Scanner Bias Research Articles

Parkinson’s Disease Recognition Using Decorrelated Convolutional Neural Networks: Addressing Imbalance and Scanner Bias in rs-fMRI Data

Parkinson’s disease (PD) is a neurodegenerative and progressive disease that impacts the nerve cells in the brain and varies from person to person. The exact cause of PD is still unknown, and the diagnosis of PD does not include a specific objective test with certainty. Although deep learning has made great progress in medical neuroimaging analysis, these methods are very susceptible to biases present in neuroimaging datasets. An innovative decorrelated deep learning technique is introduced to mitigate class bias and scanner bias while simultaneously focusing on finding distinguishing characteristics in resting-state functional MRI (rs-fMRI) data, which assists in recognizing PD with good accuracy. The decorrelation function reduces the nonlinear correlation between features and bias in order to learn bias-invariant features. The publicly available Parkinson’s Progression Markers Initiative (PPMI) dataset, referred to as a single-scanner imbalanced dataset in this study, was used to validate our method. The imbalanced dataset problem affects the performance of the deep learning framework by overfitting to the majority class. To resolve this problem, we propose a new decorrelated convolutional neural network (DcCNN) framework by applying decorrelation-based optimization to convolutional neural networks (CNNs). An analysis of evaluation metrics comparisons shows that integrating the decorrelation function boosts the performance of PD recognition by removing class bias. Specifically, our DcCNN models perform significantly better than existing traditional approaches to tackle the imbalance problem. Finally, the same framework can be extended to create scanner-invariant features without significantly impacting the performance of a model. The obtained dataset is a multiscanner dataset, which leads to scanner bias due to the differences in acquisition protocols and scanners. The multiscanner dataset is a combination of two publicly available datasets, namely, PPMI and FTLDNI—the frontotemporal lobar degeneration neuroimaging initiative (NIFD) dataset. The results of t-distributed stochastic neighbor embedding (t-SNE) and scanner classification accuracy of our proposed feature extraction–DcCNN (FE-DcCNN) model validated the effective removal of scanner bias. Our method achieves an average accuracy of 77.80% on a multiscanner dataset for differentiating PD from a healthy control, which is superior to the DcCNN model trained on a single-scanner imbalanced dataset.

Association between increased Subcutaneous Adipose Tissue Radiodensity and cancer mortality: Automated computation, comparison of cancer types, gender, and scanner bias

PurposeBody composition analysis using computed tomography (CT) is proposed as a predictor of cancer mortality. An association between subcutaneous adipose tissue radiodensity (SATr) and cancer-specific mortality was established, while gender effects and equipment bias were estimated. Methods7,475 CT studies were selected from 17 cohorts containing CT images of untreated cancer patients who underwent follow-up for a period of 2.1–118.8 months. SATr measures were collected from published data (n = 6,718) or calculated according to CT images using a deep-learning network (n = 757). The association between SATr and mortality was ascertained for each cohort and gender using the p-value from either logistic regression or ROC analysis. The Kruskal-Wallis test was used to analyze differences between gender distributions, and automatic segmentation was evaluated using the Dice score and five-point Likert quality scale. Gender effect, scanner bias and changes in the Hounsfield unit (HU) to detect hazards were also estimated. ResultsHigher SATr was associated with mortality in eight cancer types (p < 0.05). Automatic segmentation produced a score of 0.949 while the quality scale measurement was good to excellent. The extent of gender effect was 5.2 HU while the scanner bias was 10.3 HU. The minimum proposed HU change to detect a patient at risk of death was between 5.6 and 8.3 HU. ConclusionsCT imaging provides valuable assessments of body composition as part of the staging process for several cancer types, saving both time and cost. Gender specific scales and scanner bias adjustments should be carried out to successfully implement SATr measures in clinical practice.

Quantitative performance of photon-counting CT at low dose: Virtual monochromatic imaging and iodine quantification.

