A Semiautonomous Deep Learning System to Reduce False Positives in Screening Mammography.

Purpose To evaluate the ability of a semiautonomous artificial intelligence (AI) model to identify screening mammograms not suspicious for breast cancer and reduce the number of false-positive examinations. Materials and Methods The deep learning algorithm was trained using 123 248 two-dimensional digital mammograms (6161 cancers) and a retrospective study was performed on three nonoverlapping datasets of 14 831 screening mammography examinations (1026 cancers) from two U.S. institutions and one U.K. institution (2008-2017). The stand-alone performance of humans and AI was compared. Human plus AI performance was simulated to examine reductions in the cancer detection rate, number of examinations, false-positive callbacks, and benign biopsies. Metrics were adjusted to mimic the natural distribution of a screening population, and bootstrapped CIs and P values were calculated. Results Retrospective evaluation on all datasets showed minimal changes to the cancer detection rate with use of the AI device (noninferiority margin of 0.25 cancers per 1000 examinations: U.S. dataset 1, P = .02; U.S. dataset 2, P < .001; U.K. dataset, P < .001). On U.S. dataset 1 (11 592 mammograms; 101 cancers; 3810 female patients; mean age, 57.3 years ± 10.0 [SD]), the device reduced screening examinations requiring radiologist interpretation by 41.6% (95% CI: 40.6%, 42.4%; P < .001), diagnostic examinations callbacks by 31.1% (95% CI: 28.7%, 33.4%; P < .001), and benign needle biopsies by 7.4% (95% CI: 4.1%, 12.4%; P < .001). U.S. dataset 2 (1362 mammograms; 330 cancers; 1293 female patients; mean age, 55.4 years ± 10.5) was reduced by 19.5% (95% CI: 16.9%, 22.1%; P < .001), 11.9% (95% CI: 8.6%, 15.7%; P < .001), and 6.5% (95% CI: 0.0%, 19.0%; P = .08), respectively. The U.K. dataset (1877 mammograms; 595 cancers; 1491 female patients; mean age, 63.5 years ± 7.1) was reduced by 36.8% (95% CI: 34.4%, 39.7%; P < .001), 17.1% (95% CI: 5.9%, 30.1%: P < .001), and 5.9% (95% CI: 2.9%, 11.5%; P < .001), respectively. Conclusion This work demonstrates the potential of a semiautonomous breast cancer screening system to reduce false positives, unnecessary procedures, patient anxiety, and medical expenses. Keywords: Artificial Intelligence, Semiautonomous Deep Learning, Breast Cancer, Screening Mammography Supplemental material is available for this article. Published under a CC BY 4.0 license.

Development and clinical application of a deep learning model to identify acute infarct on magnetic resonance imaging.

Stroke is a leading cause of death and disability. The ability to quickly identify the presence of acute infarct and quantify the volume on magnetic resonance imaging (MRI) has important treatment implications. We developed a machine learning model that used the apparent diffusion coefficient and diffusion weighted imaging series. It was trained on 6,657 MRI studies from Massachusetts General Hospital (MGH; Boston, USA). All studies were labelled positive or negative for infarct (classification annotation) with 377 having the region of interest outlined (segmentation annotation). The different annotation types facilitated training on more studies while not requiring the extensive time to manually segment every study. We initially validated the model on studies sequestered from the training set. We then tested the model on studies from three clinical scenarios: consecutive stroke team activations for 6-months at MGH, consecutive stroke team activations for 6-months at a hospital that did not provide training data (Brigham and Women's Hospital [BWH]; Boston, USA), and an international site (Diagnósticos da América SA [DASA]; Brazil). The model results were compared to radiologist ground truth interpretations. The model performed better when trained on classification and segmentation annotations (area under the receiver operating curve [AUROC] 0.995 [95% CI 0.992-0.998] and median Dice coefficient for segmentation overlap of 0.797 [IQR 0.642-0.861]) compared to segmentation annotations alone (AUROC 0.982 [95% CI 0.972-0.990] and Dice coefficient 0.776 [IQR 0.584-0.857]). The model accurately identified infarcts for MGH stroke team activations (AUROC 0.964 [95% CI 0.943-0.982], 381 studies), BWH stroke team activations (AUROC 0.981 [95% CI 0.966-0.993], 247 studies), and at DASA (AUROC 0.998 [95% CI 0.993-1.000], 171 studies). The model accurately segmented infarcts with Pearson correlation comparing model output and ground truth volumes between 0.968 and 0.986 for the three scenarios. Acute infarct can be accurately detected and segmented on MRI in real-world clinical scenarios using a machine learning model.

