Whitepaper Features, Data, and Algorithms AI-Rad Companion Chest CT

Features, Data, and Algorithms AI-Rad Companion Chest CT VA13 Whitepaper – September 2021 SIEMENS Healthineers Whitepaper · Features, Data, and Algorithms Table of contents Introduction …………………………………………………………… 3 Product features 3 …………………………………………………… Workflow ……………………………………………………………… 6 Algorithm description……………………………………………… 7 Data requirements …………………………………………………10 Proof points: performance and clinical value ……………12 References ……………………………………………………………15 2 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper Introduction AI-Rad Companion Chest CT is a decision support The typical workflow consists out of the following steps: tool for the radiological assessment of computed tomography (CT) images of the thorax. It helps 1. Reconstructed CT images of the thorax are sent radiologists interpret CT images of the thorax more to the PACS for interpretation. quickly and more precisely (doing more by doing less) and reduces the time needed to document the 2. In parallel, they are also sent to AI-Rad Companion. findings with the help of automatic measurements. Extensions are launched automatically. It is vendor-neutral, which means that the software can evaluate image data from any CT system 3. The results of AI-Rad Companion can either be sent manufacturer. Enabled by the teamplay digital to a web-based review software or directly to PACS. health platform and using state-of-the-art image Here they can be used in combination with the original processing algorithms supported by artificial data for reporting purposes. intelligence, AI-Rad Companion Chest CT delivers value in four key areas: This whitepaper is intended to provide an overview of the product features, describe the individual 1. Accelerated interpretation and workflow efficiency algorithmic components, discuss requirements for data to be processed using the device, and to provide 2. Improved clinical outcomes and increased accuracy an overview of internal and external proof points assessing the performance and illustrating the clinical 3. Provision of additional clinically relevant information value of the application. and visual highlighting 4. Standardized results while helping to reduce inter- Product features reader variability It focuses on three main parts of the thorax: the lungs AI-Rad Companion Chest CT consists of three medical (AI-Rad Companion (Pulmonary)), the cardiovascular devices: AI-Rad Companion (Pulmonary), AI-Rad system (AI-Rad Companion (Cardiovascular)) and the Companion (Cardiovascular) and AI-Rad Companion spine (AI-Rad Companion (Musculoskeletal)). (Musculoskeletal). AI-Rad Companion Local Cloud teamplay digital health platform is not commercially available Siemens Healthcare GmbH, 2021 3 in all countries. Its future availability cannot be guaranteed. Whitepaper · Features, Data, and Algorithms AI-Rad Companion (Pulmonary) • Detection and segmentation of lung nodules The radiological assessment of pulmonary nodules and localization with respect to lung lobes is one of the most common indications for Chest CT. The radiologist needs to identify the nodules, measure • Analysis of the lung parenchyma based the diameters and – ideally – also their volume (see on segmented lung lobes with respect to: guidelines of the Fleischner Society [1]). A second • areas of low attenuation important application of Chest CT is the analysis of (low attenuation volume, or LAV) lung parenchyma. Reduced density can indicate • areas of opacity emphysema and/or COPD while increased density can • volume of lung lobes indicate inflammatory processes such as pneumonia. Exemplary outputs of AI-Rad Companion (Pulmonary) AI-Rad Companion (Pulmonary) provides the following are shown in Figure 1. In the product, the LAV-Analysis features with respect to the analysis of the lung: is called “Lung Parenchyma Analysis”, while the opacity analysis is called “Pulmonary Density”1. 1A L2 L1 L2: 8.3 mm 8.3 mm A 1B 1C F Figure 1: Outputs of the Pulmonary feature. Lung nodule detection and measurement (1A), LAV-analysis (1B), opacity detection (1C). 1 The Pulmonary Density feature is new in VA12A without FDA clearance. According to FDA policy “Enforcement Policy for Imaging Systems During the Coronavirus Disease 2019 (COVID-19) Public Health Emergency” issued in April 2020, the manufacturer is allowed to market this feature without FDA clearance. This policy is intended to remain in effect only for the duration of the public health emergency related to COVID-19 declared by the HHS, including any renewals made by the HHS Secretary in accordance with section 319(a)(2) of the Public Health Services Act (42 U.S.C. 247d(a)(2)). Pulmonary Density results are not indicated for the diag- nosis of COVID-19. Only in vitro diagnostic testing is currently the definitive method to diagnose COVID-19. 4 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper AI-Rad Companion (Cardiovascular) are far better than the results of treatment required for For the radiological assessment of the cardiovascular acute and often catastrophic disease presentations. system a large variety of dedicated CT scan protocols Thus, the identification and treatment of patients at risk exist depending on the clinical indication. The protocols for acute and catastrophic disease presentations prior differ mainly with respect to the cardiac phase in which to such an occurrence are paramount to eliminating the the acquisition is performed (controlled via ECG-gating) high morbidity and mortality associated with acute and the type and timing of contrast enhancement. presentations” and hence motivates the automatic analysis AI-Rad Companion is designed to work with any of these of the thoracic aorta on any chest CT. In the guidelines protocols, particularly the most generic non-gated and it is also described that the thoracic aorta should be non-contrast-enhanced Chest CT scans. Of course, this measured at nine predefined anatomical locations. “The also limits the analysis to features that can be reliably use of standardized measurements helps minimize errant assessed on generic chest CT data. The features are: reports of significant aneurysm growth due to technique or interreader variability in measuring technique.” [2] • Measurement of heart volume and coronary calcium volume (on unenhanced