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

Features, Data, and Algorithms AI-Rad Companion Chest CT VA40A (VA22A) SIEMENS Healthineers Whitepaper · Features, Data, and Algorithms Table of contents Introduction …………………………………………………………… 3 Product features 4 …………………………………………………… Workflow ……………………………………………………………… 7 Algorithm description……………………………………………… 8 Data requirements …………………………………………………12 Proof points: performance and clinical value ……………14 References ……………………………………………………………18 2 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper Introduction AI-Rad Companion Chest CT is a decision support It focuses on three main parts of the thorax: the lungs tool for the radiological assessment of computed (AI-Rad Companion (Pulmonary)), the cardiovascular tomography (CT) images of the thorax. It helps radiolo- system (AI-Rad Companion (Cardiovascular)) and the gists interpret CT images of the thorax more quickly spine (AI-Rad Companion (Musculoskeletal)). and more precisely (doing more by doing less) and reduces the time needed to document the findings The typical workflow consists out of the following steps: with the help of automatic measurements. It is vendor- neutral, which means that the software can evaluate 1. Reconstructed CT images of the thorax are sent image data from any CT system manufacturer. Enabled to the PACS for interpretation by the teamplay digital health platform and using state-of-the-art image processing algorithms supported 2. In parallel, they are also sent to AI-Rad Companion. by artificial intelligence, AI-Rad Companion Chest CT Extensions are launched automatically delivers value in four key areas: 3. The results of AI-Rad Companion can either be sent 1. Accelerated interpretation and workflow efficiency to a web-based review software or directly to PACS. Here they can be used in combination with the original 2. Improved clinical outcomes and increased accuracy data for reporting purposes 3. Provision of additional clinically relevant information This whitepaper is intended to provide an overview of and visual highlighting the product features, describe the individual algorithmic components, discuss requirements for data to be 4. Standardized results while helping to reduce processed using the device, and to provide an overview inter-reader variability. of internal and external proof points assessing the performance and illustrating the clinical value of the application. LO AI-Rad Companion H Local Cloud Siemens Healthineers AG, 2024 3 Whitepaper · Features, Data, and Algorithms Product Features AI-Rad Companion Chest CT consists of three medical L5 (+15%) devices: AI-Rad Companion (Pulmonary), AI-Rad L2 (+20%) Companion (Cardiovascular) and AI-Rad Companion (Musculoskeletal). AI-Rad Companion (Pulmonary) The radiological assessment of pulmonary nodules L4 (+24%) is one of the most common indications for AI-Rad Companion Chest CT. The radiologist needs to identify the nodules, measure the diameters and – ideally – also their volume (see guidelines of the Fleischner L1 (+29%) Society [1]). A second important application of AI-Rad 12:9.9 mm Companion Chest CT is the analysis of lung parenchyma. 8.9 mm Reduced density can indicate emphysema and/or COPD while increased density can indicate inflammatory processes such as pneumonia. A AI-Rad Companion (Pulmonary) provides the following features with respect to the analysis of the lung: • Detection and segmentation of lung nodules like solid and sub-solid and localization with respect to lung lobes Correlation of detected nodule with known priors1 L2 • Left Lower Lobe and quantification of changes in size 2018-10-26 Current 2018-06-27 Prior • Analysis of the lung parenchyma based on segmented Max. 2D 0 [mm] 9.9 20% 8.3 lung lobes with respect to: Min 2D 0 [mm] 8.9 +8% 8.3 • areas of low attenuation Mean 20 0 [mm] 9.4 +14% 8.3 (low attenuation volume, or LAV) Max. 30 0 [mm] 11.3 +26% 9.0 areas of opacity Volume [mm3] 284.1 (200d) 186.9 • volume of lung lobes Slice number 115 135 • Figure 1: Outputs of the Pulmonary feature. Lung nodule detection, measurement, and correlation with prior. 2A 2B E F Figure 2: Outputs of the Pulmonary feature (cont’d). LAV-analysis (2A), opacity detection (2B)2. 1 AI-Rad Companion Chest CT will correlate each segmented lung nodule with known most recent prior (min. time difference > 10 days). 2 AI-Rad Companion Chest CT is not yet commercially available in all countries and its future availability cannot be ensured. 4 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper Exemplary outputs of AI-Rad Companion (Pulmonary) calcium clusters but does not perform Agatston are shown in Figure 1 and Figure 2, respectively. scoring which requires a gated scan and is the gold In the product, the LAV-Analysis is called “Lung standard for dedicated cardiac CT scans. Parenchyma Analysis”, while the opacity analysis is called “Pulmonary Density”. However, the importance of the analysis of both coronary calcium and aorta in the context of chest CT AI-Rad Companion (Cardiovascular) is to be pointed out. Both features are listed in the For the radiological assessment of the cardiovascular recommendations by the ACR Incidental Findings system a large variety of dedicated CT scan protocols Committee [3].The 2016 SCCT/STR guidelines [4] state exist depending on the clinical indication. The protocols that coronary artery calcium “should be evaluated and differ mainly with respect to the cardiac phase in which reported on all non-contrast chest CT examinations”. the acquisition is performed (controlled via ECG-gating) Analogously, in a consensus statement the British soci- and the type and timing of contrast enhancement. eties BSCi/BSCCT and BSTI [5] “recommend that AI-Rad Companion (Cardiovascular) is designed to work coronary artery calcification is reported on all non-gated with any of these protocols, particularly the most generic thoracic CT using a simple patient-based score (none, non-gated and non-contrast-enhanced Chest CT scans. mild, moderate, severe)”. In their 2010 guidelines [2] Of course, this also limits the analysis to features the ACCF/AHA states that “many thoracic aortic diseases, that can be reliably assessed on generic chest CT data. results of treatment for stable, often asymptomatic, The features are: 3A • Measurement of heart volume and quantification of coronary calcium volume (on unenhanced data only) • Segmentation of aorta and diameter measurements (on both native and contrast-enhanced data) • at 9 landmarks according to AHA-guidelines [2] • at the location of maximum diameter of the ascending and descending aorta, respectively Exemplary outputs of AI-Rad Companion (Cardiovascular) AL are shown in Figure 3. 