Pneumothorax detection in chest radiographs: optimizing artificial intelligence system for accuracy and confounding bias reduction using in-image annotations in algorithm training.

OBJECTIVES: Diagnostic accuracy of artificial intelligence (AI) pneumothorax (PTX) detection in chest radiographs (CXR) is limited by the noisy annotation quality of public training data and confounding thoracic tubes (TT). We hypothesize that in-image annotations of the dehiscent visceral pleura for algorithm training boosts algorithm's performance and suppresses confounders. METHODS: Our single-center evaluation cohort of 3062 supine CXRs includes 760 PTX-positive cases with radiological annotations of PTX size and inserted TTs. Three step-by-step improved algorithms (differing in algorithm architecture, training data from public datasets/clinical sites, and in-image annotations included in algorithm training) were characterized by area under the receiver operating characteristics (AUROC) in detailed subgroup analyses and referenced to the well-established "CheXNet" algorithm. RESULTS: Performances of established algorithms exclusively trained on publicly available data without in-image annotations are limited to AUROCs of 0.778 and strongly biased towards TTs that can completely eliminate algorithm's discriminative power in individual subgroups. Contrarily, our final "algorithm 2" which was trained on a lower number of images but additionally with in-image annotations of the dehiscent pleura achieved an overall AUROC of 0.877 for unilateral PTX detection with a significantly reduced TT-related confounding bias. CONCLUSIONS: We demonstrated strong limitations of an established PTX-detecting AI algorithm that can be significantly reduced by designing an AI system capable of learning to both classify and localize PTX. Our results are aimed at drawing attention to the necessity of high-quality in-image localization in training data to reduce the risks of unintentionally biasing the training process of pathology-detecting AI algorithms. KEY POINTS: • Established pneumothorax-detecting artificial intelligence algorithms trained on public training data are strongly limited and biased by confounding thoracic tubes. • We used high-quality in-image annotated training data to effectively boost algorithm performance and suppress the impact of confounding thoracic tubes. • Based on our results, we hypothesize that even hidden confounders might be effectively addressed by in-image annotations of pathology-related image features.

Comparison between software volumetric breast density estimates in breast tomosynthesis and digital mammography images in a large public screening cohort.

OBJECTIVES: To compare software estimates of volumetric breast density (VBD) based on breast tomosynthesis (BT) projections to those based on digital mammography (DM) images in a large screening cohort, the Malmö Breast Tomosynthesis Screening Trial (MBTST). METHODS: DM and BT images of 9909 women (enrolled 2010-2015) were retrospectively analysed with prototype software to estimate VBD. Software calculation is based on a physics model of the image acquisition process and incorporates the effect of masking in DM based on accumulated dense tissue areas. VBD (continuously and categorically) was compared between BT [central projection (mediolateral oblique view (MLO)] and two-view DM, and with radiologists' BI-RADS density 4th ed. scores. Agreement and correlation were investigated with weighted kappa (κ), Spearman's correlation coefficient (r), and Bland-Altman analysis. RESULTS: There was a high correlation (r = 0.83) between VBD in DM and BT and substantial agreement between the software breast density categories [observed agreement, 61.3% and 84.8%; κ = 0.61 and ĸ = 0.69 for four (a/b/c/d) and two (fat involuted vs. dense) density categories, respectively]. There was moderate agreement between radiologists' BI-RADS scores and software density categories in DM (ĸ = 0.55) and BT (ĸ = 0.47). CONCLUSIONS: In a large public screening setting, we report a substantial agreement between VBD in DM and BT using software with special focus on masking effect. This automated and objective mode of measuring VBD may be of value to radiologists and women when BT is used as the primary breast cancer screening modality. KEY POINTS: • There was a high correlation between continuous volumetric breast density in DM and BT. • There was substantial agreement between software breast density categories (four groups) in DM and BT; with clinically warranted binary software breast density categories, the agreement increased markedly. • There was moderate agreement between radiologists' BI-RADS scores and software breast density categories in DM and BT.

Deep learning in computed tomography super resolution using multi-modality data training.