data only) 2A • Segmentation of aorta and diameter measurements at 9 landmarks according to AHA-guidelines [2] (on both native and contrast-enhanced data) Exemplary outputs of AI-Rad Companion (Cardiovascular) are shown in Figure 2. It is important to understand that the user should in- terpret the results of AI-Rad Companion (Cardiovascular) AL with respect to the actual scan protocol used. E.g., 2B 4 motion artifacts on a non-gated exam may hamper the accuracy of the aortic diameter measurements. Analogously, the coronary calcium analysis provides the total volume of the – potentially motion corrupted – calcium clusters but does not perform Agatstson 55 mm scoring which requires a gated scan and is the gold standard for dedicated cardiac CT scans. However, the importance of the analysis of both coro- nary calcium and aorta in the context of chest CT is to be pointed out. Both features are listed in the recommenda- tions by the ACR Incidental Findings Committee [3]. The 2016 SCCT/STR guidelines [4] state that coronary artery calcium “should be evaluated and reported on all non-contrast chest CT examinations”. Analogously, in a consensus statement the British societies BSCi/BSCCT and BSTI [5] “recommend that coronary artery calcification is reported on all non-gated thoracic CT using a simple AL patient-based score (none, mild, moderate, severe)”. 9 8 In their 2010 guidelines [2] the ACCF/AHA states that “many thoracic aortic diseases, results of treatment for Figure 2: Outputs of the Cardiovascular device. stable, often asymptomatic, but high-risk conditions Coronary calcium detection (2A), aorta analysis (2B). Siemens Healthcare GmbH, 2021 5 Whitepaper · Features, Data, and Algorithms AI-Rad Companion (Musculoskeletal) T1 T2 T3 |T4 |T5 /16 /T7 T1 15 / 15 / 18 / 181 Osteoporosis manifests as loss of bone density e.g. in the T2 16 / 16/ 16 / 148 spine, and consequently in vertebral compression fractures. 19/ 16/ 19 / 129 The International Osteoporosis Foundation (IOF) states T4 19/ 15/ 20 /142 that “there is strong evidence of widespread under-diag- T5 20 / 18 / 20 / 131 nosis of vertebral fractures” [6]. Pickhardt et al. [7] and T6 19 / 19 / 23 /110 more recently Cohen et al. [8] showed that the HU-values 17 21 / 19 / 23 / of the spine obtained from CT data acquired for other T8 20 / 19 / 24 / 109 indications can be used to identify osteoporotic patients T9 21 / 20 / 23 / 124 and called this approach “opportunistic screening for osteo- porosis”. AI-Rad Companion (Musculoskeletal) provides: OT10 24 / 23 / 24 /119 T11 24 / 23 / 26 / 107 • Labeling and segmentation of thoracic vertebras Measurements of vertebrae heights T12 11/ 91 22 22 /191 • T9 T10 • Mean vertebral body HU-density Figure 3: Output of the Musculoskeletal device: Height Exemplary outputs are shown in Figure 3. and density measurements of the thoracic vertebrae. Workflow AI-Rad Companion Chest CT offers advanced ways opened. A results table summarizes all findings and of workflow customization. By design, all results are measurements. A color-coding scheme is used to draw presented in the form of an annotated axial series, the attention to potential abnormalities. Added 3D a 3D rendering, and a concise summary table – renderings quickly provide an intuitive presentation enabling seamless integration into a verity of different overview of the type, number, and spatial context reading workflows. Moreover, a DICOM Structured of all findings. Upon confirmation of the findings, the Report with measurement results is provided. results are straightforwardly transferred to the report. 1. Efficiency gains are best accomplished when AI-Rad 2. Results of the AI-Rad Companion are best incorporated Companion is used to automate the repetitive into the primary read by synchronizing annotated and manually tedious task such as measurements. axials with the corresponding original series. As the All results are readily available the moment a case is reader scrolls through the stack, through highlighting 1 2 PATIENT NOPDFIMA TION Ang-ission Cole and Time LESIONS Lobe MAX. 2D Riged Upper Lete AORTA --- At Daghr. Pros. Arch Abe. Aerts 10d Arch Aarte Colegoy" 6 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper 3 findings their attention is drawn to potential abnor- malities. At the same time, the correctness of the AI results is easily verified through comparison with unannotated series. 3. Minimal disruption to the established workflow and unbiased reading is achieved when AI is used in a “spell checker” mode. Here, the reader would stick to their established reading patterns, but just before closing the case one last glance at the results pictogram allows for a quick and easy confirmation that indeed nothing was missed. Algorithm description Lung nodule detection (lung CAD) a likelihood value (a final confidence score above or and segmentation equal to an empirically determined threshold) are labeled as nodule candidates. Lung CAD processing is performed in several consecutive steps: Preprocessing, Candidate Generation, Classification, and Post-Filtering. The Post-Filtering step includes the application of two cascaded filters. The first one aims at removing false In the Preprocessing step the input image is standard- positives originating (a) from the colon and a second ized, and parenchyma is segmented using specialized one from (b) calcified protrusions (for example, Convolutional Neural Net (CNN, V-Net). This allows areas where the sternum meets the manubrium, spine restricting the detection of findings within the lungs malformations, and osteophytes, and so on). while optimizing the computation time. The first filter is a CNN-based classifier that has a similar structure to that of the classifier in the Classification Candidate Generation aims at achieving high sensitivity while keeping the number of candidates to a manage- step. The second filter uses three orthogonal slices able number. The isotropic volume is partitioned at the candidate location as input to three CNN-based into sub-volumes that are processed using a CNN. Then, classifiers (one per slice). The results from the three filtering and non-maximum suppression yield a list of classifiers is then combined by a max-voting mechanism. candidates for each sub-volume. A predefined threshold Any candidate deemed a false positive by either filter