Quantitative imaging techniques, such as virtual monochromatic imaging (VMI) and iodine quantification (IQ), have proven valuable diagnostic methods in several specific clinical tasks such as tumor and tissue differentiation. Recently, a new generation of computed tomography (CT) scanners equipped with photon-counting detectors (PCD) has reached clinical status. This work aimed to investigate the performance of a new photon-counting CT (PC-CT) in low-dose quantitative imaging tasks, comparing it to an earlier generation CT scanner with an energy-integrating detector dual-energy CT (DE-CT). The accuracy and precision of the quantification across size, dose, material types (including low and high iodine concentrations), displacement from iso-center, and solvent (tissue background) composition wereexplored. Quantitative analysis was performed on two clinical scanners, Siemens SOMATOM Force and NAEOTOM Alpha using a multi-energy phantom with plastic inserts mimicking different iodine concentrations and tissue types. The tube configurations in the dual-energy scanner were 80/150Sn kVp and 100/150Sn kVp, while for PC-CT both tube voltages were set to either 120 or 140 kVp with photon-counting energy thresholds set at 20/65 or 20/70 keV. The statistical significance of patient-related parameters in quantitative measurements was examined using ANOVA and pairwise comparison with the posthoc Tukey honest significance test. Scanner bias was assessed in both quantitative tasks for relevant patient-specificparameters. The accuracy of IQ and VMI in the PC-CT was comparable between standard and low radiation doses (p < 0.01). The patient size and tissue type significantly affect the accuracy of both quantitative imaging tasks in both scanners. The PC-CT scanner outperforms the DE-CT scanner in the IQ task in all cases. Iodine quantification bias in the PC-CT (-0.9 ± 0.15 mg/mL) at low doses in our study was comparable to that of DE-CT (range -2.6 to 1.5 mg/mL, published elsewhere) at a 1.7× higher dose, but the dose reduction severely biased DE-CT (4.72 ± 0.22 mg/mL). The accuracy in Hounsfield units (HU) estimation was comparable for 70 and 100 keV virtual imaging between scanners, but PC-CT was significantly underestimating virtual 40 keV HU values of dense materials in the phantom representing the extremely obesepopulation. The statistical analysis of our measurements reveals better IQ at lower radiation doses using new PC-CT. Although VMI performance was mostly comparable between the scanners, the DE-CT scanner quantitatively outperformed PC-CT when estimating HU values in the specific case of very large phantoms and dense materials, benefiting from increased X-ray tubepotentials.

Quality assessment, variability and reproducibility of anatomical measurements derived from T1-weighted brain imaging: The RIN–Neuroimaging Network case study

Initiatives for the collection of harmonized MRI datasets are growing continuously, opening questions on the reliability of results obtained in multi-site contexts.Here we present the assessment of the brain anatomical variability of MRI-derived measurements obtained from T1-weighted images, acquired according to the Standard Operating Procedures, promoted by the RIN-Neuroimaging Network. A multicentric dataset composed of 77 brain T1w acquisitions of young healthy volunteers (mean age = 29.7 ± 5.0 years), collected in 15 sites with MRI scanners of three different vendors, was considered. Parallelly, a dataset of 7 “traveling” subjects, each undergoing three acquisitions with scanners from different vendors, was also used. Intra-site, intra-vendor, and inter-site variabilities were evaluated in terms of the percentage standard deviation of volumetric and cortical thickness measures. Image quality metrics such as contrast-to-noise and signal-to-noise ratio in gray and white matter were also assessed for all sites and vendors.The results showed a measured global variability that ranges from 11% to 19% for subcortical volumes and from 3% to 10% for cortical thicknesses. Univariate distributions of the normalized volumes of subcortical regions, as well as the distributions of the thickness of cortical parcels appeared to be significantly different among sites in 8 subcortical (out of 17) and 21 cortical (out of 68) regions of i nterest in the multicentric study.The Bland-Altman analysis on “traveling” brain measurements did not detect systematic scanner biases even though a multivariate classification approach was able to classify the scanner vendor from brain measures with an accuracy of 0.60 ± 0.14 (chance level 0.33).