3B Using AI-Rad Companion Chest CT at Diagnostikum Linz, Austria, impacts patient preparation, the actual execu- tion of the CT scan, and the reading and reporting of the chest CT cases. The radiographer also doesn’t need to reconstruct the CT angiography images in different planes or do the 3D reconstructions, which usually take about 15 to 20 minutes per examination. Following this workflow in reading and reporting results in a time gain of 50 percent because the number of mouse clicks is zero. Another benefit for the reader is that the method used to calculate the diameters is the same every time the patients get the follow-up CT examinations. [36] It is important to understand that the user should inter- AL pret the results of AI-Rad Companion (Cardiovascular) with respect to the actual scan protocol used. E.g., motion artifacts on a non-gated exam may hamper the accuracy of the aortic diameter measurements. Analogously, the coronary calcium analysis provides Figure 3: Outputs of the Cardiovascular device. Coronary calcium the total volume of the – potentially motion corrupted – detection (3A), aorta analysis (3B). Siemens Healthineers AG, 2024 5 Whitepaper · Features, Data, and Algorithms but high-risk conditions are far better than the results AI-Rad Companion undergone rigorous clinical validation of treatment required for acute and often catastrophic by both FDA and CE as verification of its efficacy and disease presentations. Thus, the identification and usability. The AD’s risk and diagnostic value was assessed treatment of patients at risk for acute and catastrophic in identifying hypertension in the general population, disease presentations prior to such an occurrence in identifying the poor blood pressure (BP) controlled are paramount to eliminating the high morbidity and in the hypertension population, and in screening masked mortality associated with acute presentations” and hence hypertension in the general population respectively motivates the automatic analysis of the thoracic aorta by multiple regression analysis and receiver operating on any chest CT. In the guidelines it is also described curve analysis. [39] that the thoracic aorta should be measured at nine predefined anatomical locations. “The use of standard- AI-Rad Companion (Musculoskeletal) ized measurements helps minimize errant reports of significant aneurysm growth due to technique or inter- Osteoporosis manifests as loss of bone density e.g. reader variability in measuring technique.” [2] in the spine, and consequently in vertebral compression fractures. The International Osteoporosis Foundation In Japan 315 of 344 patients were analyzed. All patients (IOF) states that “there is strong evidence of wide- underwent non-ECG-gated and ECG-gated non-enhanced spread under-diagnosis of vertebral fractures” [6]. CT during preoperative chest screening and/or chest Pickhardt et al. [7] and more recently Cohen et al. [8] pain assessment from March 7, 2021, to March 7, 2022. showed that the HU-values of the spine obtained The accuracy of coronary calcification volume (CV) AI from CT data acquired for other indications can be was compared with that of coronary calcification volume – used to identify osteoporotic patients and called this workstation (CV-WS) and the Agatston score (AS). approach “opportunistic screening for osteoporosis”. Stratification grades based on CV-AI were compared with grades based on the AS. Cases of mismatched AI-Rad Companion (Musculoskeletal) provides: • stratification were examined to determine the limitations Labeling and segmentation of thoracic vertebras • of AI. The cut-off value with the best stratification of Measurements of vertebrae heights • CV-AI was obtained. [37] Quantification of vertebral density (in HU) In a study in China 801 patients underwent both chest Exemplary outputs are shown in Figure 4. CT scan and arterial elasticity test were enrolled. Nine horizontal diameters of the thoracic aorta (from the aortic sinuses of Valsalva to the abdominal aorta T1 T2 |T3 |T4 T5 /16 /T7 at the celiac axis origin) were measured by AI using CT. IT1 15 / 15 / 18 / 181 Patients were divided into non-arterial stiffness (AS) T2 16 / 16/ 16 / 148 (mean value of the left and right cardio ankle vascular T3 /19/ 16/ 19 / 129 T4 19 / 15/ 20 / 142 index [M.CAVI]< 8), pre-AS (8 ≤ M.CAVI < 9), and 20/ 18/ 20 /131 AS (M.CAVI ≥ 9) groups. T6 19 0 194 23 3 / 110 Arterial diameters (AD) differences compared among 17 21 / 19 / 23 135 groups, analyzed the correlation of age, ADs, and M.CAVI T8 20/ 19 / 24 / 109 or the mean pressure independent CAVI (M.CAVI0), 9 21 / 20 / 23 /124 Furthermore, we evaluated the risk predictors and the OT10 24 / 23 / 24 / 119 diagnostic value of the nine ADs for pre-AS and AS. [38] T11 24 / 23 / 26 /107 AI-Rad Companion Cardiovascular (K222360 FDA cleared) T12 11/ 9/ 22 /191 was used to perform automatic aorta measurement in thoracic CT images at nine key positions based on AHA Figure 4: Output of the Musculoskeletal device: Height and density guidelines. Data was post processed by software from measurements of the thoracic vertebrae. 6 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper Workflow AI-Rad Companion Chest CT offers advanced ways Workflow integration samples: of efficient workflow integration and customization. 1. Efficiency gains are best accomplished when AI-Rad By design, all results are presented in the form of Companion is used to automate the repetitive an annotated axial series (MPR), a 3D rendering (VRT), and manually tedious task such as measurements. a concise summary table (DICOM SC), as well as Moreover, efficiency comes when these results overlay (DICOM 6000 for Lung Nodules and Aorta are deeply integrated in the respective workflow diameter results) – enabling integration into different by DICOM SR supporting advanced interoperability reading workflows. and control of results. Best advanced integration can be achieved by 2. Minimal disruption to the established workflow the provided DICOM Structured Report (DICOM SR). and unbiased reading is achieved when AI is used This supports: in a “spell checker” mode. Here, the reader would • Toggle display of AI annotations stick to their established reading patterns, but Delete individual AI annotations just before closing the case one last glance at the • results pictogram on the first page of the results • Edit individual AI annotations • Use AI results in worklist table (DICOM SC) allows for a quick and easy confir- Use AI results by feeding directly mation that indeed nothing was missed. The results • into radiology reporting table summarizes all findings and measurements. Comparison Pulmonary nodule No previous examination available for comparison. Findings 14,5 m Aorta: Mild aortic dilatation (max. ascending diameter 4.2cm). Thorax: 10.5 mm Lung lesion(s): 6 lung lesions identified by Al-Rad Companion, as follows. CAD Annotation bestätigen Lesion L1 (Series #5, Image #53): Right upper lobe, Maximum and orthogonal diameters CAD Annotation ablehnen (axial) 10.2mm x 6.8mm, Mean diameter (axial) 8.5mm. Volume 267mm3. CAD Verifikationsstatus zurück In den Vordergrund Lesion L2 (Series #5, Image #35): Right upper lobe, Maximum and orthogonal diameters (axial) 9.7mm x 7.7mm, Mean diameter (axial) 8.7mm. Volume 259mm3. In den Hintergrund Lesion L3 (Series #5, Image #52): Right upper lobe, Maximum and orthogonal diameters (axial) 8.7mm x 7.7mm, Mean diameter (axial) 8.2mm. Volume 243mm3. Lesion L4 (Series #5, Image #104): Left lower lobe, Maximum and orthogonal diameters (axial) 8.4mm x 6.3mm, Mean diameter (axial) 7.3mm. Volume 159mm3. Example 1: Reading Workflow integration Example 2: Cancer Management Software Example 3: Radiology Report integration in Visus JiveX PACS (v5.4) based on DICOM integration – Screening Plus. by middleware support. SR outputs of AI-Rad Companion. O SECTRUA Example 4: Editability and adjudication of Example 5: Fully automatic processing per Example 6: DICOM 6000 overlay – AI results embedded in Sectra PACS + direct DICOM SC in Visus JiveX (v.5.2). Toggle On/Off. transfer to report. Siemens Healthineers AG, 2024 7 Whitepaper · Features, Data, and Algorithms A color-coding scheme is used to draw the attention axials with the corresponding original series as to potential abnormalities. Added 3D renderings well as actively usage of the provided overlay (DICOM quickly provide a presentation overview of the type, 6000). As the reader scrolls through the stack, through number, and spatial context of all findings. Upon highlighting findings their attention is drown to confirmation of the findings, the results are straight- potential abnormalities. At the same forwardly