is thus removed. is applied on the confidence score to remove the least confidence (low score) findings. All candidates exceeding this threshold are passed on to the Classification step. The algorithms have been trained using more than 2,000 manually curated CT data sets. Network layout Classification utilizes a CNN-based classifier to process diagrams have been published by Chamberlin et al. [9]. each candidate. The classifier calculates the feature After detection, nodules are segmented by an algorithm values for each candidate and uses a soft-max function to estimate the likelihood of its type as either “nodule” based on region growing. Diameter and volume measure- or “non-nodule.” Candidates meeting or exceeding ments are provided. Siemens Healthcare GmbH, 2021 7 Whitepaper · Features, Data, and Algorithms Lung lobe segmentation The detected opacities are quantified by computing the The lung lobe segmentation algorithm computes percentage of opacity (PO, per lobe and per lung) and segmentation masks of the five lung lobes for a given the percentage of high opacities (PHO, by applying CT data set of the chest. First, multi-scale deep reinforce- a threshold of -200 HU on the subset of opaque regions). ment learning [10] is used to robustly detect anatomical Based on the PO a lung severity score (LSS) is calculated landmarks in a CT volume. The carina bifurcation according to Bernheim et al. [15]: and/or sternum tip are used to identify the lung region of interest (ROI). Next, the lung ROI image is resampled LSS Condition to a 2 mm isotropic volume and fed into an adversarial Deep Image-to-Image Network (DI2IN) [11] to generate 0 PO = 0 the lung segmentation. Finally, the ROI segmentation mask is remapped to have the same dimension and the 1 0 < PO ≤ 25 % resolution as the input data. The DI2IN has been first trained on over 8,000 CT scans from a large group 2 25 % < PO ≤ 50 % of patients with various diseases, then fine-tuned with 3 50 % < PO ≤ 75 % over 1,000 scans with abnormal patterns including interstitial lung disease (ILD), pneumonia, and COVID-19. 4 PO > 75% The volume of the individual lung lobes, the left and Table 2: Thresholds for PO. right lung and of the complete lung are reported. A total LSS is computed as the sum of the individual scores per lobe. LAV analysis The LAV analysis is threshold-based, i.e. the algorithm determines all voxels below -950 Hounsfield Units (HU) Heart segmentation in the lung. The threshold of -950 HU is widely used The heart segmentation is performed using a deep for the quantification of emphysema [12]. For each lung U-shaped network [13] consisting of four convolutions lobe as well as for the complete lung (i.e. the combina- and down-sampling steps, followed by four similar tion of all lobes) the ratio of LAV (LAV %) is reported. up-sampling layers. It has been trained on over 650 CT The following thresholds are being used as default values data sets. Subsequently, the heart segmentation mask for highlighting: is used to compute the heart volume. Category Condition Coronary calcium detection I LAV% < 12.5% Using the heart mask as ROI an initial set of voxels as candidates for potentially calcified regions is obtained II 12.5 % ≤ LAV % < 25 % by thresholding at 130 HU. For each candidate voxel III 25 % ≤ LAV % < 37.5 % an image patch centered around the voxel is fed into a deep learning-based classification algorithm. The deep IV LAV % ≥ 37.5 % learning model has two components: a convolutional neural network, which takes the image patch and a Table 1: Thresholds for LAV %. precomputed coronary territory map as inputs, and a dense neural network which operates on the coordinates of the voxel. A final prediction is made by combining Opacity detection and quantification features from both components to determine whether The detection and quantification of opaque regions in the voxel belongs to the coronary arteries. The algorithm the lung – typically associated to viral pneumonia such has been trained on over 1,200 ECG-gated calcium as Covid-19 – uses a DenseUNet [13] with anisotropic scoring scans and fine-tuned on over 550 chest CTs. kernels. Details of the algorithm are described by Additional details on the computational pipeline Chaganti et al. [14]. The algorithm has been trained on and the network topology have been described by over 900 CT scans from patients with ILDs, pneumonia, Chamberlin et al. [9]. and COVID-19. 8 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper The total volume V of the detected coronary calcium Given the aorta mask, a centerline model is used is used for threshold-based categorization. Several to generate the aortic centerline. The centerline thresholds for total calcium volume have been proposed is used in combination with aortic landmarks to identify in the literature. For instance, based on the NELSON measurement planes at nine locations according to study, Mets et al. [16] showed that a coronary calcium the guidelines of the American Heart Association [2]. volume of 100 mm3 corresponds to an 8 % increased risk In each of the measurement planes, multiple diameters of cardiovascular events and 500 mm3 to an increased are computed by computing intersections of rays risk of 48 %. These volumes were used as default thresh- starting from the centerline with the aortic mask. olds in AI-Rad Companion Chest CT 1 . A third threshold Based on these diameters, the maximum in-plane at 10 mm3 is used to compensate for image noise. diameter and the diameter along a direction orthogonal Slightly different values have been used in a publication to the maximal diameter are reported. by van Assen et al [17]. An overview is provided in Table 3. The maximum diameters d are used for threshold- based categorization. Recommended thresholds 2 for Category Thresholds derived Thresholds used by from Mets et al. [16] van Assen et al. [17] the severity of a potential aortic dilation or aneurysm have been derived from the AHA-Guidelines and values I V < 10 V < 5 for the population mean and standard variation (std) given therein: II 10 ≤ V < 100 5 ≤ V < 250 III 100 ≤ V < 500 250 ≤ V < 1000 Category Condition IV V ≥ 500 V ≥ 1000 I d ≤ mean + 2*std Table 3: Thresholds for coronary calcium II d > mean + 2*std volume V in mm3. III d > 1.5*mean IV d ≥ 5.5 cm Aorta diameter measurements The aorta analysis