Sensitivity of portable low-field magnetic resonance imaging for multiple sclerosis lesions

Magnetic resonance imaging (MRI) is a fundamental tool in the diagnosis and management of neurological diseases such as multiple sclerosis (MS). New portable, low-field strength, MRI scanners could potentially lower financial and technical barriers to neuroimaging and reach underserved or disabled populations, but the sensitivity of these devices for MS lesions is unknown. We sought to determine if white matter lesions can be detected on a portable 64mT scanner, compare automated lesion segmentations and total lesion volume between paired 3T and 64mT scans, identify features that contribute to lesion detection accuracy, and explore super-resolution imaging at low-field. In this prospective, cross-sectional study, same-day brain MRI (FLAIR, T1w, and T2w) scans were collected from 36 adults (32 women; mean age, 50 ± 14 years) with known or suspected MS using Siemens 3T (FLAIR: 1 mm isotropic, T1w: 1 mm isotropic, and T2w: 0.34–0.5 × 0.34–0.5 × 3–5 mm) and Hyperfine 64mT (FLAIR: 1.6 × 1.6 × 5 mm, T1w: 1.5 × 1.5 × 5 mm, and T2w: 1.5 × 1.5 × 5 mm) scanners at two centers. Images were reviewed by neuroradiologists. MS lesions were measured manually and segmented using an automated algorithm. Statistical analyses assessed accuracy and variability of segmentations across scanners and systematic scanner biases in automated volumetric measurements. Lesions were identified on 64mT scans in 94% (31/33) of patients with confirmed MS. The average smallest lesions manually detected were 5.7 ± 1.3 mm in maximum diameter at 64mT vs 2.1 ± 0.6 mm at 3T, approaching the spatial resolution of the respective scanner sequences (3T: 1 mm, 64mT: 5 mm slice thickness). Automated lesion volume estimates were highly correlated between 3T and 64mT scans (r = 0.89, p < 0.001). Bland-Altman analysis identified bias in 64mT segmentations (mean = 1.6 ml, standard error = 5.2 ml, limits of agreement = –19.0–15.9 ml), which over-estimated low lesion volume and under-estimated high volume (r = 0.74, p < 0.001). Visual inspection revealed over-segmentation was driven venous hyperintensities on 64mT T2-FLAIR. Lesion size drove segmentation accuracy, with 93% of lesions > 1.0 ml and all lesions > 1.5 ml being detected. Using multi-acquisition volume averaging, we were able to generate 1.6 mm isotropic images on the 64mT device. Overall, our results demonstrate that in established MS, a portable 64mT MRI scanner can identify white matter lesions, and that automated estimates of total lesion volume correlate with measurements from 3T scans.

MASiVar: Multisite, multiscanner, and multisubject acquisitions for studying variability in diffusion weighted MRI.

Diffusion-weighted imaging allows investigators to identify structural, microstructural, and connectivity-based differences between subjects, but variability due to session and scanner biases is a challenge. To investigate DWI variability, we present MASiVar, a multisite data set consisting of 319 diffusion scans acquired at 3 T from b = 1000 to 3000 s/mm2 across 14 healthy adults, 83 healthy children (5 to 8 years), three sites, and four scanners as a publicly available, preprocessed, and de-identified data set. With the adult data, we demonstrate the capacity of MASiVar to simultaneously quantify the intrasession, intersession, interscanner, and intersubject variability of four common DWI processing approaches: (1) a tensor signal representation, (2) a multi-compartment neurite orientation dispersion and density model, (3) white-matter bundle segmentation, and (4) structural connectomics. Respectively, we evaluate region-wise fractional anisotropy, mean diffusivity, and principal eigenvector; region-wise CSF volume fraction, intracellular volume fraction, and orientation dispersion index; bundle-wise shape, volume, fractional anisotropy, and length; and whole connectome correlation and maximized modularity, global efficiency, and characteristic path length. We plot the variability in these measures at each level and find that it consistently increases with intrasession to intersession to interscanner to intersubject effects across all processing approaches and that sometimes interscanner variability can approach intersubject variability. This study demonstrates the potential of MASiVar to more globally investigate DWI variability across multiple levels and processing approaches simultaneously and suggests harmonization between scanners for multisite analyses should be considered before inference of group differences on subjects.

Quantitative imaging metrics derived from magnetic resonance fingerprinting using ISMRM/NIST MRI system phantom: An international multicenter repeatability and reproducibility study.