transferred to the report. time, the correctness of the AI results is easily verified through comparison with unannotated series 3. Results of the AI-Rad Companion are best incorporated supported by the toggle on/off capabilities of the into the primary read by synchronizing annotated overlays (lung nodules and aorta measurements). Algorithm Description Lung Nodule Detection (Lung CAD) and Segmentation Classification utilizes a CNN-based classifier to process Lung CAD software is a tool optimized to detect pulmo- each candidate. The classifier calculates the feature nary nodules with the following average diameters. values for each candidate and uses a soft-max function to estimate the likelihood of its type as either “nodule” Thin Slice Data or “non-nodule.” Candidates meeting or exceeding a • Solid Pulmonary Nodules (SPN) likelihood value (a final confidence score above or equal between 3 mm and 30 mm to an empirically determined threshold) are labeled as nodule candidates. • Part Solid Nodules (PSN) between 5 mm and 30 mm • Ground Glass Nodules (GGN) The Post-Filtering step includes the application of two between 5 mm and 30 mm cascaded filters. The first one aims at removing false positives originating (a) from the colon and a second The CAD findings are not labeled relative to the nodule one from (b) calcified protrusions (for example, areas type. The user should make the determination of the where the sternum meets the manubrium, spine malfor- nodule type. mations, and osteophytes, and so on). The first filter is a CNN-based classifier that has a similar structure to Lung CAD processing is performed in several con- that of the classifier in the Classification step. The second secutive steps: Preprocessing, Candidate Generation, filter uses three orthogonal slices at the candidate Classification, and Post-Filtering. location as input to three CNN-based classifiers (one per slice). The results from the three classifiers are then In the Preprocessing step the input image is standard- combined by a max-voting mechanism. Any candidate ized, and parenchyma is segmented using specialized deemed a false positive by either filter is thus removed. Convolutional Neural Net (CNN, V-Net). This allows restricting the detection of findings within the lungs The algorithms have been trained using more than while optimizing the computation time. 2000 manually curated CT data sets. Network layout diagrams have been published by Chamberlin et al. [9]. Candidate Generation aims at achieving high sensitivity while keeping the number of candidates to a manage- After detection, nodules are segmented by an algorithm able number. The isotropic volume is partitioned into based on region growing. Diameter and volume measure- sub-volumes that are processed using a CNN. Then, fil- ments are provided. tering and non-maximum suppression yield a list of candidates for each sub-volume. A predefined threshold Lung Nodule Follow-up is applied on the confidence score to remove the least The lung nodule follow-up feature correlates nodules confidence (low score) findings. All candidates exceeding detected in the current scan (data1 at time point T1) this threshold are passed on to the Classification step. with nodules detected in a previous Chest CT exam of the same patient (data0 at time point T0) and calculates temporal size changes3. 3 AI-Rad Companion Chest CT will correlate each segmented lung nodule with known most recent prior (min. time difference > 10 days). 8 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper The algorithms used to establish the correlation In the RECIST guidelines [11] a decrease in diameter by has been designed based on the following assumptions at least 30% is defined as (partial) response to treatment, and requirements: while an increase by at least 20% is defined as progres- • The algorithm shall be executed at T1, i.e., when sive disease. If none of the criteria are met (i.e., diameter processing data1, and assumes that the previous data change within the interval (-30%, 20%)) the disease is set data0 has already been processed by the software considered stable. In analogy to these ranges, diameter at an earlier point in time measurements are highlighted by colors as follows: • AI-Rad Companion (Pulmonary) has been designed as a cloud-based and on-edge product. Thus, it should be minimalistic in terms of the amount of data from Category Condition T0 that must be available at T1, i.e., when processing data1. In particular it should not require that the full green dr ≤ -30% original data set data0 is available at T1 • Identification of lesion pairs should be independent yellow -30% < dr < 20% of changes in lesion size or shape over time and should red dr ≥ 20% rely on the lesion location as a robust feature instead in order to prevent a bias towards matching nodules of Table 1: Thresholds for diameter ratios dr. the same size or shape. Based on these requirements the algorithm works as follows: Lung Lobe Segmentation 1. At T0, multi-scale deep reinforcement learning algo- rithm [10] is used to extract anatomical landmarks The lung lobe segmentation algorithm computes from data0, and an AI-algorithm performs segmen- segmentation masks of the five lung lobes for a given tation of the lung lobes (see Section on Lung Lobe CT data set of the chest. Segmentation below). The list of landmarks, a mesh grid of the lung lobes, and the detected lesion coor- First, multi-scale deep reinforcement learning [10] dinates and size properties (diameters and volume) is used to robustly detect anatomical landmarks in a CT are stored in a secure and encrypted long-term volume. The carina bifurcation and/or sternum tip are storage (LTS). used to identify the lung region of interest (ROI). Next, the lung ROI image is resampled to a 2 mm isotropic 2. At T1, the LTS is queried for a prior exam of the volume and fed into an adversarialdeep Image-to-Image same patient and if available the stored data is Network (DI2IN) [12] to generate the lung segmentation. loaded. Landmarks and lung lobe meshes from T0 Finally, the ROI segmentation mask is remapped to are coregistered with the anatomical landmarks have the same dimension and the resolution as the input and lung lobes meshes extracted from data1 using data. The DI2IN has been first trained on over 8,000 CT an affine registration, resulting in a coordinate scans from a large group of patients with various transform F. The affine registration is capable of diseases, then fine-tuned with over 1,000 scans with handling shift, rotation, scale and shear, in partic- abnormal patterns including interstitial lung disease ular also capturing different breathing states of (ILD), pneumonia, and COVID-19. the two exams. The volume of the individual lung lobes, the left and 3. The coordinate transform F is used to map the T0- right lung and of the complete lung are reported. nodules into the T1-coordinate system. Distances between potential nodule pairs are computed LAV Analysis in the T1-coordinate system and a 3D distance The LAV analysis is threshold-based, i.e. the algorithm threshold is used to identify nodule pairs. determines all voxels below -950 Hounsfield Units (HU) in the lung. The threshold of -950 HU is widely used 4. For each identified pair the temporal size changes for the quantification of emphysema [13]. For each lung are computed. For the diameter-based metrics lobe as well as for the complete lung (i.e. the combina- the size change is expressed in terms of percentage tion of all lobes) the ratio of LAV (LAV%) is reported. change, i.e., as diameter ratio (dr). If the nodule The following thresholds are being used as default values volume has increased the change