pipeline consists of three steps: Table 4: Thresholds for aortic diameters d. landmark detection, aorta segmentation, and diameter measurements. The aorta module works for both native and contrast- enhanced data with and without ECG-gating. Six aortic landmarks (Aortic Root, Aortic Arch Center, Brachiocephalic Artery Bifurcation, Left Common Carotid Artery, Left Subclavian Artery, and Celiac Trunk) are Vertebra labeling and density detected automatically based on Deep Reinforcement measurement Learning [10]. The twelve thoracic vertebrae are localized and The aortic root is used to define a ROI for the segmen- labeled using an algorithm based on wavelet features, tation algorithm. Within the ROI the segmentation is AdaBoost, and local geometry constraints [18]. performed using an adversarial DI2IN in a symmetric Around each vertebra center cylindric regions of interest convolutional encoder-decoder architecture [11]. The are used to measure the average HU-density of the front part is a convolutional encoder-decoder network trabecular bone. with feature concatenation, and the backend is deep supervision network through multi-level. Blocks inside DI2IN consist of convolutional and upscaling layers. Vertebra segmentation and height The algorithm has been trained on over 1,000 CT data measurement sets including native and contrast-enhanced scans. The vertebra centers are also used to determine ROIs for the vertebra segmentation. Within the ROI the segmentation is performed using a DI2IN 1 In the software version available in the United States no default values in a symmetric convolutional encoder-decoder are provided. architecture [11]. The algorithm has been trained 2 In the software version available in the United States the thresholds cannot be adapted by the user. on over 4,300 thoracic vertebrae. Siemens Healthcare GmbH, 2021 9 Whitepaper · Features, Data, and Algorithms From the segmentation masks the sagittal midplane is extracted and within this plane height measurements Category Condition at anterior, medial, and posterior location. Afterwards, I hr ≥ 80 % the height ratios hr are computed by comparing heights of neighboring vertebrae using the Genant severity II 80 % > hr ≥ 75 % grading method [19]. Although originally developed on chest radiographs, the Genant method is a widely used III 75 % > hr ≥ 60 % also in CT imaging [6]: IV hr < 60% Table 5: Thresholds for vertebra height ratios hr. Data requirements Technical requirements generation step is based on HU-thresholding and the AI-Rad Companion Chest CT uses a single DICOM series threshold is not valid for contrast-enhanced scans nor as input for all modules. In general, the algorithms are for kVp < 100. The topic has been discussed in detail intended to work with any chest CT series. However, by Vonder et al. [20] and in a corresponding there are a couple of technical properties required for Siemens Healthineers Whitepaper on calcium quanti- the device to process the cases: fication on dedicated cardiac CT data [21]. • Primary axial images (image orientation 1\0\0\0\1\0) Scan Parameter Recommendations • Volume scans without gaps, no gantry tilt Besides the coronary calcium detection, HU-thresholding • Slice thickness ≤ 3 mm (for MSK ≤ 2 mm), is also used in the LAV-analysis of the lung parenchyma. preferably ≤ 1 mm, see recommendations below As a consequence, the results of these two features • Matrix size 512 x 512 are sensitive to image noise. Image noise in CT data • Photometric interpretation: MONOCHROME 2 depends on many parameters, most prominently on slice • 16 bit, no lossy compression, samples per pixel: 1 thickness, reconstruction kernel, and dose. Hence the • Rescale slope ≤ 5 combination of thin slices, hard kernels, and low dose may result in very noisy images. For such data the The cardiac module (heart segmentation and coronary cardiac module would reject the case (if there are too calcium detection) has the additional requirements many calcium candidates) and the LAV analysis may be that the images are without contrast enhancement and confounded by noise-related LAV-patches [22]. kVp ≥ 100. That is because the initial candidate Reconstruction kernel Soft to medium kernel Hard kernel Slice thickness ≤ 1 mm 1–2 mm 2–3 mm ≤ 1 mm 1–2 mm 2–3 mm Lung nodules Lung Parenchyma (LAV and opacities) Aorta Heart and Coronaries Vertebrae Table 6: Recommended scan parameters for AI-Rad Companion Chest CT. fully supported supported but results might be suboptimal not supported 10 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper Publication Patient cohort Feature(s) studied Study size Scanner Model(s) Scan parameters Chamberlin Lung cancer Lung nodules, SOMATOM go.Top, protocol according to ACR-STR et al. [9] screening cor. calcium 117 Definition AS+, Definition LDCT guidelines. Flash, and Force slice thickness: 1.0 mm Paired Cardiac and 95 SOMATOM Definition slice thickness: van Assen et al. [17] Chest CTs, consec- Cor. calcium Flash, Definition AS+, 1.0 mm–3.0 mm, utive Chest CTs + 168 and Force medium sharp kernel Fischer SOMATOM Definition slice thickness 1.5 mm, et al. [22] Emphysema LAV 141 Flash, Force, and Emotion comparing two kernels: lung (B60s) and soft tissue (B31s) Yacoub et al. [23] Consecutive cases all 100 SOMATOM Definition slice thickness: 1.0 mm, Flash, and Force soft tissue kernel Lung nodules, Rückel et al. [24] Emergency CT aorta diam., cor. 105 SOMATOM Force slice thickness: 0.75 mm, calcium, heart size, soft tissue kernel Br36d vert. heights Fischer et al. [25] COPD Lung lobes, LAV 137 SOMATOM Definition slice thickness 1.5 mm, Flash, Force, and Emotion lung kernel SOMATOM Definition Rückel Aortic aneurysm et al. [27] Aorta diam. 18 x 2 Flash, Force, and Defini- slice thickness: follow-up tion AS+, GE Optima 0.6 mm –3.0 mm, CT660, Discovery 750 HD soft tissue kernel Lobe volume, PO, Weikert COVID-19 patients PHO, LSS, LAV, SOMATOM Definition AS+, et al. [28] heart size,cor. calcium, 120 slice thickness: 1.0 mm, and Force soft tissue kernel aorta diam SOMATOM Definition Homayounieh COVID-19 patients Lobe volume, Flash, Force, and Defini- slice thickness: et al. [29] PO, PHO 241 tion Edge, Emotion 16, GE 1.0 mm–2.0 mm, Discovery 750 HD soft tissue kernel B20f Abadia Lung nodules in cases w/ complex Lung lobes, 103 SOMATOM Force slice thickness: 1.0 mm, et