxyglucose positron emission tomography (FDG-PET) accumulation within areas of contrast enhancement on post-therapeutic volumetric MRI. METHODS: A total of 10 patients with Grade III or IV glioma were treated with resection followed by intracavitary radiation therapy with (131)I-labelled antitenascin monoclonal antibody. Patients underwent serial FDG-PET and 1.5 T MR imaging. For each patient, MR and FDG-PET image volumes at each time point were aligned using a rigid-body normalised mutual information algorithm. Contrast-enhancing regions of interest (ROIs) were defined using a semi-automated k-means clustering technique. Activity within the ROI on the co-registered PET scan was calculated as a ratio (mean activity ratio; MAR) to activity in contralateral normal-appearing white matter (NAWM). The PET lesion was defined as the portion of the ROI associated with activity greater than two standard deviations above the mean in NAWM. Survival was assessed using the logrank test. RESULTS: Larger contrast-enhancing ROIs were strongly associated with an increased MAR (r = 0.51; p<0.002). Enhancing lesions with an MAR >1.2 were associated with decreased survival (p<0.016). In nine patients who died, the MAR on PET correlated inversely with survival duration (r = -0.43; p<0.01), whereas PET lesion volume did not. CONCLUSION: Following intracavitary radiation therapy, the development of contrast-enhancing lesions that are associated with high mean FDG-PET accumulation suggests poor prognosis.

Trichothiodystrophy with dysmyelination and central osteosclerosis.

Trichothiodystrophy (TTD) is a rare group of autosomal recessive disorders of DNA repair unified by the presence of sulfur-deficient brittle hair. We report a 3-year-old boy with classic clinical features of TTD, including ichthyosis, alopecia, developmental delay, and tiger-tail banding of the hair shaft on polarizing microscopy. Brain MR imaging showed both diffuse dysmyelination and osteosclerosis, findings that, in combination, may be specific for TTD.

Ruptured maxillary retention cyst: cause of unilateral rhinorrhea after trauma.

This study describes a case of a patient with traumatic rupture of a maxillary sinus retention cyst, which had an interesting clinical presentation of unilateral rhinorrhea, mimicking a CSF leak. The diagnosis was made fortuitously by comparison of a posttraumatic CT brain examination with a CT sinus study performed 1 day earlier.

Quantitative assessment of diffusion abnormalities in posterior reversible encephalopathy syndrome.

BACKGROUND AND PURPOSE: Previous studies have shown that lesions in posterior reversible encephalopathy syndrome are often isointense on diffusion-weighted MR images. We hypothesized that 1) apparent diffusion coefficient (ADC) maps using various thresholds would show larger abnormalities in posterior white matter (WM) and 2) isointense appearance of lesions on isotropic diffusion-weighted images results from a balance of T2 prolongation effects and diffusibility effects. METHODS: T2-weighted MR images from 11 patients were reviewed. Hyperintense lesions were located in both anterior and posterior WM in eight patients and solely in posterior WM in three patients. The ADC maps were produced by use of ADC values > or = 3 SD and > or = 10 SD above the mean value of normal WM. Lesions on diffusion-weighted images were classified as isointense or hypointense. ADC values within lesions (ADC(L)) were compared with those of normal WM (ADC(N)), and compared for isointense lesions and hypointense lesions. RESULTS: The distribution of lesions with ADC values > or = 3 SD was essentially identical to that on T2-weighted images. Regions with ADC values > or = 10 SD were found in both anterior WM and posterior WM in two patients and solely in posterior WM in nine patients. On diffusion-weighted images, lesions appeared isointense in seven patients and hypointense in four patients. Mean ADC(L)/ADC(N) for all lesions was 1.81; for hypointense lesions, 2.30. CONCLUSION: Vasogenic edema was more severe in posterior WM. Isointense lesions result from a balance of T2 effects and increased water diffusibility. Hypointense lesions have higher ADC values, which are not balanced by T2 effects.

Serial MR imaging of pineal cysts: implications for natural history and follow-up.