is expressed as for highlighting: volume doubling time. Siemens Healthineers AG, 2024 9 Whitepaper · Features, Data, and Algorithms Coronary Calcium Detection Category Condition Using the heart mask as ROI an initial set of voxels as I LAV% < 12.5% candidates for potentially calcified regions is obtained by thresholding at 130 HU. For each candidate voxel II 12.5% ≤ LAV% < 25% an image patch centered around the voxel is fed into a deep learning-based classification algorithm. The deep III 25% ≤ LAV% < 37.5% learning model has two components: a convolutional neural network, which takes the image patch and IV LAV% ≥ 37.5% a precomputed coronary territory map as inputs, and a dense neural network which operates on the coordi- Opacity Detection and Quantification nates of the voxel. A final prediction is made by combining features from both components to determine The detection and quantification of opaque regions whether the voxel belongs to the coronary arteries. in the lung – typically associated to viral pneumonia such as Covid-19 – uses a DenseUNet [14] with The algorithm has been trained on over 1,200 ECG- anisotropic kernels. Details of the algorithm are gated calcium scoring scans and finetuned on over described by Chaganti et al. [15]. The algorithm has 550 chest CTs. Additional details on the computational been trained on over 900 CT scans from patients pipeline and the network topology have been described with ILDs, pneumonia, and COVID-19. by Chamberlin et al. [9]. The total volume V of the detected coronary calcium is used for threshold-based The detected opacities are quantified by computing the categorization. Several thresholds for total calcium percentage of opacity (PO, per lobe and per lung) and volume have been proposed in the literature. For the percentage of high opacities (PHO, by applying instance, based on the NELSON study, Mets et al. [17] a threshold of -200 HU on the subset of opaque regions). showed that a coronary calcium volume of 100 mm4 Based on the PO a lung severity score (LSS) is calculated corresponds to an 8% increased risk of cardiovascular according to Bernheim et al. [16]: events and 500 mm4 to an increased risk of 48%. These volumes were used as default thresholds in AI-Rad Companion Chest CT.4 LSS Condition 0 PO = 0 Category Thresholds derived Thresholds used by from Mets et al. [17] van Assen et al. [18] 1 0 < PO ≤ 25% I V < 10 V < 5 2 25% < PO ≤ 50% II 10 ≤ V < 100 5 ≤ V < 250 3 50% < PO ≤ 75% III 100 ≤ V < 500 250 ≤ V < 1000 4 PO > 75% IV V ≥ 500 V ≥ 1000 Table 3: Thresholds for P0. Table 4: Thresholds for coronary calcium volume V in mm4. A total LSS is computed as the sum of the individual scores per lobe. A third threshold at 10 mm is used to compensate for image noise. Slightly different values have been used Heart Segmentation in a publication by van Assen et al [18]. An overview is provided in Table 4. The heart segmentation is performed using a deep U-shaped network [14] consisting of four convolutions Aorta Diameter Measurements and down-sampling steps, followed by four similar up-sampling layers. It has been trained on over 650 CT The aorta analysis pipeline consists of three steps: data sets. Subsequently, the heart segmentation mask landmark detection, aorta segmentation, and diameter is used to compute the heart volume. measurements. 4 In the software version available in the United States no default values are provided. 10 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper Six aortic landmarks (Aortic Root, Aortic Arch Center, Brachiocephalic Artery Bifurcation, Left Common Carotid Category Condition Artery, Left Subclavian Artery, and Celiac Trunk) are detected automatically based on Deep Reinforcement I d ≤ mean + 2*std Learning [10]. II d > mean + 2*std The aortic root is used to define a ROI for the segmen- III d > 1.5*mean tation algorithm. Within the ROI the segmentation is performed using an adversarial DI2IN in a symmetric IV d ≥ 5.5 cm convolutional encoder-decoder architecture [12]. The front part is a convolutional encoder-decoder Table 5: Thresholds for aortic diameters d. network with feature concatenation, and the backend is deep supervision network through multi-level. The aorta module works for both native and contrast- Blocks inside DI2IN consist of convolutional and enhanced data with and without ECG-gating. upscaling layers. The algorithm has been trained on over 1,250 CT data sets including native and contrast-enhanced Vertebra Labeling and Density Measurement scans. Given the aorta mask, a centerline model The twelve thoracic vertebrae are localized and labeled is used to generate the aortic centerline. The centerline using an algorithm based on wavelet features, AdaBoost, is used in combination with aortic landmarks to identify and local geometry constraints [19]. Around each measurement planes at nine locations according vertebra center cylindric regions of interest are used to to the guidelines of the American Heart Association [2]. measure the average HU-density of the trabecular bone. The measurement planes for the maximum diameter Vertebra Segmentation and Height Measurement of the ascending and descending aorta are identified by computing the area of each cross-sectional The vertebra centers are also used to determine plane along the centerline in the respective ranges. ROIs for the vertebra segmentation. Within the ROI For maximum diameter of the ascending aorta this the segmentation is performed using a DI2IN range is defined as Sinotubular Junction (AHA-location in a symmetric convolutional encoder-decoder archi- #2) to Proximal Aortic Arch (AHA-location #4). For tecture [12]. The algorithm has been trained on over maximum diameter of the descending aorta the range 7,300 thoracic vertebrae. is defined as Proximal Descending Aorta (AHA-location #6) to the end of the field of view. The centerline From the segmentation masks the sagittal midplane location of the maximum in-plane area is chosen as is extracted and within this plane height measurements the location of the maximum diameter of the ascending at anterior, medial, and posterior location. Afterwards, and descending aorta, respectively. the height ratios hr are computed by comparing heights of neighboring vertebrae using the Genant severity In each of the eleven measurement planes (nine planes grading method [20]. Although originally developed on according to AHA-locations plus the two locations chest radiographs, the Genant method is a widely used at maximum diameter of the ascending and descending also in CT imaging [6]: aorta, respectively), multiple diameters are computed by computing intersections of rays starting from Category Condition the centerline with the aortic mask. Based on these diameters, the maximum in-plane diameter is reported. I hr ≥ 80% The maximum diameters d are used for threshold-based II 80% > hr ≥ 75% categorization. Recommended thresholds for the severity of a potential aortic dilation or aneurysm have III 75% > hr ≥ 60% been derived from the AHA-Guidelines and values for the population mean and standard variation (std) IV hr < 60% given therein5: Table 6: Thresholds for vertebra ratios hr. 5 In the software version available in the United States the thresholds cannot be adapted by the user. Siemens Healthineers AG, 2024 11 Whitepaper · Features, Data, and Algorithms Data Requirements Technical Requirements Scan Parameter Recommendations AI-Rad Companion Chest CT uses a single DICOM series Besides the coronary calcium detection, HU-thresholding as input for all modules. In general, the algorithms is also used in the LAV-analysis of the lung parenchyma. are intended to work with any chest CT series. However, As a consequence, the