al. [30] lung disease Lung nodules + 40 sharp body kernel Table 7: Scan parameters used in various publications using AI-Rad Companion Chest CT. On the other hand, thin slices, i.e. high spatial resolution Companion Chest CT. To achieve optimal results for in z-direction, are required for most of the algorithms, in all modules, it is recommended to use a thin slice with particular for accurate vertebrae height measurements a soft to medium kernel. (ideally slice thickness should be ≤ 1 mm), detailed delin- eation of lung lobes and accurate lung nodule volumetry. Table 7 summarizes scan parameters used in various clinical studies using AI-Rad Companion Chest CT. Details In summary, Table 6 displays the recommendations about these and other studies will be discussed in the of scan parameters for the individual modules of AI-Rad subsequent section. Siemens Healthcare GmbH, 2021 11 Whitepaper · Features, Data, and Algorithms Considerations regarding deviation, but also due to imaging artefacts like motion patient population or noise – one or more algorithms might fail or produce Table 7 also illustrates that AI-Rad Companion incorrect result. In that context it is also important to Chest CT has been used to analyze a broad spectrum note that the output images generated by AI-Rad of patient cohorts: Companion Chest CT are designed in a way that the user can easily verify the correctness of the result. An example • low dose lung cancer screening [9] would be the sagittal MPR of the spine, see Figure 4. • consecutive cohorts, independent of particular clinical indications [23; 17] or with an indication unrelated to the features of AI-Rad Companion Chest CT like data T1 15/ 16/ 11 /277 from the emergency department [24] 2 18/ 15/ 18 /256 T3 19/ 15/ 16 /246 • patients with known disease patterns relevant for 19 / 16/ 19 / 247 the feature of investigation, like emphysema/COPD [25; 22], osteoporosis [26], aortic aneurysms [27] T5 18 / 18/ 18 /240 COVID-19 [28; 29] T6 18/ 16/ 20 /244 17 • patients with known comorbidities that make the 19/ 16/ 21 /235 assessment of the feature under investigation more T8 19/ 16/ 21 /224 challenging, such as the detection of lung nodules in the presence of, e.g. ILD [30]. T9 20 / 19/ 21 / 232 The broad spectrum illustrates the versatile and generic Figure 4: AI-Rad Companion Chest CT output: design of the algorithms of AI-Rad Companion Chest CT. Sagittal view of the spine including height and density On the other hand, one would always find cases where – measurement. The incorrect height measurement for T1 due to severe pathology, comorbidity, or anatomical (due to image noise) can easily be verified by the user. Proof points: performance and clinical value AI-Rad Companion Chest CT offers advanced ways and 92.9 % (presence of lung nodules and presence of of workflow customization. By design, all results are coronary calcifications, respectively) and a specificity presented in the form of an annotated axial series, of 70.8 % and 96.0 %, respectively. The authors also use a 3D rendering, and a concise summary table – enabling the results for predicting of lung cancer and major seamless integration into a verity of different reading adverse cardiac events at 1-year follow-up yielding AUC- workflows. Moreover, a DICOM Structured Report with values of 0.942 and 0.911, respectively, emphasizing measurement results is provided. that “this information can be used to improve diagnostic ability, facilitate intervention, improve morbidity and Accuracy and clinical value mortality, and decrease healthcare costs”. In the study by Chamberlin et al. [9] N= 117 lung cancer screening exams were processed by AI-Rad Companion Focusing on the other end of the spectrum of patient Chest CT and analyzed with respect to lung nodules and cohorts, namely patients with complex lung disease such coronary calcium. The agreement with expert reader as ILD, COPD, bronchitis, edema, and pulmonary embo- has been found excellent (Cohen’s kappa of lung nodule lism, Abadia et al. [30] investigated the accuracy of the detection: 0.846, intraclass correlation coefficient of lung nodule detection and localization (N = 103 plus 40 coronary calcium volume: 0.904) at a sensitivity of 100 % controls). On a patient level AI-Rad Companion Chest CT showed a sensitivity of 89.4 % and a specificity of 82.5 %. 12 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper On the individual nodule level sensitivity was 67.7 %, • 32.4 % aortic ectasia similar to the accuracy reported for experienced • 1.9 % actionable lung nodules radiologists. • 12.4 % vertebra fractures On an unspecific but representative patient population, The authors point out that “In particular, the integration i.e. N= 100 consecutive cases, Yacoub et al. [23] of different specialized algorithms in a single software reported sensitivity and specificity of all features of solution is promising to avoid clinically too narrow AI AI-Rad Companion Chest CT, see Table 8. applications. But also, with regard to less urgent appli- cations of medical imaging, it should also be mentioned Sensitivity that especially non-radiology clinicians might even take N Specificity positive more benefit from AI-assisted image analysis compared cases AI Report AI Report to anyway well-trained radiologists, e.g., in clinical settings without 24/7 radiology coverage or long turn- Lung 83 92.8% 97.6% 82.4% 100% around times for radiology reporting.” nodules Consecutive patients (N= 168) were also enrolled Emphysema 31 80.6% 74.2% 66.7% 97.1% in a study on coronary calcium detection by van Assen et al. [17]. Here the coronary calcium volume computed Aortic by AI-Rad Companion Chest CT was compared against dilation 27 96.3% 25.9% 81.4% 100% the calcium volume obtained from manual calcium scoring. The correlation was found excellent (logarithmic Coronary 59 89.8% 75.4% 100% 94.9% correlation coefficient 0.923). By applying volume- Calcium thresholds (see Table 3) the AI-computed calcium volume Vertebra was categorized into no, mild, moderate, and severe. compression 9 100% 100% 63.7% 100% The categories were compared against qualitative visual rating by an experienced cardiac radiologist. Results Table 8: Sensitivity and specificity of AI-Rad Companion are shown in Table 9.82 % of all cases were correctly Chest CT and radiological reports on N= 100 