To compare the bias and inherent reliability of the quantitative (T1 and T2 ) imaging metrics generated from the magnetic resonance fingerprinting (MRF) technique using the ISMRM/NIST system phantom in an international multicenter setting. ISMRM/NIST MRI system phantom provides standard reference T1 and T2 relaxation values (vendor-provided) for each of the 14 vials in T1 and T2 arrays. MRF-SSFP scans repeated over 30days on GE 1.5 and 3.0T scanners at three collaborative centers. MRF estimated T1, and T2 values averaged over 30days were compared with the phantom vendor-provided and spin-echo (SE) based convention gold standard (GS) method. Repeatability and reproducibility were characterized by the within-case coefficient of variation (wCV) of the MRF data acquired over 30days, along with the biases. For the wide ranges of MRF estimated T1 values, vials #1-8 (T1 relaxation time between 2033 and 184ms) exhibited a wCV less than 3% and 4%, respectively, on 3.0 and 1.5T scanners. T2 values in vials #1-8 (T2 relaxation, 1044-45ms) have shown wCV to be <7% on both 3.0 and 1.5T scanners. A stronger linear correlation overall for T1 (R2 =0.9960 and 0.9963 at center-1 and center-2 on 3.0T scanner, and R2 =0.9951 and 0.9988 at center-1 and center-3 on 1.5T scanner) compared to T2 (R2 =0.9971 and 0.9972 at center-1 and center-2 on 3.0T scanner, and R2 =0.9815 and 0.9754 at center-1 and center-3 on 1.5T scanner). Bland-Altman (BA) analysis showed MRF based T1 and T2 values were within the limit of agreement (LOA) except for one data point. The mean difference or bias and 95% lower bound (LB) and upper bound (UB) LOA are reported in the format; mean bias: 95% LB LOA: 95% UB LOA. The biases for T1 values were 21.34: -50.00: 92.69, 21.32: -47.29: 89.94ms, and for T2 values were -19.88: -42.37: 2.61, -19.06: -43.58: 5.45ms on 3.0T scanner at center-1 and center-2, respectively. Similarly, on 1.5T scanner biases for T1 values were 26.54: -53.41: 106.50, 9.997: -51.94: 71.94ms, and for T2 values were -23.84: -135.40: 87.76, -37.30: 134.30: 59.73ms at center-1 and center-3, respectively. Additionally, the correlation between the SE based GS and MRF estimated T1 and T2 values (R2 =0.9969 and 0.9977) showed a similar trend as we observed between vendor-provided and MRF estimated T1 and T2 values (R2 =0.9963 and 0.9972). In addition to correlation, BA analysis showed that all the vials are within the LOA between the GS and vendor-provided for the T1 values and except one vial for T2 . All the vials are within the LOA between GS and MRF except one vial in T1 and T2 array. The wCV for reproducibility was <3% for both T1 and T2 values in vials #1-8, for all the 14 vials, wCV calculated for reproducibility was <4% for T1 values and <5% for T2 . This study shows that MRF is highly repeatable (wCV <4% for T1 and <7% for T2 ) and reproducible (wCV < 3% for both T1 and T2 ) in certain vials (vials #1-8).

Abstract PO-053: Machine learning CADx process for classification of lung nodules below the Lung-RADS 4A threshold in LDCT scans

