Lessons learned from independent external validation of an AI tool to detect breast cancer using a representative UK data set.

OBJECTIVE: To pilot a process for the independent external validation of an artificial intelligence (AI) tool to detect breast cancer using data from the NHS breast screening programme (NHSBSP). METHODS: A representative data set of mammography images from 26,000 women attending 2 NHS screening centres, and an enriched data set of 2054 positive cases were used from the OPTIMAM image database. The use case of the AI tool was the replacement of the first or second human reader. The performance of the AI tool was compared to that of human readers in the NHSBSP. RESULTS: Recommendations for future external validations of AI tools to detect breast cancer are provided. The tool recalled different breast cancers to the human readers. This study showed the importance of testing AI tools on all types of cases (including non-standard) and the clarity of any warning messages. The acceptable difference in sensitivity and specificity between the AI tool and human readers should be determined. Any information vital for the clinical application should be a required output for the AI tool. It is recommended that the interaction of radiologists with the AI tool, and the effect of the AI tool on arbitration be investigated prior to clinical use. CONCLUSION: This pilot demonstrated several lessons for future independent external validation of AI tools for breast cancer detection. ADVANCES IN KNOWLEDGE: Knowledge has been gained towards best practice procedures for performing independent external validations of AI tools for the detection of breast cancer using data from the NHS Breast Screening Programme.

Automated measurement of hippocampal atrophy using fluid-registered serial MRI in AD and controls.

OBJECTIVE: To assess hippocampal atrophy rates calculated from fluid registration methods. METHODS: Hippocampi were segmented on baseline and registered-repeat scans of 32 probable Alzheimer disease (AD) subjects and 55 controls. Fluid-based atrophy rates were calculated. RESULTS: In AD patients, the mean (SD) atrophy rates for manual, fluidly propagated, and Jacobian methods were 5.09 (3.59), 5.34 (3.43), and 3.55 (2.70) (percentage per year). In controls, atrophy rates were 1.31 (2.00), 0.89 (0.75), and 0.56 (1.12) (percentage per year). In AD, fluid propagation and manual rates were similar in means (P = 0.55) and variances (P = 0.71). Jacobian rates were smaller in mean (P = 0.002) and variance (P = 0.026) than in manual rates. In controls, fluid-propagated rates were similar in mean to manual rates (P = 0.12), but less variable (P < 0.0001). Jacobian rates were smaller in mean (P = 0.014) and less variable (P < 0.0001) than in manual rates. Both fluid methods were superior to manual measures in separating AD from controls (P < 0.0001). CONCLUSIONS: Fluid-based methods may be useful in large serial hippocampal studies.

Quantitative magnetization transfer imaging in Alzheimer disease.

PURPOSE: To prospectively measure magnetization transfer (MT) parameters, along with established atrophy parameters, in patients with Alzheimer disease (AD) and in age- and sex-matched control subjects. MATERIALS AND METHODS: Participants provided informed consent, and additional assent was obtained from next of kin of all patients with AD. The study was approved by the local ethics committee. Fourteen patients with AD (seven men; mean age, 67.2 years+/-6.5 [standard deviation]) and 14 control subjects (nine men; mean age, 65.5 years+/-9.4) underwent volumetric T1-weighted magnetic resonance and MT imaging. Whole-brain and total hippocampal volumes were adjusted for total intracranial volume. MT images were processed to derive four fundamental parameters in the hippocampal region by using the two-pool model of the MT phenomenon. Pearson correlation coefficients were used to assess the association between volumetric and MT parameters and Mini-Mental State Examination (MMSE) results. Logistic regression models were used to investigate whether combinations of parameters associated with MMSE could help provide better group discrimination. RESULTS: Patients with AD had significantly reduced whole-brain (P=.001) and total hippocampal (P<.001) volumes compared with those of control subjects. Two MT parameters were significantly reduced in the hippocampal region of patients: 1/(RAT2A)--that is, ratio of relaxation times of free proton pool, where RA equals 1/T1A and is the inverse of the longitudinal relaxation time of the free proton pool (P=.01)--and f*b, which equals fb/[RA(1-fb)], where fb is the restricted proton fraction (P<.001). Among patients with AD, whole-brain volume and hippocampal were correlated with MMSE results. When both parameters were included in a logistic regression model, only hippocampal was significantly associated with case-control status (P=.03). CONCLUSION: Certain MT parameters may serve as useful biomarkers of AD.

Differentiating AD from aging using semiautomated measurement of hippocampal atrophy rates.