consecutive classified with all wrongly classified scans being attributed cases as reported by Yacoub et al. [23]. to an adjacent category. The authors state that “such results align well with Expert\AI No Mild Moderate Severe the general recommendation to maximize sensitivity when AI is being used in radiology to detect abnor- No 60 6 0 0 malities, even at the expense of lower specificity, in order Mild 7 44 0 0 to ensure that fewer abnormal findings are missed. Our findings illustrate that the use of AI for diagnostic Moderate 0 6 14 5 reading provides rather a support tool which is not intended to replace the role of a radiologist.” They Severe 0 0 4 20 conclude that “incorporating AI support into radiology workflows can provide significant added value to clinical Table 9: Category agreement between manual radiology reporting”. qualitative assessment and AI determined calcium volume as reported by van Assen et al. [17]. The low sensitivity of the radiologists in particular among the incidental findings has also been studied by Rückel In a second arm of the study, N= 95 patients were identi- et al. [24] in the particular time-critical setting of emer- fied which underwent both dedicated coronary calcium gency CT. The following abnormalities were missing in scoring exams (non-contrasted, ECG-gated cardiac CTs) the original reports but identified by AI-Rad Companion and chest CTs within 1.5 years. For those patients, Chest CT and confirmed by radiologists in a consecutive conventional calcium scoring was performed according cohort of N= 105 whole-body emergency CTs: to Agatston on the cardiac CTs and compared to the calcium volume computed by AI-Rad Companion • 23.8 % increased heart size Chest CT on the chest CT data. By design, the agreement • 16.2 % coronary calcifications of these results will be lower, simply because the data Siemens Healthcare GmbH, 2021 13 Whitepaper · Features, Data, and Algorithms compared originates from different acquisitions from Two publications by Fischer et al. [25; 22] study different time points. Nevertheless, the correlation the results of the lung lobe-based LAV analysis in between manual Agatston score and calcium volume emphysema/COPD patients (N= 141 and N= 137, computed by AI-Rad Companion Chest CT was found respectively). The correlation of LAV with spirometry- excellent (logarithmic correlation coefficient 0.921). based Tiffeneau index was -0.86, and 0.88 with GOLD When comparing threshold-based categories (volume stages, respectively. The LAV of the upper lobes threshold as in Table 3 vs. standard Agatston risk “was also able to most clearly distinguish mild and categories), 70 % of all cases were classified correctly, moderate forms of COPD. This is particularly relevant in only 5 % the prediction was more than one category due to the fact that early disease processes often off. Moreover, a misclassification into the “no calcium” elude conventional pulmonary function diagnostics. category, which – according to the authors [17] – Earlier detection of COPD is a crucial element for “would have the largest impact on patient treatment, positively altering the course of disease progression since these patients will be considered to have no/little through various therapeutic options” [25]. cardiac risk”, occurred only in 3 % of the cases. In the course of the Covid-19 pandemic two papers Particular features of AI-Rad Companion Chest CT by Weikert et al. [28] and Homayounieh et al. [29] inves- were also studied by Savage et al. [26], correlating the tigated the use of AI-Rad Companion Chest CT features average HU-density of the vertebrae computed by the for the prediction of patient management and patient software with T-scores obtained from dual-energy X-ray outcome in COVID-19 patients: Homayounieh et al. [29] absorptiometry (DEXA) on N= 65 patients yielding used a combination of lung lobe volumes, PO and PHO significant difference between healthy and osteoporotic yielding a “higher AUC for predicting ICU admission than (i.e. T < -2.5) patients. This is supported by work subjective severity scores” (N= 241). Weikert et al. [28] by Cohen et al. [8] using manual HU-measurements. added also cardiovascular metrics obtained from AI-Rad The authors found that a threshold of 110 HU Companion Chest CT, namely heart volume, coronary could be used to identify osteoporotic patients with calcium volume and aortic diameters, together with a specificity of 93 %. lab-findings yielding excellent predictions (AUC = 0.91, Average reading time (sec) 900 800 700 600 500 400 300 200 100 0 Lung nodules in complex cases by Abadia et al. [29] Aorta Follow-up by Rückel et al. [26] unaided AI-aided Figure 5: Average reading times with and without support of AI-Rad Companion Chest CT. 14 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper N= 120). In the work of Biebau et al [31], visual scores • Opacity quantification: Opaque regions were of lung injury were compared against AI-based scoring detected with a sensitivity of 89.4 % at 0.544 average of the LSS on N= 182 consecutive Covid-19 patients false positives per case. Correlation coefficient yielding a very good correlation of 0.89. for PO was 0.945. 95 %-Limits of agreement (LoA) of manual measurements of PO per lobe by two Efficiency and standardization radiologists was established at 15.8 %. Ratio of Increasing efficiency of the radiological workflow automatic PO measurements lying within the LoA is key to manage increasing workload and at the same was 93.0 %. N= 149. time saving healthcare cost. In the aforementioned study by Abadia et al. [30] on patients with complex • Heart segmentation: Average DICE coefficient lung disease average reading time for lung nodules was was 0.93. N= 274. 