results of these two features are there are a couple of technical properties required for sensitive to image noise. Image noise in CT data depends the device to process the cases: on many parameters, most prominently on slice thick- ness, reconstruction kernel, and dose. Hence the • Primary axial images (image orientation 1\0\0\0\1\0) combination of thin slices, hard kernels, and low dose • Volume scans without gaps, no gantry tilt may result in very noisy images. For such data the • Slice thickness ≤ 3 mm (for MSK ≤ 2 mm, preferably cardiac module would reject the case (if there are too ≤ 1 mm or below), see recommendations below many calcium candidates) and the LAV analysis may be • Matrix size 512 × 512 confounded by noise-related LAV-patches [23]. • Photometric interpretation: MONOCHROME 2 • 16 bit, no lossy compression, samples per pixel: 1 On the other hand, thin slices, i.e. high spatial resolution • Rescale slope ≤ 5 in z-direction, are required for most of the algorithms, in particular for accurate vertebrae height measurements The cardiovascular module (heart segmentation (ideally slice thickness should be ≤ 1 mm), detailed delin- and coronary calcium detection) has the additional eation of lung lobes and accurate lung nodule volumetry. requirements that the images are without contrast enhancement and kVp ≥ 100. That is because the initial In summary, Table 7 displays the recommendations candidate generation step is based on HU-thresholding of scan parameters for the individual modules of AI-Rad and the threshold is not valid for contrast-enhanced Companion Chest CT. To achieve optimal results for scans nor for kVp < 100. The topic has been discussed all modules, it is recommended to use a thin slice with in detail by Vonder et al. [21] and in a corresponding a soft to medium kernel. In addition, Table 8 summarizes Siemens Healthineers Whitepaper on calcium quanti- scan parameters used in various clinical studies using fication on dedicated cardiac CT data [22]. AI-Rad Companion Chest CT. Details about these and other studies will be discussed in the subsequent section. 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 7: Recommended scan parameters for AI-Rad Companion Chest CT. fully supported supported but results might be suboptimal not supported 12 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper Publication Patient cohort Feature(s) studies Study size Scanner model(s) Scan parameters Chamberlin Lung cancer Lung nodules, SOMATOM go.Top, protocol according et al. [9] screening cor. calcium 117 Definition AS+, to ACR-STR LDCT guidelines. Definition Flash, and Force slice thickness: 1.0 mm Paired Cardiac and et al. [18] Chest CTs, consec- Cor. calcium 95 SOMATOM Definition Flash, slice thickness: van Assen utive Chest CTs + 168 Definition AS+, and Force 1.0 mm – 3.0 mm, medium sharp kernel Fischer slice thickness: 1.5 mm, et al. [23] Emphysema SOMATOM Definition Flash, LAV 141 Force, and Emotion comparing two kernels: lung (B60s) and soft tissue (B31s) Yacoub et al. [24] Consecutive cases all 100 SOMATOM Definition Flash, slice thickness: 1.0 mm, and Force soft tissue kernel Rückel Lung nodules, Emergency CT aorta diam., cor. calcium, 105 SOMATOM Force slice thickness: 0.75 mm, et al. [25] heart size, vert. heights soft tissue kerne Br36d Fischer COPD Lung lobes, LAV 137 SOMATOM Definition Flash, slice thickness: 1.5 mm, et al. [26] Force, and Emotion lung kernel SOMATOM Definition Flash, Rückel Aortic aneurysm et al. [28] follow-up Aorta diam. 18 × 2 Force, and Definition AS+, slice thickness: GE Optima CT660, 0.6 mm – 3.0 mm, Discovery 750 HD soft tissue kernel Weikert Lobe volume, PO, PHO, et al. [29] COVID-19 patients LSS, LAV, heart size, 120 SOMATOM Definition AS+ slice thickness: 1.0 mm, cor. calcium, aorta diam soft tissue kernel SOMATOM Definition Flash, Homayounieh Lobe volume, Force, and Definition Edge, slice thickness: et al. [30] COVID-19 patients PO, PHO 241 Emotion 16, 1.0 mm–2.0 mm, GE Discovery 750 HD soft tissue kernel B20f Abadia Lung nodules in et al. [31] cases w/ complex Lung lobes, 103 slice thickness: 1.0 mm, lung disease Lung nodules + 40 SOMATOM Force sharp body kernel GE: ASIR at 40% Detail kernel Philips: iDose 4, strength 3, Ebrahimian GE Discovery 750 HD, kernel B, et al. [32] Emphysema LAV 113 Philips iCT, SOMATOM Definition Edge Siemens: Admire strength 2, I31f, slice thickness: 0.625 mm – 1.25 mm Table 8: Scan parameters used in various publications using AI-Rad Companion Chest CT. Siemens Healthineers AG, 2024 13 Whitepaper · Features, Data, and Algorithms Considerations Regarding Patient Population it is also important to note that the output images Table 8 also illustrates that AI-Rad Companion generated by AI-Rad Companion Chest CT are designed Chest CT has been used to analyze a broad spectrum in a way that the user can easily verify the correctness of patient cohorts: of the result. An example would be the sagittal MPR • low dose lung cancer screening [9], of the spine, see Figure 5. • consecutive cohorts, independent of particular clinical indications [24; 18] or with an indication unrelated to the features of AI-Rad Companion Chest CT like data /156 from the emergency department [25], 13/ 17 / T2 14 / 13 / 17 • patients with known disease patterns relevant for T3 16 / 14 / / 122 the feature of investigation, like emphysema/COPD T4 / 15/ 18 / 136 [26; 23], osteoporosis [27], aortic aneurysms [28] T5 18 / 17 / 18 / 110 COVID-19 [29; 30], or T6 18/ 15/ 19 / 120 patients with known comorbidities that make the 17 16 / 18/ 19 / 116 • assessment of the feature under investigation more T8 19 / 16/ 20 / 120 challenging, such as the detection of lung nodules 19 16 / 16 / 19 / 125 in the presence of, e.g. ILD [31]. T10 21 / 18 / 20 / 120 The broad spectrum illustrates the versatile and T11 23/ 19/ 24 /132 generic design of the algorithms of AI-Rad Companion T12 24 / 20 / 25 / 69 Chest CT. On the other hand, one would always find cases where – due to severe pathology, comorbidity, or anatomical deviation, but also due to imaging arte- Figure 5: AI-Rad Companion (Musculoskeletal) output: Sagittal facts like motion or noise – one or more algorithms view of the spine including height and density measurement. might fail or produce incorrect result. In that context Proof Points: Performance and Clinical Value AI-Rad Companion Chest CT delivers value in four coronary calcium volume: 0.904) at a sensitivity of 100% main categories efficiency, accuracy, additional clinically and 92.9% (presence of lung nodules and presence relevant information, and standardization. These of coronary calcifications, respectively) and a specificity improvements of the radiologist’s daily work need of 70.8% and 96.0%, respectively. The authors also to be interpreted within the context of particular clinical use the results for predicting of lung cancer and major use cases. Moreover, the foundation for the improve- adverse cardiac events at 1-year follow-up yielding ments in all four categories lies in an excellent algorithm AUC-values of 0.942 and 0.911, respectively, empha- performance. Hence also the scientific evidence, sizing that “this information can be used to improve in terms of peer-reviewed journal publications but also diagnostic ability, facilitate intervention, improve mor- internal performance tests, clusters around Accuracy bidity and mortality, and decrease healthcare costs”. and Clinical Value for a particular use case, Efficiency and Standardization, and Standalone performance Focusing on the other end of the spectrum of patient of the individual algorithm components. cohorts, namely patients with complex lung disease such as ILD, COPD, bronchitis, edema, and pulmonary Accuracy and Clinical Value embolism, Abadia et al. [31] investigated the accuracy In the study by Chamberlin et al. [9] N = 117 lung cancer of the lung nodule detection and localization screening exams were processed by AI-Rad Companion (N = 103 plus 40 controls). On a patient level AI-Rad Chest CT and analyzed with respect to lung nodules and Companion Chest CT showed a sensitivity of 89.4% and coronary calcium. The agreement with expert reader a specificity of 82.5%. On the individual nodule level has been found excellent (Cohen’s kappa of lung nodule sensitivity was 67.7%, similar to the accuracy reported detection: 0.846, intraclass correlation coefficient of for experienced radiologists. 