A Multisite Study of a Breast Density Deep Learning Model for Full-Field Digital Mammography and Synthetic Mammography.

PURPOSE: To develop a Breast Imaging Reporting and Data System (BI-RADS) breast density deep learning (DL) model in a multisite setting for synthetic two-dimensional mammographic (SM) images derived from digital breast tomosynthesis examinations by using full-field digital mammographic (FFDM) images and limited SM data. MATERIALS AND METHODS: A DL model was trained to predict BI-RADS breast density by using FFDM images acquired from 2008 to 2017 (site 1: 57 492 patients, 187 627 examinations, 750 752 images) for this retrospective study. The FFDM model was evaluated by using SM datasets from two institutions (site 1: 3842 patients, 3866 examinations, 14 472 images, acquired from 2016 to 2017; site 2: 7557 patients, 16 283 examinations, 63 973 images, 2015 to 2019). Each of the three datasets were then split into training, validation, and test. Adaptation methods were investigated to improve performance on the SM datasets, and the effect of dataset size on each adaptation method was considered. Statistical significance was assessed by using CIs, which were estimated by bootstrapping. RESULTS: Without adaptation, the model demonstrated substantial agreement with the original reporting radiologists for all three datasets (site 1 FFDM: linearly weighted Cohen κ [κw] = 0.75 [95% CI: 0.74, 0.76]; site 1 SM: κw = 0.71 [95% CI: 0.64, 0.78]; site 2 SM: κw = 0.72 [95% CI: 0.70, 0.75]). With adaptation, performance improved for site 2 (site 1: κw = 0.72 [95% CI: 0.66, 0.79], 0.71 vs 0.72, P = .80; site 2: κw = 0.79 [95% CI: 0.76, 0.81], 0.72 vs 0.79, P < .001) by using only 500 SM images from that site. CONCLUSION: A BI-RADS breast density DL model demonstrated strong performance on FFDM and SM images from two institutions without training on SM images and improved by using few SM images.Supplemental material is available for this article.Published under a CC BY 4.0 license.

Probabilistic Graphical Models for Dynamic PET: A Novel Approach to Direct Parametric Map Estimation and Image Reconstruction.

In the context of dynamic emission tomography, the conventional processing pipeline consists of independent image reconstruction of single-time frames, followed by the application of a suitable kinetic model to time-activity curves (TACs) at the voxel or region-of-interest level. Direct 4D positron emission tomography (PET) reconstruction, by contrast, seeks to move beyond this scheme and incorporate information from multiple time frames within the reconstruction task. Established direct methods are based on a deterministic description of voxelwise TACs, captured by the chosen kinetic model, considering the photon counting process the only source of uncertainty. In this paper, we introduce a new probabilistic modeling strategy based on the key assumption that activity time course would be subject to uncertainty even if the parameters of the underlying dynamic process are known. This leads to a hierarchical model that we formulate using the formalism of probabilistic graphical modeling. The inference is addressed using a new iterative algorithm, in which kinetic modeling results are treated as prior expectation of activity time course, rather than as a deterministic match, making it possible to control the trade-off between a data-driven and a model-driven reconstruction. The proposed method is flexible to an arbitrary choice of (linear and nonlinear) kinetic models, it enables the inclusion of arbitrary (sub)differentiable priors for parametric maps, and it is simple to implement. Computer simulations and an application to a real-patient scan show how the proposed method is able to generalize over conventional indirect and direct approaches, providing a bridge between them by properly tuning the impact of the kinetic modeling step on image reconstruction.

Comparison of the clinical performance of upper abdominal PET/DCE-MRI with and without concurrent respiratory motion correction (MoCo).

PURPOSE: To compare the clinical performance of upper abdominal PET/DCE-MRI with and without concurrent respiratory motion correction (MoCo). METHODS: MoCo PET/DCE-MRI of the upper abdomen was acquired in 44 consecutive oncologic patients and compared with non-MoCo PET/MRI. SUVmax and MTV of FDG-avid upper abdominal malignant lesions were assessed on MoCo and non-MoCo PET images. Image quality was compared between MoCo DCE-MRI and non-MoCo CE-MRI, and between fused MoCo PET/MRI and fused non-MoCo PET/MRI images. RESULTS: MoCo PET resulted in higher SUVmax (10.8 ± 5.45) than non-MoCo PET (9.62 ± 5.42) and lower MTV (35.55 ± 141.95 cm3) than non-MoCo PET (38.11 ± 198.14 cm3; p < 0.005 for both). The quality of MoCo DCE-MRI images (4.73 ± 0.5) was higher than that of non-MoCo CE-MRI images (4.53±0.71; p = 0.037). The quality of fused MoCo-PET/MRI images (4.96 ± 0.16) was higher than that of fused non-MoCo PET/MRI images (4.39 ± 0.66; p < 0.005). CONCLUSION: MoCo PET/MRI provided qualitatively better images than non-MoCo PET/MRI, and upper abdominal malignant lesions demonstrated higher SUVmax and lower MTV on MoCo PET/MRI.