2:44 min ± 0:54 min without support of AI-Rad • Companion Chest CT. After a month of washout-period a Coronary calcium detection: Logarithmic random subset of N= 20 patients of the original study correlation coefficient of total coronary calcium were reevaluated with support of AI-Rad Companion volume was 0.96. N= 381. Chest CT. Here average reading time was reduced to • 0:36 min, i.e. a significant reduction by 78 %. Moreover, Aorta diameters: Average absolute error in aorta “the expert reported increased confidence for lung diameters was 1.6 mm across all nine measurement nodule detection for all 20 cases” [30]. locations and varied between 0.9 mm and 2.4 mm per location. N= 193. Average reading time was also in the focus in a study by Rückel et al. [27] on N= 18 patients with aortic ectasia • Vertebra HU-density: 95 %-Limits of agreement (LoA) undergoing follow-up assessments (two timepoints of manual density measurements by four radiologists per patient). Reading of the two time-points studies was established at 64.1 HU. Ratio of automatic was performed by three radiologists with and without vertebra density measurements lying within the LoA support of AI-Rad Companion Chest CT. Average was 98.8 %. N= 150. reading time was reduced from 13:01 min to 4:46 min • corresponding to a significant reduction by 63 %. In addi- Vertebrae heights: LoA of manual height measure- tion, AI assistance reduced total diameter inter-reader ments by four radiologists was established at 2.86 mm variability by 42.5 %. Figure 5 summarizes time savings for slice thickness ≤ 1.0 mm, and at 3.20 mm for reported by the various studies. slice thickness > 1.0 mm, respectively. Ratio of auto- matic vertebra height measurements lying within Standalone performance the LoA was 95.1 % for slice thickness ≤ 1.0 mm and 87.5 % for slice thickness > 1.0 mm. N= 150. Besides the validation of AI-Rad Companion Chest CT in studies performed by academic sites, internal standalone performance tests on the individual algorithms have been performed: References • Lung nodule detection: For nodule size range of 4 to 30 mm sensitivity was 93.1 % at 1 false positives 1 Guidelines for Management of Incidental Pulmonary per case (median), N= 316. Nodules Detected on CT Images: From the Fleischner Society 2017. MacMahon, H., Naidich, D. P., Goo, J. M., Lee, K. S., Leung, A., Mayo, J. R., Mehta, A. C., • Lung lobe segmentation: Average DICE coefficients for the individual lung lobes ranged between Ohno, Y., Powell, C. A., Prokop, M., Rubin, G. D., 0.95 and 0.98. Mean surface distance ranged between Schaefer-Prokop, C. M., Travis, W. D., Van Schil, 0.5 mm and 1.0 mm. Volume error was between P. E., & Bankier, A. A. 1, 2017, Radiology, Vol. 284, 1.5 % and 3.5 %. N= 4,500. pp. 228-243. Siemens Healthcare GmbH, 2021 15 Whitepaper · Features, Data, and Algorithms 2 2010 ACCF/AHA/AATS/ACR/ASA/SCA/SCAI/SIR/STS/SVM 9 Automated detection of lung nodules and coronary guidelines for the diagnosis and management of artery calcium using artificial intelligence on low- patients with Thoracic Aortic Disease. Hiratzka, L. F., dose CT scans for lung cancer screening: accuracy et al. 13, 2010, Circulation, Vol. 121, pp. 266-369. and prognostic value. Chamberlin, J., Kocher, M. R., 3 Managing Incidental Findings on Thoracic CT: Medias- Waltz, J., Snoddy, M., Stringer, N., Stephenson, J., tinal and Cardiovascular Findings. A White Paper of Sahbaee, P., Sharma, P., Rapaka, S., Schoepf, U. J., the ACR Incidental Findings Committee. Munden, R. F., Abadia, A. F., Sperl, J., Hoelzer, P., Mercer, M., Carter, B. W., Chiles, C., MacMahon, H., Black, W. C., Somayaji, N., Aquino, G., & Burt, J. R. 2021, BMC Ko, J. P., McAdams, H. P., Rossi, S. E., Leung, A. N., Medicine, Vol. 19, p. 55. Boiselle, P. M., Kent, M. S., Brown, K., Dyer, D. S., 10 MultiScale Deep Reinforcement Learning for Real- Hartman, T. E., Goodman, E. M., Naidich, D. P., Time 3D-Landmark Detection in CT Scans. Ghesu, F. Kazerooni, E. A., Berland, L. 8, 2018, J Am Coll C., Georgescu, B., Zheng, Y., Grbic, S., Maier, A., Radiol., Vol. 16, pp. 1097-1096. Hornegger, J., & Comaniciu, D. 1, 2019, IEEE Trans 4 2016 SCCT/STR guidelines for coronary artery calcium Pattern Anal Mach Intell., Vol. 41, pp. 176-189. scoring of noncontrast noncardiac chest CT scans: 11 Automatic Liver Segmentation Using an Adversarial A report of the Society of Cardiovascular Computed Image-to-Image Network. Yang, D., Xu, D., Zhou, S. K., Tomography and Society of Thoracic Radiology. Georgescu, B., Chen, M., Grbic, S., Metaxas, D., & Hecht, H. S., Cronin, P., Blaha, M. J., Budoff, M. J., Comaniciu, D. 2017. MICCAI. pp. 507-515. Kazerooni, E. A., Narula, J., Yankelevitz, D., & Abbara, 12 Quantifying the extent of emphysema: factors associ- S. 5, 2017, J Thorac Imaging, Vol. 32, pp. W54-W66. ated with radiologists' estimations and quantitative 5 Reporting incidental coronary, aortic valve and indices of emphysema severity using the ECLIPSE cardiac calcification on non-gated thoracic computed cohort. Gietema, H. A., Müller, N. L., Fauerbach, P. V., tomography, a consensus statement from the BSCI/ Sharma, S., Edwards, L. D., Camp, P. G., & Coxson, H. O. BSCCT and BSTI. Williams, M. C., Abbas, A., Tirr, E., 6, 2011, Academic radiology, Vol. 18, pp. 661-671. Alam, S., Nicol, E., Shambrook, J., Schmitt, M., 13 U-Net: Convolutional Networks for Biomedical Image Hughes, G. M., Stirrup, J., Holloway, B., Gopalan, D., Segmentation. Ronneberger, O., Fischer, P., & Brox, T. Deshpande, A., Weir-McCall, J., Agrawal, B., 2015. MICCAI. pp. 234-241. Rodrigues, J., Brady, A., Roditi, G., Robinson, G., & 14 Automated Quantification of CT Patterns Associated Bull, R. 1117, 2021, Br J Radiol, Vol. 94, p. 20200894. with COVID-19 from Chest CT. Chaganti, S., 6 Adams, J.E., Lenchik, L., Roux, C., & Genant, H. K. Balachandran, A., Chabin, G., Cohen, S., Flohr, T., Radiological Assessment of Vertebral Fracture. Inter- Georgescu, B., Grenier, P., Grbic, S., Liu, S., national Osteoporosis Foundation Vertebral Fracture Mellot, F., Murray, N., Nicolaou, S., Parker, W., Initiative Resource Document Part II. 2010, pp. 1-49. Re, T., Sanelli, P., Sauter, A. W., Xu, Z., Yoo, Y., 7 Opportunistic screening for osteoporosis using Ziebandt, V., & Comaniciu, D. 4, 2020, Radiology: abdominal computed tomography scans obtained Artificial Intelligence, Vol. 2. for other indications. Annals of internal medicine. 