14 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper 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. [24] 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 9. applications. But also, with regard to less urgent appli- cations of medical imaging, it should also be mentioned that especially non-radiology clinicians might even take N Sensitivity Specificity more benefit from AI-assisted image analysis compared positive to anyway well-trained radiologists, e.g., in clinical cases AI Report AI Report settings without 24/7 radiology coverage or long turn- around times for radiology reporting.” Lung nodules 83 92.8% 97.6% 82.4% 100% Consecutive patients (N = 168) were also enrolled in a study on coronary calcium detection by Emphysema 31 80.6% 74.2% 66.7% 97.1% van Assen et al. [18]. Here the coronary calcium volume computed by AI-Rad Companion Chest CT Aortic was compared against the calcium volume obtained dilation 27 96.3% 25.9% 81.4% 100% from manual calcium scoring. The correlation was found excellent (logarithmic correlation coefficient Coronary 94.9% 0.923). By applying volume-thresholds (see Table 4) calcium 59 89.8% 75.4% 100% the I-computed calcium volume was categorized into Vertebra no, mild, moderate, and severe. compression 9 100% 100% 63.7% 100% The categories were compared against qualitative visual Table 9: Sensitivity and specificity of AI-Rad Companion Chest CT rating by an experienced cardiac radiologist. Results are and radiological reports on N = 100 consecutive cases as reported shown in Table 10. 82% of all cases were correctly classi- by Yacoub et al. [24]. fied with all wrongly classified scans being attributed to an adjacent category. The authors state that “such results align well with the general recommendation to maximize sensitivity when AI is being used in radiology to detect abnormalities, Expert\AI No Mild Moderate Severe even at the expense of lower specificity, in order to ensure that fewer abnormal findings are missed. No 60 6 0 0 “Our findings illustrate that the use of AI for diagnostic Mild 7 44 0 0 reading provides rather a support tool which is Moderate 0 6 14 5 not intended to replace the role of a radiologist.” They conclude that “incorporating AI support into radiology Severe 0 0 4 20 workflows can provide significant added value to clinical radiology reporting”. Table 10: Category agreement between manual qualitative assessment and AI determined calcium volume as reported The low sensitivity of the radiologists in particular by van Assen et al. [18]. among the incidental findings has also been studied by Rückel et al. [25] in the particular time-critical In a second arm of the study, N = 95 patients were setting of emergency CT. The following abnormalities identi- fied which underwent both dedicated coronary were missing in the original reports but identified calcium scoring exams (non-contrasted, ECG-gated by AI-Rad Companion Chest CT and confirmed by radio- cardiac CTs) and chest CTs within 1.5 years. For those logists in a consecutive cohort of N = 105 whole-body patients, conventional calcium scoring was performed emergency CTs: according 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 agree- • 16.2% coronary calcifications, ment of these results will be lower, simply because the • 32.4% aortic ectasia, data compared originates from different acquisitions • 1.9% actionable lung nodules, and from different time points. Nevertheless, the correlation • 12.4% vertebra fractures between manual Agatston score and calcium volume Siemens Healthineers AG, 2024 15 Whitepaper · Features, Data, and Algorithms computed by AI-Rad Companion Chest CT was found predicting ICU admission than subjective severity scores” excellent (logarithmic correlation coefficient 0.921). (N = 241). Weikert et al. [29] added also cardiovascular When comparing threshold-based categories (volume metrics obtained from AI-Rad Companion Chest CT, threshold as in Table 4 vs. standard Agatston risk namely heart volume, coronary calcium volume, categories), 70% of all cases were classified correctly, and aortic diameters, together with lab-findings yielding in only 5% the prediction was more than one category excellent predictions (AUC = 0.91, N = 120). In the off. Moreover, a misclassification into the “no calcium” work of Biebau et al [33], visual scores of lung injury category, which – according to the authors [18] – were compared against AI-based scoring of the LSS “would have the largest impact on patient treatment, on N = 182 consecutive Covid-19 patients yielding a very since these patients will be considered to have no/little good correlation of 0.89. cardiac risk”, occurred only in 3% of the cases. Efficiency and Standardization Particular features of AI-Rad Companion Chest CT were also studied by Savage et al. [27], correlating the Increasing efficiency of the radiological workflow average HU-density of the vertebrae computed by is key to manage increasing workload and at the same the software with T-scores obtained from dual-energy time saving healthcare cost. In the aforementioned X-ray absorptiometry (DEXA) on N = 65 patients study by Abadia et al. [31] on patients with complex yielding significant difference between healthy and lung disease average reading time for lung nodules osteoporotic (i.e. T < -2.5) patients. This is supported was 2:44 min ± 0:54 min without support of AI-Rad by work by Cohen et al. [8] using manual HU-measure- Companion Chest CT. After a month of washout-period ments. The authors found that a threshold of 110 HU a random subset of N = 20 patients of the original study could be used to identify osteoporotic patients with were reevaluated with support of AI-Rad Companion a specificity of 93%. Chest CT. Here average reading time was reduced to 0:36 min, i.e. a significant reduction by 78%. Moreover, The routine chest CT with AI-Rad Companion is of great “the expert reported increased confidence for lung value in opportunistic screening for osteopenia or osteo- nodule detection for all 20 cases” [31]. porosis, which can quickly screen the population at high risk of osteoporosis without increasing radiation dose, The potential of AI-Rad Companion Chest CT to reduce thus reducing the incidence of osteoporotic fracture. [40] reading time has been evaluated in a study by Yacoub et al. [34]: In this prospective study, chest CT Two publications by Fischer et al. [26; 23] study reading times by three radiologists were assessed. the results of the lung lobe-based LAV analysis in N=390 consecutive CT scans were enrolled, and each emphysema/COPD patients (N = 141 and N = 137, reader was assigned an equal number of cases with respectively). The correlation of LAV with spirometry- and without AI-Rad Companion Chest CT results. Mean based Tiffeneau index was -0.86, and 0.88 with reading using AI-support was reduced by 92.9 sec GOLD stages, respectively. The LAV of the upper lobes (22.1%). Müller et al. [35] performed a prospective study “was