Concurrent Respiratory Motion Correction of Abdominal PET and Dynamic Contrast-Enhanced-MRI Using a Compressed Sensing Approach.

We present an approach for concurrent reconstruction of respiratory motion-compensated abdominal dynamic contrast-enhanced (DCE)-MRI and PET data in an integrated PET/MR scanner. The MR and PET reconstructions share the same motion vector fields derived from radial MR data; the approach is robust to changes in respiratory pattern and does not increase the total acquisition time. Methods: PET and DCE-MRI data of 12 oncologic patients were simultaneously acquired for 6 min on an integrated PET/MR system after administration of 18F-FDG and gadoterate meglumine. Golden-angle radial MR data were continuously acquired simultaneously with PET data and sorted into multiple motion phases on the basis of a respiratory signal derived directly from the radial MR data. The resulting multidimensional dataset was reconstructed using a compressed sensing approach that exploits sparsity among respiratory phases. Motion vector fields obtained using the full 6-min (MC6-min) and only the last 1 min (MC1-min) of data were incorporated into the PET reconstruction to obtain motion-corrected PET images and in an MR iterative reconstruction algorithm to produce a series of motion-corrected DCE-MR images (moco_GRASP). The motion-correction methods (MC6-min and MC1-min) were evaluated by qualitative analysis of the MR images and quantitative analysis of SUVmax and SUVmean, contrast, signal-to-noise ratio (SNR), and lesion volume in the PET images. Results: Motion-corrected MC6-min PET images demonstrated 30%, 23%, 34%, and 18% increases in average SUVmax, SUVmean, contrast, and SNR and an average 40% reduction in lesion volume with respect to the non-motion-corrected PET images. The changes in these figures of merit were smaller but still substantial for the MC1-min protocol: 19%, 10%, 15%, and 9% increases in average SUVmax, SUVmean, contrast, and SNR; and a 28% reduction in lesion volume. Moco_GRASP images were deemed of acceptable or better diagnostic image quality with respect to conventional breath-hold Cartesian volumetric interpolated breath-hold examination acquisitions. Conclusion: We presented a method that allows the simultaneous acquisition of respiratory motion-corrected diagnostic quality DCE-MRI and quantitatively accurate PET data in an integrated PET/MR scanner with negligible prolongation in acquisition time compared with routine PET/DCE-MRI protocols.

A machine learning method for fast and accurate characterization of depth-of-interaction gamma cameras.

Measuring the depth-of-interaction (DOI) of gamma photons enables increasing the resolution of emission imaging systems. Several design variants of DOI-sensitive detectors have been recently introduced to improve the performance of scanners for positron emission tomography (PET). However, the accurate characterization of the response of DOI detectors, necessary to accurately measure the DOI, remains an unsolved problem. Numerical simulations are, at the state of the art, imprecise, while measuring directly the characteristics of DOI detectors experimentally is hindered by the impossibility to impose the depth-of-interaction in an experimental set-up. In this article we introduce a machine learning approach for extracting accurate forward models of gamma imaging devices from simple pencil-beam measurements, using a nonlinear dimensionality reduction technique in combination with a finite mixture model. The method is purely data-driven, not requiring simulations, and is applicable to a wide range of detector types. The proposed method was evaluated both in a simulation study and with data acquired using a monolithic gamma camera designed for PET (the cMiCE detector), demonstrating the accurate recovery of the DOI characteristics. The combination of the proposed calibration technique with maximum- a posteriori estimation of the coordinates of interaction provided a depth resolution of ≈1.14 mm for the simulated PET detector and ≈1.74 mm for the cMiCE detector. The software and experimental data are made available at http://occiput.mgh.harvard.edu/depthembedding/.

PET/MRI in the Presence of Metal Implants: Completion of the Attenuation Map from PET Emission Data.