BACKGROUND: One of the limitations in leveraging the potential of artificial intelligence in X-ray imaging is the limited availability of annotated training data. As X-ray and CT shares similar imaging physics, one could achieve cross-domain data sharing, so to generate labeled synthetic X-ray images from annotated CT volumes as digitally reconstructed radiographs (DRRs). To account for the lower resolution of CT and the CT-generated DRRs as compared to the real X-ray images, we propose the use of super-resolution (SR) techniques to enhance the CT resolution before DRR generation. PURPOSE: As spatial resolution can be defined by the modulation transfer function kernel in CT physics, we propose to train a SR network using paired low-resolution (LR) and high-resolution (HR) images by varying the kernel's shape and cutoff frequency. This is different to previous deep learning-based SR techniques on RGB and medical images which focused on refining the sampling grid. Instead of generating LR images by bicubic interpolation, we aim to create realistic multi-detector CT (MDCT) like LR images from HR cone-beam CT (CBCT) scans. METHODS: We propose and evaluate the use of a SR U-Net for the mapping between LR and HR CBCT image slices. We reconstructed paired LR and HR training volumes from the same CT scans with small in-plane sampling grid size of 0.20 × 0.20 mm 2 $0.20 \times 0.20 \, {\rm mm}^2$ . We used the residual U-Net architecture to train two models. SRUN R e s K $^K_{Res}$ : trained with kernel-based LR images, and SRUN R e s I $^I_{Res}$ : trained with bicubic downsampled data as baseline. Both models are trained on one CBCT dataset (n = 13 391). The performance of both models was then evaluated on unseen kernel-based and interpolation-based LR CBCT images (n = 10 950), and also on MDCT images (n = 1392). RESULTS: Five-fold cross validation and ablation study were performed to find the optimal hyperparameters. Both SRUN R e s K $^K_{Res}$ and SRUN R e s I $^I_{Res}$ models show significant improvements (p-value < $<$ 0.05) in mean absolute error (MAE), peak signal-to-noise ratio (PSNR) and structural similarity index measures (SSIMs) on unseen CBCT images. Also, the improvement percentages in MAE, PSNR, and SSIM by SRUN R e s K $^K_{Res}$ is larger than SRUN R e s I $^I_{Res}$ . For SRUN R e s K $^K_{Res}$ , MAE is reduced by 14%, and PSNR and SSIMs increased by 6 and 8%, respectively. To conclude, SRUN R e s K $^K_{Res}$ outperforms SRUN R e s I $^I_{Res}$ , which the former generates sharper images when tested with kernel-based LR CBCT images as well as cross-modality LR MDCT data. CONCLUSIONS: Our proposed method showed better performance than the baseline interpolation approach on unseen LR CBCT. We showed that the frequency behavior of the used data is important for learning the SR features. Additionally, we showed cross-modality resolution improvements to LR MDCT images. Our approach is, therefore, a first and essential step in enabling realistic high spatial resolution CT-generated DRRs for deep learning training.

Adversarial robustness improvement for X-ray bone segmentation using synthetic data created from computed tomography scans.

Deep learning-based image analysis offers great potential in clinical practice. However, it faces mainly two challenges: scarcity of large-scale annotated clinical data for training and susceptibility to adversarial data in inference. As an example, an artificial intelligence (AI) system could check patient positioning, by segmenting and evaluating relative positions of anatomical structures in medical images. Nevertheless, data to train such AI system might be highly imbalanced with mostly well-positioned images being available. Thus, we propose the use of synthetic X-ray images and annotation masks forward projected from 3D photon-counting CT volumes to create realistic non-optimally positioned X-ray images for training. An open-source model (TotalSegmentator) was used to annotate the clavicles in 3D CT volumes. We evaluated model robustness with respect to the internal (simulated) patient rotation α on real-data-trained models and real&synthetic-data-trained models. Our results showed that real&synthetic- data-trained models have Dice score percentage improvements of 3% to 15% across different α groups compared to the real-data-trained model. Therefore, we demonstrated that synthetic data could be supplementary used to train and enrich heavily underrepresented conditions to increase model robustness.