Abstract Lung Cancer screening trials have demonstrated significant mortality reduction. Low-Dose Computed Tomography (LDCT) screening can frequently discover many small nodules in at risk participants. However classification of these, sub-cm nodules as cancerous or benign is a challenging task even for expert clinicians. In this work we use machine learning (ML) techniques to differentiate, cancerous (clinically confirmed) and benign nodules (&amp;gt;5 years of follow-up). Data for this study is drawn from a screening study (PanCan) from which we selected 613 distinct nodules (141 cancerous, and ~size matched 472 benign). We analyzed texture and shape features (~170) that are extracted from the nodule with and without perimeter transition pixels to control for perimeter effects. Features are also extracted from the ring of parenchyma surrounding the nodule to account for tumor effects on surrounding tissue. From the equivalent location in the opposite lung, parenchymal characteristics were extracted which we use to normalize the nodule texture features in an effort to reduce scanner bias. Our preliminary results for machine learning classification have shown model accuracies of up to ~80% (feature selection and classification algorithm dependent). Radiomic feature data can be combined with patient demographic variables such as age, sex, and smoking status to further improve our models, reaching ~84% classification accuracy. When normalized by the opposite lung, 58% of texture features showed improved classification ability. A single feature from the ring of parenchyma surrounding the nodule achieved 73.5% accuracy. Others have shown that nodule area/volume is a good classifier; for this data set it gave 68% accuracy. Data exploration to study classification accuracies among different subsets of patients reveals a number of interesting trends. We have noted for instance that in patients without emphysema, tumors can be classified with much higher accuracy (&amp;gt;90%) than in those who suffer from it. This and other results detailing classification among specific patient groupings may prove relevant in any potential future clinical implementation of these results. Citation Format: Rohan Abraham, Ian Janzen, Saeed Seyyedi, Sukhinder Khattra, John Mayo, Ren Yuan, Renelle Meyers, Stephen Lam, Calum MacAulay. Machine learning CADx process for classification of lung nodules below the Lung-RADS 4A threshold in LDCT scans [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-053.

Scanner invariant representations for diffusion MRI harmonization.

PurposeIn the present work, we describe the correction of diffusion‐weighted MRI for site and scanner biases using a novel method based on invariant representation.Theory and MethodsPooled imaging data from multiple sources are subject to variation between the sources. Correcting for these biases has become very important as imaging studies increase in size and multi‐site cases become more common. We propose learning an intermediate representation invariant to site/protocol variables, a technique adapted from information theory‐based algorithmic fairness; by leveraging the data processing inequality, such a representation can then be used to create an image reconstruction that is uninformative of its original source, yet still faithful to underlying structures. To implement this, we use a deep learning method based on variational auto‐encoders (VAE) to construct scanner invariant encodings of the imaging data.ResultsTo evaluate our method, we use training data from the 2018 MICCAI Computational Diffusion MRI (CDMRI) Challenge Harmonization dataset. Our proposed method shows improvements on independent test data relative to a recently published baseline method on each subtask, mapping data from three different scanning contexts to and from one separate target scanning context.ConclusionsAs imaging studies continue to grow, the use of pooled multi‐site imaging will similarly increase. Invariant representation presents a strong candidate for the harmonization of these data.

Measuring temporal stability of positron emission tomography standardized uptake value bias using long-lived sources in a multicenter network.

Positron emission tomography (PET) is a quantitative imaging modality, but the computation of standardized uptake values (SUVs) requires several instruments to be correctly calibrated. Variability in the calibration process may lead to unreliable quantitation. Sealed source kits containing traceable amounts of [Formula: see text] were used to measure signal stability for 19 PET scanners at nine hospitals in the National Cancer Institute's Quantitative Imaging Network. Repeated measurements of the sources were performed on PET scanners and in dose calibrators. The measured scanner and dose calibrator signal biases were used to compute the bias in SUVs at multiple time points for each site over a 14-month period. Estimation of absolute SUV accuracy was confounded by bias from the solid phantoms' physical properties. On average, the intrascanner coefficient of variation for SUV measurements was 3.5%. Over the entire length of the study, single-scanner SUV values varied over a range of 11%. Dose calibrator bias was not correlated with scanner bias. Calibration factors from the image metadata were nearly as variable as scanner signal, and were correlated with signal for many scanners. SUVs often showed low intrascanner variability between successive measurements but were also prone to shifts in apparent bias, possibly in part due to scanner recalibrations that are part of regular scanner quality control. Biases of key factors in the computation of SUVs were not correlated and their temporal variations did not cancel out of the computation. Long-lived sources and image metadata may provide a check on the recalibration process.