We present a novel technique for accurate whole-body attenuation correction in the presence of metallic endoprosthesis, on integrated non-time-of-flight (non-TOF) PET/MRI scanners. The proposed implant PET-based attenuation map completion (IPAC) method performs a joint reconstruction of radioactivity and attenuation from the emission data to determine the position, shape, and linear attenuation coefficient (LAC) of metallic implants. Methods: The initial estimate of the attenuation map was obtained using the MR Dixon method currently available on the Siemens Biograph mMR scanner. The attenuation coefficients in the area of the MR image subjected to metal susceptibility artifacts are then reconstructed from the PET emission data using the IPAC algorithm. The method was tested on 11 subjects presenting 13 different metallic implants, who underwent CT and PET/MR scans. Relative mean LACs and Dice similarity coefficients were calculated to determine the accuracy of the reconstructed attenuation values and the shape of the metal implant, respectively. The reconstructed PET images were compared with those obtained using the reference CT-based approach and the Dixon-based method. Absolute relative change (aRC) images were generated in each case, and voxel-based analyses were performed. Results: The error in implant LAC estimation, using the proposed IPAC algorithm, was 15.7% ± 7.8%, which was significantly smaller than the Dixon- (100%) and CT- (39%) derived values. A mean Dice similarity coefficient of 73% ± 9% was obtained when comparing the IPAC- with the CT-derived implant shape. The voxel-based analysis of the reconstructed PET images revealed quantification errors (aRC) of 13.2% ± 22.1% for the IPAC- with respect to CT-corrected images. The Dixon-based method performed substantially worse, with a mean aRC of 23.1% ± 38.4%. Conclusion: We have presented a non-TOF emission-based approach for estimating the attenuation map in the presence of metallic implants, to be used for whole-body attenuation correction in integrated PET/MR scanners. The Graphics Processing Unit implementation of the algorithm will be included in the open-source reconstruction toolbox Occiput.io.

Attenuation correction synthesis for hybrid PET-MR scanners: application to brain studies.

Attenuation correction is an essential requirement for quantification of positron emission tomography (PET) data. In PET/CT acquisition systems, attenuation maps are derived from computed tomography (CT) images. However, in hybrid PET/MR scanners, magnetic resonance imaging (MRI) images do not directly provide a patient-specific attenuation map. The aim of the proposed work is to improve attenuation correction for PET/MR scanners by generating synthetic CTs and attenuation maps. The synthetic images are generated through a multi-atlas information propagation scheme, locally matching the MRI-derived patient's morphology to a database of MRI/CT pairs, using a local image similarity measure. Results show significant improvements in CT synthesis and PET reconstruction accuracy when compared to a segmentation method using an ultrashort-echo-time MRI sequence and to a simplified atlas-based method.

Efficient determination of the uncertainty for the optimization of SPECT system design: a subsampled fisher information matrix.

System designs in single photon emission tomography (SPECT) can be evaluated based on the fundamental trade-off between bias and variance that can be achieved in the reconstruction of emission tomograms. This trade off can be derived analytically using the Cramer-Rao type bounds, which imply the calculation and the inversion of the Fisher information matrix (FIM). The inverse of the FIM expresses the uncertainty associated to the tomogram, enabling the comparison of system designs. However, computing, storing and inverting the FIM is not practical with 3-D imaging systems. In order to tackle the problem of the computational load in calculating the inverse of the FIM, a method based on the calculation of the local impulse response and the variance, in a single point, from a single row of the FIM, has been previously proposed for system design. However this approximation (circulant approximation) does not capture the global interdependence between the variables in shift-variant systems such as SPECT, and cannot account e.g., for data truncation or missing data. Our new formulation relies on subsampling the FIM. The FIM is calculated over a subset of voxels arranged in a grid that covers the whole volume. Every element of the FIM at the grid points is calculated exactly, accounting for the acquisition geometry and for the object. This new formulation reduces the computational complexity in estimating the uncertainty, but nevertheless accounts for the global interdependence between the variables, enabling the exploration of design spaces hindered by the circulant approximation. The graphics processing unit accelerated implementation of the algorithm reduces further the computation times, making the algorithm a good candidate for real-time optimization of adaptive imaging systems. This paper describes the subsampled FIM formulation and implementation details. The advantages and limitations of the new approximation are explored, in comparison with the circulant approximation, in the context of design optimization of a parallel-hole collimator SPECT system and of an adaptive imaging system (similar to the commercially available D-SPECT).

Attenuation correction synthesis for hybrid PET-MR scanners.