Artificial Intelligence in Chest Radiography Reporting Accuracy: Added Clinical Value in the Emergency Unit Setting Without 24/7 Radiology Coverage.

OBJECTIVES: Chest radiographs (CXRs) are commonly performed in emergency units (EUs), but the interpretation requires radiology experience. We developed an artificial intelligence (AI) system (precommercial) that aims to mimic board-certified radiologists' (BCRs') performance and can therefore support non-radiology residents (NRRs) in clinical settings lacking 24/7 radiology coverage. We validated by quantifying the clinical value of our AI system for radiology residents (RRs) and EU-experienced NRRs in a clinically representative EU setting. MATERIALS AND METHODS: A total of 563 EU CXRs were retrospectively assessed by 3 BCRs, 3 RRs, and 3 EU-experienced NRRs. Suspected pathologies (pleural effusion, pneumothorax, consolidations suspicious for pneumonia, lung lesions) were reported on a 5-step confidence scale (sum of 20,268 reported pathology suspicions [563 images × 9 readers × 4 pathologies]) separately by every involved reader. Board-certified radiologists' confidence scores were converted into 4 binary reference standards (RFSs) of different sensitivities. The RRs' and NRRs' performances were statistically compared with our AI system (trained on nonpublic data from different clinical sites) based on receiver operating characteristics (ROCs) and operating point metrics approximated to the maximum sum of sensitivity and specificity (Youden statistics). RESULTS: The NRRs lose diagnostic accuracy to RRs with increasingly sensitive BCRs' RFSs for all considered pathologies. Based on our external validation data set, the AI system/NRRs' consensus mimicked the most sensitive BCRs' RFSs with areas under ROC of 0.940/0.837 (pneumothorax), 0.953/0.823 (pleural effusion), and 0.883/0.747 (lung lesions), which were comparable to experienced RRs and significantly overcomes EU-experienced NRRs' diagnostic performance. For consolidation detection, the AI system performed on the NRRs' consensus level (and overcomes each individual NRR) with an area under ROC of 0.847 referenced to the BCRs' most sensitive RFS. CONCLUSIONS: Our AI system matched RRs' performance, meanwhile significantly outperformed NRRs' diagnostic accuracy for most of considered CXR pathologies (pneumothorax, pleural effusion, and lung lesions) and therefore might serve as clinical decision support for NRRs.

An Artificial Intelligence-Based Chest X-ray Model on Human Nodule Detection Accuracy From a Multicenter Study.

Importance: Most early lung cancers present as pulmonary nodules on imaging, but these can be easily missed on chest radiographs. Objective: To assess if a novel artificial intelligence (AI) algorithm can help detect pulmonary nodules on radiographs at different levels of detection difficulty. Design, Setting, and Participants: This diagnostic study included 100 posteroanterior chest radiograph images taken between 2000 and 2010 of adult patients from an ambulatory health care center in Germany and a lung image database in the US. Included images were selected to represent nodules with different levels of detection difficulties (from easy to difficult), and comprised both normal and nonnormal control. Exposures: All images were processed with a novel AI algorithm, the AI Rad Companion Chest X-ray. Two thoracic radiologists established the ground truth and 9 test radiologists from Germany and the US independently reviewed all images in 2 sessions (unaided and AI-aided mode) with at least a 1-month washout period. Main Outcomes and Measures: Each test radiologist recorded the presence of 5 findings (pulmonary nodules, atelectasis, consolidation, pneumothorax, and pleural effusion) and their level of confidence for detecting the individual finding on a scale of 1 to 10 (1 representing lowest confidence; 10, highest confidence). The analyzed metrics for nodules included sensitivity, specificity, accuracy, and receiver operating characteristics curve area under the curve (AUC). Results: Images from 100 patients were included, with a mean (SD) age of 55 (20) years and including 64 men and 36 women. Mean detection accuracy across the 9 radiologists improved by 6.4% (95% CI, 2.3% to 10.6%) with AI-aided interpretation compared with unaided interpretation. Partial AUCs within the effective interval range of 0 to 0.2 false positive rate improved by 5.6% (95% CI, -1.4% to 12.0%) with AI-aided interpretation. Junior radiologists saw greater improvement in sensitivity for nodule detection with AI-aided interpretation as compared with their senior counterparts (12%; 95% CI, 4% to 19% vs 9%; 95% CI, 1% to 17%) while senior radiologists experienced similar improvement in specificity (4%; 95% CI, -2% to 9%) as compared with junior radiologists (4%; 95% CI, -3% to 5%). Conclusions and Relevance: In this diagnostic study, an AI algorithm was associated with improved detection of pulmonary nodules on chest radiographs compared with unaided interpretation for different levels of detection difficulty and for readers with different experience.