Manual segmentation of the hippocampus is the gold standard in volumetric hippocampal magnetic resonance imaging (MRI) analysis; however, this is difficult to achieve reproducibly. This study explores whether application of local registration and calculation of the hippocampal boundary shift integral (HBSI) can reduce random variation compared with manual measures. Hippocampi were outlined on the baseline and registered-repeat MRIs of 32 clinically diagnosed Alzheimer's disease (AD) patients and 47 matched controls (37-86 years) with a wide range of scanning intervals (175-1173 days). The scans were globally registered using 9 degrees of freedom and subsequently locally registered using 6 degrees of freedom and HBSI was then calculated automatically. HBSI significantly reduced the mean rate (P < 0.01) and variation in controls (P < 0.001) and increased group separation between AD cases and controls. When comparing HBSI atrophy rates with manually derived atrophy rates at 90% sensitivity, specificities were 98% and 81%, respectively. From logistic regression models, a 1% increase in HBSI atrophy rates was associated with an 11-fold (CI 3, 36) increase in the odds of a diagnosis of AD. For manually derived atrophy rates, the equivalent odds ratio was 3 (CI 2,4). We conclude that HBSI-derived atrophy rates reduce operator time and error, and are at least as effective as the manual equivalent as a diagnostic marker and are a potential marker of progression in longitudinal studies and trials.

Using nine degrees-of-freedom registration to correct for changes in voxel size in serial MRI studies.

Quantitative longitudinal brain magnetic resonance (MR) studies may be confounded by scanner-related drifts in voxel sizes. Total intracranial volume (TIV) normalisation is commonly used to correct serial cerebral volumetric measurements for these drifts. We hypothesised that automated rigid-body registration of whole brain incorporating automatic scaling correction might also correct for such fluctuations, and might be a more practical alternative. Twenty-three subjects (12 patients with Alzheimer's disease [AD] and 11 controls) had at least two serial T1-weighted volumetric brain MR scans. Ten scans from the control subjects were artificially scaled (stretched) by 1.5, 3.0, 4.6 and 6.1%. A 9-degrees-of-freedom (9dof) registration was used to register the scaled scans back onto the original scans and corresponding scaling factors compared to TIV measurements. A further nine 1-year repeat scans from the AD subjects were artificially scaled and registered (9dof) to baseline. The two correction methods were further assessed using multiple serial scans for each of the 23 subjects (resulting in 49 scan pairs). All serial scans were registered (9dof) to baseline. TIV was measured on all scans. It was found that the 9dof registration successfully recovered the artificially generated scaling changes. Scaling correction using 9dof registration did not alter the amount of brain atrophy measured over the 1-year period in the AD subjects. The 9dof volume scaling factors were very similar to the TIV ratios (repeat TIV over baseline TIV), but less variable (p < 0.001), in both artificial and 'real' scenarios. In the latter, the volume scaling factors allowed identification of two time-points in which a 3% change in voxel size had occurred. Both the 9dof brain registration and TIV correction were successfully able to correct for these fluctuations. Significant shifts in voxel size are a problem in longitudinal brain imaging studies. It is important that such changes are adjusted for: 9dof registration, which is automated and computationally inexpensive, may be superior to the more labour-intensive TIV correction for this purpose.

Correction of differential intensity inhomogeneity in longitudinal MR images.

Longitudinal MR imaging is increasingly being used to measure cerebral atrophy progression in dementia and other neurological disorders. Differences in intensity inhomogeneity between serial scans can confound these measurements. This differential bias also distorts nonlinear registration and makes both manual and automated segmentation of tissue type less reliable. A technique is described for the correction of this differential bias that makes no assumptions about signal distribution, bias field or signal homogeneity. Instead, the bias field calculation is performed on the basis that the remaining structure in the difference image of registered serial scans has small-scale structure. The differential bias field is of much larger scale and can thus be obtained by applying an appropriate filter to the difference image. The serial scan pair is then corrected for the differential bias field and atrophy measurement can be performed on the corrected scan pair. Application of a known, simulated bias field to real serial MR images was shown to alter atrophy measurements significantly. The differential correction method recovered the applied differential bias field and thereby improved atrophy measurements. This method was then applied to serial imaging in patients with dementia using a set of serial scan pairs with visually identified, significant differential bias and a set of scan pairs with negligible differential bias. Differential bias correction specifically reduced the variance of the atrophy measure significantly for the scans with significant differential bias.