15 Chest CT Findings in Coronavirus Disease-19 Pickhardt, P. J., Pooler, B. D., Lauder, T., del Rio, A. M., (COVID-19): Relationship to Duration of Infection. Bruce, R. J., & Binkley, N. 8, 2013, Ann Intern Med, Bernheim, A., Mei, X., Huang, M., Yang, Y., Fayad, Z. A., Vol. 158, pp. 588-595. Zhang, N., Diao, K., Lin, B., Zhu, X., Li, K., Li, S., 8 Opportunistic screening for osteoporosis and Shan, H., Jacobi, A., & Chung, M. 3, 2020, Radiology, osteopenia by routine computed tomography scan: Vol. 295, p. 200463. A heterogeneous, multiethnic, middle-eastern 16 Lung cancer screening CT-based prediction of population validation study. Cohen, A., Foldes, A. J., cardiovascular events. Mets, O. M., Vliegenthart, R., Hiller, N., Simanovsky, N., & Szalat, A. 2021, Eur J Gondrie, M. J., Viergever, M. A., Oudkerk, M., Radiol, Vol. 136, p. 109568. de Koning, H. J., Mali, W. P., Prokop, M., van Klaveren, R. J., van der Graaf, Y., Buckens, C. F., Zanen, P., Lammers, J. W., Groen, H. J., Isgum, I., & de Jong, P. A. 8, 2013, JACC: Cardiovascular Imaging, Vol. 6, pp. 899-907. 16 Siemens Healthcare GmbH, 2021 Features, Data, and Algorithms · Whitepaper 17 Automatic coronary calcium scoring in chest CT using 25 Artificial Intelligence-based Fully Automated Per Lobe a deep neural network in direct comparison with non- Segmentation and Emphysema-quantification Based contrast cardiac CT: A validation study. van Assen, M., on Chest Computed Tomography Compared With Martin, S. S., Varga-Szemes, A., Rapaka, S., Cimen, S., Global Initiative for Chronic Obstructive Lung Disease Sharma, P., Sahbaee, P., De Cecco, C. N., Vliegenthart, Severity of Smokers. Fischer, A. M., Varga-Szemes, A., R., Leonard, T. J., Burt, J. R., & Schoepf, U. J. 2021, Martin, S. S., Sperl, J. I., Sahbaee, P., Neumann, D., Eur J Radiol, Vol. 134, p. 109428. Gawlitza, J., Henzler, T., Johnson, C. M., Nance, J. W., 18 Zhan, Y., Jian, B., Maneesh, D., & Zhou, X.S. Cross- Schoenberg, S. O., & Schoepf, U. J. Suppl 1, 2020, Modality Vertebrae Localization and Labeling Using J Thorac Imaging, Vol. 35, pp. S23-S34. Learning- Based Approaches. [ed.] S. Li and J. Yao. 26 Utilizing Artificial Intelligence to Determine Bone Spinal Imaging and Image Analysis. Lecture Notes in Mineral Density Via Chest Computed Tomography. Computational Vision and Biomechanics. s.l.: Savage, R. H., van Assen, M., Martin, S. S., Springer, 2015, Vol. 18. Sahbaee, P., Griffith, L. P., Giovagnoli, D., Sperl, J. I., 19 Vertebral fracture assessment using a semiquantita- Hopfgartner, C., Kärgel, R., & Schoepf, U. J. 2020 : tive technique. Genant, H. K., Wu, C. Y., van Kuijk, C., s.n., J Thorac Imaging, Vol. 35, pp. S35-S39. & Nevitt, M. C. 9, 1993, J Bone Miner Res., Vol. 8, 27 Artificial intelligence assistance improves reporting pp. 1137-48. efficiency of thoracic aortic aneurysm CT follow-up. 20 The impact of dose reduction on the quantification of Rueckel, J., Reidler, P., Fink, N., Sperl, J., Geyer, T., coronary artery calcifications and risk categorization: Fabritius, M. P., Ricke, J., Ingrisch, M., & Sabel, B. O. A systematic review. Vonder, M., van der Werf, N. R., 2021, Eur J Radiol, Vol. 134, p. 109424. Leiner, T., Greuter, M., Fleischmann, D., 28 Prediction of Patient Management in COVID-19 Using Vliegenthart, R., Oudkerk, M., & Willemink, M. J. Deep Learning-Based Fully Automated Extraction of 5, 2018, J Cardiovasc Comput Tomogr, Vol. 12, Cardiothoracic CT Metrics and Laboratory Findings. pp. 352-363. Weikert, T., Rapaka, S., Grbic, S., Re, T., Chaganti, S., 21 Siemens Healthineers. Agatston Score calcium Winkel, D. J., Anastasopoulos, C., Niemann, T., quantification with arbitrary tube voltage. 2018. Wiggli, B. J., Bremerich, J., Twerenbold, R., 22 Comparison of Artificial Intelligence-Based Fully Sommer, G., Comaniciu, D., & Sauter, A. W. 2021, Automatic Chest CT Emphysema Quantification to Korean J Radiol, Vol. 10.3348/kjr.2020.0994. Pulmonary Function Testing. Fischer, A. M., Varga- 29 Multicenter Assessment of CT Pneumonia Analysis Szemes, A., van Assen, M., Griffith, L. P., Sahbaee, P., Prototype for Predicting Disease Severity and Patient Sperl, J. I., Nance, J. W., & Schoepf, U. J. 5, 2020, Outcome. Homayounieh, F., Bezerra Cavalcanti Am J Roentgenol, Vol. 214, pp. 1065-1071. Rockenbach, M. A., Ebrahimian, S., Doda Khera, R., 23 Performance of an Artificial Intelligence-Based Bizzo, B. C., Buch, V., Babaei, R., Karimi Mobin, H., Platform Against Clinical Radiology Reports for the Mohseni, I., Mitschke, M., Zimmermann, M., Evaluation of Noncontrast Chest CT. Yacoub, B., Durlak, F., Rauch, F., Digumarthy, S. R., & Kalra, M. K. Kabakus, I. M., Schoepf, U. J., Giovagnoli, V. M., 2021, J Digit Imaging, Vols. doi: 10.1007/s10278- Fischer, A. M., Wichmann, J. L., Martinez, J. D., 021-00430-9. Sharma, P., Rapaka, S., Sahbaee, P., Hoelzer, P., 30 Diagnostic accuracy and performance of artificial Burt, J. R., Varga-Szemes, A., & Emrich, T. 21, 2021, intelligence in detecting lung nodules in patients with Acad. Radiol., Vols. S1076-6332, pp. 00070-2. complex lung disease: a non-inferiority study. 24 Reduction of missed thoracic findings in emergency Abadia, A.F. ACCEPTED IN JTI - IN PRESS - TO BE whole-body. Rueckel, J., Sperl, J., Kaestle, S., UPDATED. Hoppe, B., Fink, N., Rudolph, J., Schwarze, V., Geyer, 31 Comparing Visual Scoring of Lung Injury with T., Strobl, F., Ricke, J., Ingrisch, M., & Sabel, B. a Quantifying AI-Based Scoring in Patients with 6, 2021, Quantitative Imaging in Medicine and COVID-19. Biebau, C., Dubbeldam, A., Cockmartin, L., Surgery, Vol. 11, pp. 2486-2498. Coudyzer, W., Coolen, J., Verschakelen, J. & De Wever, W. 1, 2021, Journal of the Belgian Society of Radiology, Vol. 105, p. 16. Siemens Healthcare GmbH, 2021 17 AI-Rad Companion Chest CT is not commercially available in all countries, and its future availability cannot be ensured. The information in this document contains general tech- nical descriptions of specifications and options as well as standard and optional features which do not always have to be present in individual cases, and which may not be commercially available in all countries. Due to regulatory reasons their future availability cannot be guaranteed. Please contact your local Siemens organization for further details. Siemens reserves the right to modify the design, pack- aging, specifications, and options described herein without prior notice. Please contact your local Siemens sales representative for the most current information. Note: Any technical data contained in this document may vary within defined tolerances. Original images always lose a certain amount of detail when reproduced. Siemens Healthineers Headquarters Siemens Healthcare GmbH Henkestr. 127 91052 Erlangen, Germany Phone: +49 9131 84-0 siemens-healthineers.com Restricted Published by Siemens Healthcare GmbH, Germany · ©Siemens Healthcare GmbH, 2021