Volumetric breast density measurement for personalized screening: accuracy, reproducibility, consistency, and agreement with visual assessment.

Assessment of breast density at the point of mammographic examination could lead to optimized breast cancer screening pathways. The onsite breast density information may offer guidance of when to recommend supplemental imaging for women in a screening program. A software application (Insight BD, Siemens Healthcare GmbH) for fast onsite quantification of volumetric breast density is evaluated. The accuracy of the method is assessed using breast tissue equivalent phantom experiments resulting in a mean absolute error of 3.84%. Reproducibility of measurement results is analyzed using 8427 exams in total, comparing for each exam (if available) the densities determined from left and right views, from cranio-caudal and medio-lateral oblique views, from full-field digital mammograms (FFDM) and digital breast tomosynthesis (DBT) data and from two subsequent exams of the same breast. Pearson correlation coefficients of 0.937, 0.926, 0.950, and 0.995 are obtained. Consistency of the results is demonstrated by evaluating the dependency of the breast density on women's age. Furthermore, the agreement between breast density categories computed by the software with those determined visually by 32 radiologists is shown by an overall percentage agreement of 69.5% for FFDM and by 64.6% for DBT data. These results demonstrate that the software delivers accurate, reproducible, and consistent measurements that agree well with the visual assessment of breast density by radiologists.

Comparison of screening performance metrics and patient dose of two mammographic image acquisition modes in the Danish National Breast Cancer Screening Programme.

INTRODUCTION: In this study, screening performance metrics and radiation dose were compared for two image acquisition modes for breast cancer screening with MAMMOMAT Inspiration (Siemens Healthcare GmbH, Forchheim, Germany). This mammography system can operate without an anti-scatter grid in place but using software scatter correction instead. This grid-less acquisition mode (PRIME) requires less patient dose due to the increase in primary radiation reaching the detector. This study retrospectively analyses data from the Region of Southern Denmark where the grid-less mode has been installed in November 2013 and replaced grid-based screening. METHODS AND MATERIALS: A total of 72,188 screening cases from the same geographical region in Denmark were included in the study. They were subdivided into two study populations: cases acquired before and after installation of the grid-less acquisition mode. Sensitivity and specificity of breast cancer screening were calculated for the two populations; thus representing the performance of grid-less and grid-based screening. To measure the entrance surface air kerma (ESAK) additional phantom tests were carried out. Polymethylmethacrylate (PMMA) attenuation plates with different thicknesses (20-70 mm in steps of 10 mm) simulated the compressed breast (21 mm-90 mm) and a solid-state dosimeter was used. RESULTS: Statistical testing of the results showed that screening with grid-less acquisition provides equivalent performance with respect to sensitivity and specificity compared to grid-based screening. The specificity was 98.11% (95% confidence interval (CI) from 97.93% to 98.29%) and 97.96% (95% CI from 97.84% to 98.09%) for screening with grid-less acquisition and grid-based acquisition, respectively. The cancer detection rate as a measure for sensitivity was equal (0.55%) for grid-less screening and grid-based screening. An average glandular dose saving between 13.5% and 36.4% depending on breast thickness in grid-less acquisition was obtained compared to grid-based acquisition. CONCLUSION: Statistically significant equivalence was shown with an equivalence margin of 0.12% points for cancer detection rate and with an equivalence margin of 0.40% points for specificity. A marked patient dose savings in grid-less acquisition of up to 36% compared to grid-based acquisition was achieved. It can be concluded that grid-less acquisition with software scatter correction is an alternative to grid-based acquisition in mammography.