also able to most clearly distinguish mild and with N=90 cases and two readers. Here no time saving moderate forms of COPD. This is particularly relevant was reported but additional actionable findings due to the fact that early disease processes often were found in 12.5% of the cases as well as qualitative elude conventional pulmonary function diagnostics. improvements: change of case impression (12.5%), Earlier detection of COPD is a crucial element for better case overview (55%) and increased diagnostic positively altering the course of disease progression confidence (20%). through various therapeutic options” [26]. Ebrahamian et al. [32] showed that the LAV-based quantification Average reading time was also in the focus in a study of emphysema is of similar quality than visual assess- by Rückel et al. [28] on N = 18 patients with aortic ment by radiologists (N=113). ectasia undergoing follow-up assessments (two time- points per patient). Reading of the two time-points In the course of the Covid-19 pandemic two papers studies was performed by three radiologists with and by Weikert et al. [29] and Homayounieh et al. [30] without support of AI-Rad Companion Chest CT. Average inves- tigated the use of AI-Rad Companion Chest CT reading time was reduced from 13:01 min to 4:46 min features for the prediction of patient management corresponding to a significant reduction by 63%. In addi- and patient outcome in COVID-19 patients: tion, AI assistance reduced total diameter inter-reader Homayounieh et al. [30] used a combination of lung variability by 42.5%. Figure 6 summarizes time savings lobe volumes, PO and PHO yielding a “higher AUC for reported by the various studies. 16 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper Average reading time (sec) 900 800 700 600 500 400 300 200 100 unaided 0 AI-aided Chest CT reading Lung nodules in complex Aorta Follow-up by Yacoub et al. [34] cases by Abadia et al. [31] by Rückel et al. [28] Figure 6: Average reading times with and without support of AI-Rad Companion Chest CT. Standalone Performance • Heart segmentation: Average DICE coefficient was Besides the validation of AI-Rad Companion Chest CT in 0.93. N = 274. studies performed by academic sites, internal standalone • performance tests on the individual algorithms have Coronary calcium detection: Logarithmic correlation been performed: coefficient of total coronary calcium volume was 0.96. N = 381. • Lung nodule detection: For nodule size range of • 4 to 30 mm sensitivity was 93.1% at 1 false positives Aorta diameters: Average absolute error in aorta per case (median), N = 316. diameters was 1.6 mm across all nine measurement locations and varied between 1.2 mm and 2.2 mm Lung nodule follow-up: Sensitivity of nodule per location. N=193. • matching: 94.3%, positive predictive value 99.1%, • Vertebra HU-density: 95%-Limits of agreement (LoA) N=199. of manual density measurements by four radiologists • Lung lobe segmentation: Average DICE coefficients was established at 64.1 HU. Ratio of automatic for the individual lung lobes ranged between 0.95 vertebra density measurements lying within the LoA and 0.98. Mean surface distance ranged between was 98.8%. N = 150. 0.5 mm and 1.0 mm. Volume error was between 1.5% • and 3.5%. N = 4,500. Vertebrae heights: LoA of manual height measure- ments by four radiologists was established at 2.86 mm • Opacity quantification: Opaque regions were for slice thickness ≤ 1.0 mm, and at 3.20 mm for detected with a sensitivity of 89.4% at 0.544 average slice thickness > 1.0 mm, respectively. Ratio of auto- false positives per case. Correlation coefficient for matic vertebra height measurements lying within PO was 0.945. 95%-Limits of agreement (LoA) of the LoA was 95.5% for slice thickness ≤ 1.0 mm and manual measurements of PO per lobe by two radiolo- 92.6% for slice thickness > 1.0 mm. N=150. gists was established at 15.8%. Ratio of automatic PO measurements lying within the LoA was 93.0%. N = 149. Siemens Healthineers AG, 2024 17 Whitepaper · Features, Data, and Algorithms References 1 Guidelines for Management of Incidental Pulmonary 6 Adams, J.E., Lenchik, L., Roux, C., & Genant, H. K. Nodules Detected on CT Images: From the Fleischner Radiological Assessment Of Vertebral Fracture. Inter- Society 2017. MacMahon, H., Naidich, D. P., national Osteoporosis Foundation Vertebral Fracture Goo, J. M., Lee, K. S., Leung, A., Mayo, J. R., Initiative Resource Document Part II. 2010, pp. 1-49. Mehta, A. C., Ohno, Y., Powell, C. A., Prokop, M., Rubin, G. D., Schaefer-Prokop, C. M., Travis, W. D., 7 Opportunistic screening for osteoporosis using Van Schil, P. E., & Bankier, A. A. abdominal computed tomography scans obtained 1, 2017, Radiology, Vol. 284, pp. 228-243. for other indications. Annals of internal medicine. Pickhardt, P. J., Pooler, B. D., Lauder, T., del Rio, A. M., 2 2010 ACCF/AHA/AATS/ACR/ASA/SCA/SCAI/SIR/STS/SVM Bruce, R. J., & Binkley, N. guidelines for the diagnosis and management of 8, 2013, Ann Intern Med, Vol. 158, pp. 588-595. patients with Thoracic Aortic Disease. Hiratzka, L. F., et al. 13, 2010, Circulation, Vol. 121, pp. 266-369. 8 Opportunistic screening for osteoporosis and osteopenia by routine computed tomography scan: 3 Managing Incidental Findings on Thoracic CT: A heterogeneous, multiethnic, middle-eastern popu- Mediastinal and Cardiovascular Findings. A White lation validation study. Cohen, A., Foldes, A. J., Hiller, Paper of the ACR Incidental Findings Committee. N., Simanovsky, N., & Szalat, A. Munden, R. F., Carter, B. W., Chiles, C., MacMahon, 2021, Eur J Radiol, Vol. 136, p. 109568. H., Black, W. C., Ko, J. P., McAdams, H. P., Rossi, S. E., Leung, A. N., Boiselle, P. M., Kent, M. S., Brown, K., 9 Automated detection of lung nodules and coronary Dyer, D. S., Hartman, T. E., Goodman, E. M., artery calcium using artificial intelligence on l Naidich, D. P., Kazerooni, E. A., Berland, L. ow- dose CT scans for lung cancer screening: 8, 2018, J Am Coll Radiol., Vol. 16, pp. 1097-1096. accuracy and prognostic value. Chamberlin, J., Kocher, M. R., Waltz, J., Snoddy, M., Stringer, N., 4 2016 SCCT/STR guidelines for coronary artery calcium Stephenson, J., Sahbaee, P., Sharma, P., Rapaka, S., scoring of noncontrast noncardiac chest CT scans: Schoepf, U. J., Abadia, A. F., Sperl, J., Hoelzer, P., A report of the Society of Cardiovascular Computed Mercer, M., Somayaji, N., Aquino, G., & Burt, J. R. Tomography and Society of Thoracic Radiology. 2021, BMC Med, Vol. 19, p. 55. Hecht, H. S., Cronin, P., Blaha, M. J., Budoff, M. J., Kazerooni, E. A., Narula, J., Yankelevitz, D., & Abbara, 10 MultiScale Deep Reinforcement Learning for S. 5, 2017, J Thorac Imaging, Vol. 32, pp. W54-W66. Real-Time 3D-Landmark Detection in CT Scans. Ghesu, F. C., Georgescu, B., Zheng, Y., Grbic, S., 5 Reporting incidental coronary, aortic valve and Maier, A., Hornegger, J., & Comaniciu, D. cardiac calcification on non-gated thoracic computed 1, 2019, IEEE Trans Pattern Anal Mach Intell., tomography, a consensus statement from the BSCI/ Vol. 41, pp. 176-189. BSCCT and BSTI. Williams, M. C., Abbas, A., Tirr, E., Alam, S., Nicol, E., Shambrook, J., Schmitt, M., Hughes, G. M., Stirrup, J., Holloway, B., Gopalan, D., Deshpande, A., Weir-McCall, J., Agrawal, B., Rodrigues, J., Brady, A., Roditi, G., Robinson, G., & Bull, R. 1117, 2021, Br J Radiol, Vol. 94, p. 20200894. 