The combination of functional and anatomical imaging technologies such as Positron Emission Tomography (PET) and Computed Tomography (CT) has shown its value in the preclinical and clinical fields. In PET/CT hybrid acquisition systems, CT-derived attenuation maps enable a more accurate PET reconstruction. However, CT provides only very limited soft-tissue contrast and exposes the patient to an additional radiation dose. In comparison, Magnetic Resonance Imaging (MRI) provides good soft-tissue contrast and the ability to study functional activation and tissue microstructures, but does not directly provide patient-specific electron density maps for PET reconstruction. The aim of the proposed work is to improve PET/MR reconstruction by generating synthetic CTs and attenuation-maps. The synthetic images are generated through a multi-atlas information propagation scheme, locally matching the MRI-derived patient's morphology to a database of pre-acquired MRI/CT pairs. Results show improvements in CT synthesis and PET reconstruction accuracy when compared to a segmentation method using an Ultrashort-Echo-Time MRI sequence.

Markov random field and Gaussian mixture for segmented MRI-based partial volume correction in PET.

In this paper we propose a segmented magnetic resonance imaging (MRI) prior-based maximum penalized likelihood deconvolution technique for positron emission tomography (PET) images. The model assumes the existence of activity classes that behave like a hidden Markov random field (MRF) driven by the segmented MRI. We utilize a mean field approximation to compute the likelihood of the MRF. We tested our method on both simulated and clinical data (brain PET) and compared our results with PET images corrected with the re-blurred Van Cittert (VC) algorithm, the simplified Guven (SG) algorithm and the region-based voxel-wise (RBV) technique. We demonstrated our algorithm outperforms the VC algorithm and outperforms SG and RBV corrections when the segmented MRI is inconsistent (e.g. mis-segmentation, lesions, etc) with the PET image.

An anatomically driven anisotropic diffusion filtering method for 3D SPECT reconstruction.

In this study, we aim to reconstruct single-photon emission computed tomography images using anatomical information from magnetic resonance imaging as a priori knowledge about the activity distribution. The trade-off between anatomical and emission data is one of the main concerns for such studies. In this work, we propose an anatomically driven anisotropic diffusion filter (ADADF) as a penalized maximum likelihood expectation maximization optimization framework. The ADADF method has improved edge-preserving denoising characteristics compared to other smoothing penalty terms based on quadratic and non-quadratic functions. The proposed method has an important ability to retain information which is absent in the anatomy. To make our approach more stable to the noise-edge classification problem, robust statistics have been employed. Comparison of the ADADF method is performed with a successful anatomically driven technique, namely, the Bowsher prior (BP). Quantitative assessment using simulated and clinical neuroreceptor volumetric data show the advantage of the ADADF over the BP. For the modelled data, the overall image resolution, the contrast, the signal-to-noise ratio and the ability to preserve important features in the data are all improved by using the proposed method. For clinical data, the contrast in the region of interest is significantly improved using the ADADF compared to the BP, while successfully eliminating noise.

Steady-state model of the radio-pharmaceutical uptake for MR-PET.

This work explores a fully-automated algorithm for estimation of the uptake of radio-pharmaceutical in brain MR-PET imaging. The algorithm is based on a model of the pharmaceutical uptake coupled with probabilistic models of the PET and MR acquisition systems. In contrast to algorithms that attempt to correct for the partial volume effect (PVE), the problem is tackled here in the reconstruction by means of a probabilistic model of the pharmaceutical uptake. We make use of hybrid Bayesian networks to describe the joint probabilistic model and to obtain an efficient optimisation algorithm. We describe solutions adopted in order to mitigate the effect of local maxima and to reduce the sensitivity to the initialisation of the parameters, rendering the algorithm fully automatic. The algorithm is evaluated on simulated MR-PET data and on the reconstruction of clinical PET FDG acquisitions.

4-D generative model for PET/MRI reconstruction.

We introduce a 4-dimensional joint generative probabilistic model for estimation of activity in a PET/MRI imaging system. The model is based on a mixture of Gaussians, relating time dependent activity and MRI image intensity to a hidden static variable, allowing one to estimate jointly activity, the parameters that capture the interdependence of the two images and motion parameters. An iterative algorithm for optimisation of the model is described. Noisy simulation data, modeling 3-D patient head movements, is obtained with realistic PET and MRI simulators and with a brain phantom from the BrainWeb database. Joint estimation of activity and motion parameters within the same framework allows us to use information from the MRI images to improve the activity estimate in terms of noise and recovery.