Initial evaluation of image performance of a 3-D x-ray system: phantom-based comparison of 3-D tomography with conventional computed tomography.

Phantom-based initial performance assessment of a prototype three-dimensional (3-D) x-ray system and comparison of 3-D tomography with computed tomography (CT) were proposed. A 3-D image quality phantom was scanned with a prototype version of 3-D cone-beam CT imaging implemented on a twin robotic x-ray system using three trajectories (163 deg = table, 188 deg = upright, and 200 deg = side), six tube voltages (60, 70, 81, 90, 100, and 121 kV), and four detector doses (0.348, 0.696, 1.740, and [Formula: see text]). CT was obtained with a clinical protocol. Spatial resolution (line pairs/cm) and soft-tissue-contrast resolution were assessed by two independent readers. Radiation dose was assessed. Descriptive and analysis of variance (ANOVA) ([Formula: see text]) were performed. With 3-D tomography, a maximum of 16 lp/cm was visible and best soft-tissue-contrast resolution was 2 mm at 30 Hounsfield units (HU) for 160 projections. With CT, 10 lp/cm was visible and soft-tissue-contrast resolution was 4 mm at 20 HU. The upright trajectory yielded significantly better spatial resolution and soft tissue contrast, and the side trajectory yielded significantly higher soft tissue contrast than the table trajectory ([Formula: see text]). Radiation dose was higher in 3-D tomography (45 to 704 mGycm) than CT (44 mGycm). Three-dimensional tomography renders overall equal or higher spatial resolution and comparable soft tissue contrast to CT for medium- and high-dose protocols in the side and upright trajectories, but with higher radiation doses.

Initial Assessment of a Prototype 3D Cone-Beam Computed Tomography System for Imaging of the Lumbar Spine, Evaluating Human Cadaveric Specimens in the Upright Position.

OBJECTIVES: The aims of this study were to assess feasibility, image quality, and radiation dose and to estimate the optimal dose protocol for the lumbar spine of cadaveric specimens with different body mass indices (BMIs) in the upright position using a prototype 3-dimensional cone-beam computed tomography (CT) software implemented on a robotic x-ray system and compare with CT. MATERIALS AND METHODS: The lumbar spine of 5 formalin-fixed human cadaveric specimens (BMI, 22-35 kg/m) was prospectively assessed in the upright position using prototype software for 3-dimensional tomography implemented on a robotic x-ray system. Specimens were scanned with varying kilovolt values (70, 81, 90, 100, 109, 121 kV) and thereafter with 80 kV (BMI ≤30 kg/m) and 121 kV (BMI >30 kg/m) and varying dose levels (DLs; 0.278, 0.435, 0.548, 0.696, 0.87, 1.09). Computed tomography data were acquired with a standard clinical protocol. Two independent readers rated visibility of the cortex, endplates, facet joints, trabeculae, neuroforamina, posterior alignment, and spinal canal as well as nerve roots. Radiation dose was measured with a cylindrical CTDI phantom. Descriptive statistics and analysis of variance were used (P < 0.05). RESULTS: Average intraclass correlation was excellent (0.94). The lowest technically possible kilovolt and the highest technically possible DL yielded the best image quality. In specimens with a BMI of 30 kg/m or less, depiction of all structures was good and comparable to CT, except for nerve roots. For specimens with a BMI greater than 30 kg/m, image quality was limited. CONCLUSIONS: Three-dimensional cone-beam CT of the lumbar spine in cadaveric specimens in the upright position is feasible. An optimal dose protocol was estimated. Depiction of osseous structures is comparable to CT in specimens with BMI of 30 kg/m or less. Image quality is limited for soft tissue structures and specimens with BMI greater than 30 kg/m.

Dynamic iterative reconstruction for interventional 4-D C-arm CT perfusion imaging.