OBJECTIVE: The purpose of this study was to examine the frequency of change in size of pineal cysts on serial MR studies. MATERIALS AND METHODS: Thirty-two patients (19 females, 13 males) with a diagnosis of pineal cyst at any time who underwent brain MR imaging more than once in a period of at least 6 months were identified by computerized search of radiology reports. Four patients underwent MR imaging to follow up pineal cysts, whereas the remaining patients were imaged for a variety of indications, including intracerebral neoplasms. Measurements of maximal cyst dimension on both initial and latest follow-up studies were obtained in all patients, and cyst volumes were calculated in 23 patients. RESULTS: Length of follow-up ranged from 6 months to 9 years. All cysts were considered incidental and none were treated. Maximal cyst dimensions ranged from 0.5 to 2.2 cm. On average, there was no significant change in cyst volume. The maximal dimension of the cyst did not change in 24 (75%) of 32 patients. Two cysts resolved completely on follow-up, three cysts decreased by 2-4 mm, two cysts enlarged by 2-3 mm, and one cyst formed and grew to 1.2 cm. CONCLUSION: Whereas the size of pineal cysts as a whole remained unchanged on serial MR studies, cysts may either form or involute in individual patients. Small increases in cyst size did occur but were not associated with specific clinical findings. These findings suggest that typical pineal cysts may be followed up on a clinical basis alone rather than on imaging.

Vertex epidural hematomas: imaging findings and diagnostic pitfalls.

PURPOSE: Our purpose was to show the computed tomography (CT) and magnetic resonance (MR) imaging features of vertex epidural hematomas (EDHs) and emphasize pitfalls in the diagnosis of this entity. SUBJECTS AND METHODS: The neuroradiologic studies of four patients (CT in four, MR imaging and MR venography in one) were evaluated for EDH shape, size and appearance. RESULTS: EDHs were biconvex in three patients and crescentic in one patient. CT appearances included a collection that was hyperdense (two patients), generally isodense with a few regions of hyperdensity (one patient) and mixed hyperdense and hypodense (one patient). MR imaging findings in one patient consisted of hyperintense signal on T1-weighted images and hypointense signal on T2-weighted images. Inferior displacement of the superior sagittal sinus was seen in two patients. Diagnosis of a small vertex EDH was difficult on routine axial CT in one patient, but apparent on MR imaging and MR venography. CONCLUSIONS: Small vertex EDHs can be difficult to diagnose on routine CT. MR imaging or thin section CT should be performed to exclude the diagnosis in patients with trauma to the skull vertex.

Hippocampal sulcal cavities on MRI: relationship to age and apolipoprotein E genotype.

Hippocampal sulcal cavities are usually considered incidental findings on brain MRI. In a group of 92 elderly volunteers, the authors graded the number and size of hippocampal sulcal cavities with brain MRI to obtain a cavity score. Cavity scores increased with age, and were higher in subjects carrying the APOE epsilon4 or epsilon2 alleles than in subjects with the epsilon3/3 genotype. Age-related morphologic changes in the hippocampus may be mediated by the APOE genotype.

Comparison of patient age with MR imaging features of gangliogliomas.

OBJECTIVE: The purpose of this study was to compare MR imaging features of gangliogliomas in children less than 10 years old with those seen in patients at least 10 years old. MATERIALS AND METHODS: Our study population consisted of 15 female patients and 10 male patients with a mean age of 20 years. The early childhood group was composed of six children with a mean age of 5.5 years. The older group was composed of 19 patients with a mean age of 25.6 years. We assessed tumor volume, tumor location, percentage of tumor that was cystic, pattern of contrast enhancement, and degree of edema. RESULTS: The temporal lobe was the most common tumor location in both groups. Mean tumor volume in the early childhood group was 83 cm3, which was significantly larger than the mean tumor volume (9.78 cm3) for the older group (p = 0.001). Cystic tumors were more common in the early childhood group (83%) than in the older group (63%), and the average percentage of cysts in the cystic tumors was much higher in the early childhood group (67%) than in the older group (30%). Contrast enhancement was seen in five of six early childhood tumors and 13 of 16 tumors in older patients. Four of six tumors in the early childhood group and five of 19 tumors in the older patient group had associated edema. CONCLUSION: The mean tumor volume of gangliogliomas in the early childhood group was significantly larger than that of the older patient group. This finding may be indicative of differences in tumor growth patterns in the two groups, ability of the hemicranium to adjust to mass effect in childhood, or sampling error as a result of a relatively small sample size.

Magnetic resonance spectroscopy in Alzheimer's disease: focus on N-acetylaspartate.

This paper reviews published post-mortem brain and in-vivo proton magnetic resonance spectroscopy (1H-MRS) studies in Alzheimer's disease (AD) and focuses on the emerging role of N-acetylaspartate (NAA) as a prognostic marker of neuronal function. Post-mortem brain studies have reported significantly lower NAA levels in AD brains than in control brains, and some have correlated the low levels with neuropathological findings (i.e. amyloid plaques and neurofibrillary tangles). Similarly, almost all published in-vivo studies have reported lower NAA levels in AD patients compared to elderly controls. While some studies have found changes in metabolite levels that were considered useful for the diagnosis of AD, most have found that 1H-MRS provided little or no advantages over other, more common diagnostic tools. Instead, recent studies in AD and other neuropsychiatric disorders suggest that NAA may be more useful as a prognostic marker for monitoring neurodegeneration, stabilization, or improvement, and for evaluating therapeutic response to novel drugs.