Combining Multiple Laser Scans of Spotted Microarrays by Means of a Two-Way ANOVA Model

Assessment of gene expression on spotted microarrays is based on measurement of fluorescence intensity emitted by hybridized spots. Unfortunately, quantifying fluorescence intensity from hybridized spots does not always correctly reflect gene expression level. Low gene expression levels produce low fluorescence intensities which tend to be confounded with the local background while high gene expression levels produce high fluorescence intensities which rapidly reach the saturation level. Most algorithms that combine data acquired at different voltages of the photomultiplier tube (PMT) assume that a change in scanner setting transforms the intensity measurements by a multiplicative constant. In this paper we introduce a new model of spot foreground intensity which integrates a PMT voltage independent scanner optical bias. This new model is used to implement a "Combining Multiple Scan using a Two-way ANOVA" (CMS2A) method, which is based on a maximum likelihood estimation of the scanner optical bias. After having computed scanner bias, coefficients of the two-way ANOVA model are used for correcting the saturated spots intensities obtained at high PMT voltage by using their counterpart values at lower PMT voltages. The method was compared to state-of-the-art multiple scan algorithms, using data generated from the MAQC study. CMS2A produced fold-changes that were highly correlated with qPCR fold-changes. As the scanner optical bias is accurately estimated within CMS2A, this method allows also avoiding fold-change compression biases whatever the value of this optical bias.

SU‐FF‐T‐316: 3 Years of Data Collection and Analysis of Film Flatness and Symmetry Using Vidar/ RIT Software From Different Centers Nation Wide

Purpose: To analyze accuracy and dependability of RIT software / Vidar Scanner in film flatness and symmetry measurement nationwide. Method and Materials: EDR and V films were exposed to 200 MU 6X &amp; 15X radiations at depth of 10cm in solid water phantom, 100cmSSD with filed size of 20×20cm. Measurements were repeated with different phantom orientation, phantom of different vendors, and different packs of films, different processor, and different linac machines. Water beam scanning was done with PTW water tank at the lowest speed, central axis and at steps of 3cm off axis. Profiles were analyzed using Memphysto software. Exposed films were scanned at different 90' intervals and flipped in each orientation for comparison. Measurements were repeated at different centers with different linac and different RIT software and Vidar Scanner Results: RIT's film flatness scanning shows an area of increased intensity on the left hand side of the scan regardless of using different Linac machine, Vidar scanner or the orientation of the phantom being used. No mirror images were obtained in RIT when the flatness film was scanned flipped or if rotated at different angles. Results of ‘Scan info’ test performed weekly on Vidar scanner for three months showed no failure and passed the overall performance and water beam scanning profile collected by PTW showed persistent uniform symmetry and flatness using TG‐45 guidelines. Conclusions: None of the mentioned effects were seen while scanning a patient's QA film. Only in scanning of flatness film, one would notice the above effects. RIT believes that a systemic scanner bias (within Vidar production specifications) may be causing this effect and is working on methods to compensate for any systemic scanner bias.

Replicability of diffusion tensor imaging measurements of fractional anisotropy and trace in brain

To evaluate within-scanner and between-scanner reliability of fractional anisotropy (FA) and trace (sum of the diagonal elements of the diffusion tensor) as measured by diffusion tensor imaging (DTI). Ten young healthy adults were scanned on three separate days, on two different systems made by the same manufacturer. One scan was acquired at one site, and two scans were acquired on two different occasions on another scanner at another site. Three levels of analysis were used to compare the DTI metrics: 1) a voxel-by-voxel analysis of all supratentorial brain (gray matter + white matter + cerebrospinal fluid) and of supratentorial white matter; 2) a slice-by-slice analysis of supratentorial white matter; and 3) a single-region analysis of the corpus callosum. The voxel-by-voxel analysis of all supratentorial brain found that FA and trace measures and correlations were equivalently and significantly higher within than across scanners. For supratentorial white matter, FA was similar within and across scanners, whereas trace demonstrated across-scanner bias. A similar pattern was observed for the slice-by-slice comparison. For the single-region analysis of the corpus callosum, within-scanner FA and trace measures were highly reproducible for FA (CV = 1.9%) and trace (CV = 2.6%), but both DTI measures showed a systematic mean bias across scanners (CV = 4.5% for FA and CV = 7.5% for trace). These estimates of measurement variation and scanner bias can be used to predict effect sizes for longitudinal and multisite studies using diffusion tensor imaging.