Tissue perfusion measurement using C-arm angiography systems capable of CT-like imaging (C-arm CT) is a novel technique with potentially high benefit for catheter guided treatment of stroke in the interventional suite. However, perfusion C-arm CT (PCCT) is challenging: the slow C-arm rotation speed only allows measuring samples of contrast time attenuation curves (TACs) every 5-6 s if reconstruction algorithms for static data are used. Furthermore, the peak values of the TACs in brain tissue typically lie in a range of 5-30 HU, thus perfusion imaging is very sensitive to noise. We present a dynamic, iterative reconstruction (DIR) approach to reconstruct TACs described by a weighted sum of basis functions. To reduce noise, a regularization technique based on joint bilateral filtering (JBF) is introduced. We evaluated the algorithm with a digital dynamic brain phantom and with data from six canine stroke models. With our dynamic approach, we achieve an average Pearson correlation (PC) of the PCCT canine blood flow maps to co-registered perfusion CT maps of 0.73. This PC is just as high as the PC achieved in a recent PCCT study, which required repeated injections and acquisitions.

Interventional 4-D C-arm CT perfusion imaging using interleaved scanning and partial reconstruction interpolation.

Tissue perfusion measurement during catheter-guided stroke treatment in the interventional suite is currently not possible. In this work, we present a novel approach that uses a C-arm angiography system capable of computed tomography (CT)-like imaging (C-arm CT) for this purpose. With C-arm CT one reconstructed volume can be obtained every 4-6 s which makes it challenging to measure the flow of an injected contrast bolus. We have developed an interleaved scanning (IS) protocol that uses several scan sequences to increase temporal sampling. Using a dedicated 4-D reconstruction approach based on partial reconstruction interpolation (PRI) we can optimally process our data. We evaluated our combined approach (IS-PRI) with simulations and a study in five healthy pigs. In our simulations, the cerebral blood flow values (unit: ml/100 g/min) were 60 (healthy tissue) and 20 (pathological tissue). For one scan sequence the values were estimated with standard deviations of 14.3 and 2.9, respectively. For two interleaved sequences the standard deviations decreased to 3.6 and 1.5, respectively. We used perfusion CT to validate the in vivo results. With two interleaved sequences we achieved promising correlations ranging from r=0.63 to r=0.94. The results suggest that C-arm CT tissue perfusion imaging is feasible with two interleaved scan sequences.

A model for filtered backprojection reconstruction artifacts due to time-varying attenuation values in perfusion C-arm CT.

Filtered backprojection is the basis for many CT reconstruction tasks. It assumes constant attenuation values of the object during the acquisition of the projection data. Reconstruction artifacts can arise if this assumption is violated. For example, contrast flow in perfusion imaging with C-arm CT systems, which have acquisition times of several seconds per C-arm rotation, can cause this violation. In this paper, we derived and validated a novel spatio-temporal model to describe these kinds of artifacts. The model separates the temporal dynamics due to contrast flow from the scan and reconstruction parameters. We introduced derivative-weighted point spread functions to describe the spatial spread of the artifacts. The model allows prediction of reconstruction artifacts for given temporal dynamics of the attenuation values. Furthermore, it can be used to systematically investigate the influence of different reconstruction parameters on the artifacts. We have shown that with optimized redundancy weighting function parameters the spatial spread of the artifacts around a typical arterial vessel can be reduced by about 70%. Finally, an inversion of our model could be used as the basis for novel dynamic reconstruction algorithms that further minimize these artifacts.

Deconvolution-Based CT and MR Brain Perfusion Measurement: Theoretical Model Revisited and Practical Implementation Details.

Deconvolution-based analysis of CT and MR brain perfusion data is widely used in clinical practice and it is still a topic of ongoing research activities. In this paper, we present a comprehensive derivation and explanation of the underlying physiological model for intravascular tracer systems. We also discuss practical details that are needed to properly implement algorithms for perfusion analysis. Our description of the practical computer implementation is focused on the most frequently employed algebraic deconvolution methods based on the singular value decomposition. In particular, we further discuss the need for regularization in order to obtain physiologically reasonable results. We include an overview of relevant preprocessing steps and provide numerous references to the literature. We cover both CT and MR brain perfusion imaging in this paper because they share many common aspects. The combination of both the theoretical as well as the practical aspects of perfusion analysis explicitly emphasizes the simplifications to the underlying physiological model that are necessary in order to apply it to measured data acquired with current CT and MR scanners.