White matter lesion contrast in fast spin-echo fluid-attenuated inversion recovery imaging: effect of varying effective echo time and echo train length.

OBJECTIVE: Our aim was to determine whether the contrast between white matter lesions and normal-appearing white matter in fast spin-echo fluid-attenuated inversion recovery (FLAIR) images can be improved by lengthening the effective TE and the echo train length. SUBJECTS AND METHODS: Thirty patients with various white matter lesions were imaged using fast spin-echo FLAIR sequences (TR = 10,002 msec; inversion time = 2200) on a 1.5-T MR imaging system. For 14 patients, fast spin-echo FLAIR sequences with a TE of 165 msec and echo train length of 32 (fast spin-echo FLAIR 165/32) were compared with fast spin-echo FLAIR sequences with a TE of 125 msec and echo train length of 24 (fast spin-echo FLAIR 125/24). For 16 other patients, fast spin-echo FLAIR 165/32 sequences were compared with fast spin-echo FLAIR sequences with a TE of 145 msec and echo train length of 28 (fast spin-echo FLAIR 145/28). Signal difference-to-noise ratios were calculated between the lesions and normal-appearing white matter for a typical lesion in each patient. RESULTS: In both groups, a small but statistically significant increase in the signal difference-to-noise ratio was found on the fast spin-echo FLAIR sequences using the longer TE and echo train length. In the first group, signal difference-to-noise ratio increased from 18.7 +/- 4.7 (mean +/- SD) for fast spin-echo FLAIR 125/24 to 20.1 +/- 4.5 for fast spin-echo FLAIR 165/32 (p < .05). In the second group, the signal difference-to-noise ratio increased from 15.4 +/- 4.0 for fast spin-echo FLAIR 145/28 to 16.8 +/- 4.6 for fast spin-echo FLAIR 165/32 (p <.01). In addition, fast spin-echo FLAIR sequences with a longer TE and echo train length were obtained more rapidly (6 min for fast spin-echo FLAIR 125/24, 5 min 20 sec for fast spin-echo FLAIR 145/28, and 4 min 41 sec for fast spin-echo FLAIR 165/32). CONCLUSION: Lengthening the TE to 165 msec and echo train length to 32 in fast spin-echo FLAIR imaging allows both a mild improvement in the contrast between white matter lesions and normal-appearing white matter and shorter imaging times.

Antiphospholipid antibodies: findings at arteriography.

PURPOSE: The purpose of this study was to determine the frequency and types of abnormalities at arteriography in patients with antiphospholipid antibodies (APA) and ischemic cerebrovascular events. METHODS: Twenty-three patients with APA and ischemic cerebrovascular events who underwent arteriography were identified. Patients over the age of 65 years were excluded. No patients met diagnostic criteria for systemic lupus erythematosus. All angiograms were reviewed by two neuroradiologists. RESULTS: Seventeen patients (74%) between the ages of 28 and 64 years (average age, 40 years) had abnormal angiograms. Sixteen patients had arterial abnormalities and one had dural sinus thrombosis. Ten had solely intracranial abnormalities (nine arterial and one venous), six had solely extracranial arterial abnormalities, and one had both intracranial and extracranial arterial abnormalities. Intracranial arterial abnormalities included stem or branch occlusions of the cerebral or basilar arteries, which were generally solitary (six patients), and findings suggestive of vasculitis (four patients). Four patients had stenoses of the origins of two or more great vessels. Two patients had extracranial internal carotid artery stenoses or occlusions that were not typical of atheromatous disease, considered to be embolic in one patient. In another patient, a stenosis of the origin of the internal carotid artery was present that appeared typical of atheromatous disease. Infarctions were seen on CT or MR studies in 13 of 17 patients with abnormal angiograms. CONCLUSION: In our group of patients, typical atheromatous lesions at the common carotid artery bifurcation were rare. Some lesions that are infrequent in the general stroke population (eg, vasculitis-like findings and stenoses at the origin of great vessels) were common. Patients with APA and cerebrovascular events appear to differ from the general stroke population with regard to types of arterial abnormalities seen at arteriography.