18 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper 11 New guidelines to evaluate the response to 17 Lung cancer screening CT-based prediction of treatment in solid tumors. European Organization cardio-vascular events. Mets, O. M., Vliegenthart, R., for Research and Treatment of Cancer, National Gondrie, M. J., Viergever, M. A., Oudkerk, M., de Cancer Institute of the United States, National Cancer Koning, H. J., Mali, W. P., Prokop, M., van Klaveren, R. Institute of Canada. Therasse, P., Arbuck, S. G., J., van der Graaf, Y., Buckens, C. F., Zanen, P., Eisenhauer, E. A., Wanders, J., Kaplan, R. S., Lammers, J. W., Groen, H. J., Isgum, I., & de Jong, P. A. Rubinstein, L., Verweij, J., Van Glabbeke, M., van 8, 2013, JACC Cardiovasc Imaging, Vol. 6, pp. 899-907. Oosterom, A. T., Christian, M. C., & Gwyther, S. G. 3, 2000, J Natl Cancer Inst, Vol. 92. 18 Automatic coronary calcium scoring in chest CT using a deep neural network in direct comparison 12 Automatic Liver Segmentation Using an with non-contrast cardiac CT: A validation study. Adversarial Image-to-Image Network. Yang, D., Xu, van Assen, M., Martin, S. S., Varga-Szemes, A., D., Zhou, S. K., Georgescu, B., Chen, M., Grbic, S., Rapaka, S., Cimen, S., Sharma, P., Sahbaee, P., Metaxas, D., & Comaniciu, D. De Cecco, C. N., Vliegenthart, R., Leonard, T. J., 2017. MICCAI. pp. 507-515. Burt, J. R., & Schoepf, U. J. 2021, Eur J Radiol, Vol. 134, p. 109428. 13 Quantifying the extent of emphysema: factors associated with radiologists' estimations and 19 Zhan, Y., Jian, B., Maneesh, D., & Zhou, X.S. quantitative indices of emphysema severity using Cross-Modality Vertebrae Localization and Labeling the ECLIPSE cohort. Gietema, H. A., Müller, N. L., Using Learning-Based Approaches. [ed.] S. Li and J. Fauerbach, P. V., Sharma, S., Edwards, L. D., Camp, P. Yao. Spinal Imaging and Image Analysis. Lecture G., & Coxson, H. O. Notes in Computational Vision and Biomechanics. 6, 2011, Acad Radiol, Vol. 18, pp. 661-671. s.l. : Springer, 2015, Vol. 18. 14 U-Net: Convolutional Networks for Biomedical Image 20 Vertebral fracture assessment using a semiquantita- Segmentation. Ronneberger, O., Fischer, P., & Brox, T. tive technique. Genant, H. K., Wu, C. Y., 2015. MICCAI. pp. 234-241. van Kuijk, C., & Nevitt, M. C. 9, 1993, J Bone Miner Res., Vol. 8, pp. 1137-48. 15 Automated Quantification of CT Patterns Associated with COVID-19 from Chest CT. Chaganti, S., 21 The impact of dose reduction on the quantification Balachandran, A., Chabin, G., Cohen, S., Flohr, T., of coronary artery calcifications and risk Georgescu, B., Grenier, P., Grbic, S., Liu, S., categorization: A systematic review. Vonder, M., Mellot, F., Murray, N., Nicolaou, S., Parker, W., van der Werf, N. R., Leiner, T., Greuter, M., Re, T., Sanelli, P., Sauter, A. W., Xu, Z., Yoo, Y., Fleischmann, D., Vliegenthart, R., Oudkerk, M., Ziebandt, V., & Comaniciu, D. & Willemink, M. J. 4, 2020, Radiol Artif Intell, Vol. 2, p. e200048. 5, 2018, J Cardiovasc Comput Tomogr, Vol. 12, pp. 352-363. 16 Chest CT Findings in Coronavirus Disease-19 (COVID-19): Relationship to Duration of Infection. 22 Siemens Healthineers. Agatston Score calcium Bernheim, A., Mei, X., Huang, M., Yang, Y., Fayad, Z. quantification with arbitrary tube voltage. 2018. A., Zhang, N., Diao, K., Lin, B., Zhu, X., Li, K., Li, S., Shan, H., Jacobi, A., & Chung, M. 3, 2020, Radiology, Vol. 295, p. 200463. Siemens Healthineers AG, 2024 19 Whitepaper · Features, Data, and Algorithms 23 Comparison of Artificial Intelligence-Based Fully 28 Artificial intelligence assistance improves reporting Automatic Chest CT Emphysema Quantification efficiency of thoracic aortic aneurysm CT follow-up. to Pulmonary Function Testing. Fischer, A. M., Rueckel, J., Reidler, P., Fink, N., Sperl, J., Geyer, T., Varga-Szemes, A., van Assen, M., Griffith, L. P., Fabritius, M. P., Ricke, J., Ingrisch, M., & Sabel, B. O. Sahbaee, P., Sperl, J. I., Nance, J. W., & Schoepf, U. J. 2021, Eur J Radiol, Vol. 134, p. 109424. 5, 2020, Am J Roentgenol, Vol. 214, pp. 1065-1071. 29 Prediction of Patient Management in COVID-19 Using 24 Performance of an Artificial Intelligence-Based Deep Learning-Based Fully Automated Extraction of Platform Against Clinical Radiology Reports for Cardiothoracic CT Metrics and Laboratory Findings. the Evaluation of Noncontrast Chest CT. Yacoub, B., Weikert, T., Rapaka, S., Grbic, S., Re, T., Chaganti, S., Kabakus, I. M., Schoepf, U. J., Giovagnoli, V. M., Winkel, D. J., Anastasopoulos, C., Niemann, T., Fischer, A. M., Wichmann, J. L., Martinez, J. D., Wiggli, B. J., Bremerich, J., Twerenbold, R., Sharma, P., Rapaka, S., Sahbaee, P., Hoelzer, P., Sommer, G., Comaniciu, D., & Sauter, A. W. Burt, J. R., Varga-Szemes, A., & Emrich, T. 6, 2021, Korean J Radiol, Vol. 22, pp. 994-1004. 21, 2021, Acad Radiol, Vols. S1076-6332, pp. 00070-2. 30 Multicenter Assessment of CT Pneumonia Analysis Prototype for Predicting Disease Severity and Patient 25 Reduction of missed thoracic findings in emergency Outcome. Homayounieh, F., Bezerra Cavalcanti whole-body. Rueckel, J., Sperl, J., Kaestle, S., Rockenbach, M. A., Ebrahimian, S., Doda Khera, R., Hoppe, B., Fink, N., Rudolph, J., Schwarze, V., Bizzo, B. C., Buch, V., Babaei, R., Karimi Mobin, H., Geyer, T., Strobl, F., Ricke, J., Ingrisch, M., & Sabel, B. Mohseni, I., Mitschke, M., Zimmermann, M., Durlak, 6, 2021, Quant Imaging Med Surg, Vol. 11, F., Rauch, F., Digumarthy, S. R., & Kalra, M. K. pp. 2486-2498. 2, 2021, J Digit Imaging, Vol. 32, pp. 320–329. 26 Artificial Intelligence-based Fully Automated Per Lobe 31 Diagnostic accuracy and performance of artificial Segmentation and Emphysema-quantification Based intelligence in detecting lung nodules in patients on Chest Computed Tomography Compared With with complex lung disease: a non-inferiority study. Global Initiative for Chronic Obstructive Lung Disease Abadia, A. F., Yacoub, B., Stringer, N., Snoddy, M., Severity of Smokers. Fischer, A. M., Varga-Szemes, A., Kocher, M., Schoepf, U. J., Aquino, G. J.,Kabakus, I., Martin, S. S., Sperl, J. I., Sahbaee, P., Neumann, D., Dargis, D., Hoelzer, P., Sperl, J. I., Sahbaee, P., Gawlitza, J., Henzler, T., Johnson, C. M., Nance, J. W., Vingiani, V., Mercer, M., & Burt, J. R. Schoenberg, S. O., & Schoepf, U. J. 3, 2022, J Thorac Imaging, Vol. 37, pp. 154–161. Suppl 1, 2020, J Thorac Imaging, Vol. 35, pp. S23-S34. 32 Artificial Intelligence has Similar Performance to Subjective Assessment of Emphysema Severity 27 Utilizing Artificial Intelligence to Determine Bone on Chest CT. Ebrahimian, S., Digumarthy, S., Mineral Density Via Chest Compute Tomography. Bizzo, B., Primak, A., Zimmermann, M., Tarbiah, M. Savage, R. H., van Assen, M., Martin, S. S., M., Kalra, M. K., & Dreyer, K. J. Sahbaee, P., Griffith, L. P., Giovagnoli, D., Sperl, J. I., Oct 2021, Acad Radiol, pp. S1076-6332(21)00421-9. Hopfgartner, C., Kärgel, R., & Schoepf, U. J. 2020, J Thorac Imaging, Vol. 35, pp. S35-S39. 20 Siemens Healthineers AG, 2024 Features, Data, and Algorithms · Whitepaper 33 Comparing Visual Scoring of Lung Injury with 39 Artificial intelligences measuring the aortic diameter a Quan- tifying AI-Based Scoring in Patients assist in identifying adverse blood pressure status with COVID-19. Biebau, C., Dubbeldam, A., including masked hypertension. Wang, Y., Bai, L., Cockmartin, L., Coudyzer, W., Coolen, J., Yang, J., Lu, Y., Fan, W., Nie, Z., Yu, J., Wen, K., Verschakelen, J. & De Wever, W. Wang, R., He, L., Yang, F., Qi, B. 1, 2021, J Belg Soc Radiol, Vol. 105, p. 16. Postgraduate Medicine, 2022, Vol. 134, No. 1, pp. 111-121 34 Artificial Intelligence Enhances Time Efficiency in Reading Chest CT Scans – A Randomized Prospective 40 Opportunistic osteoporosis screening using chest CT Study. Yacoub, B., Varga-Szemes, A., Emrich, T., with artificial intelligence. Yang, J., Liao, M., Wang, Y. Brandt, V., O'Doherty, J., Sahbaee, P., Hoelzer, P., Multicenter study in department of Radiology and Sperl, J., Sullivan, A., Kabakus, I., Burt, J., Baruah, D. Geriatrics, 2022, Vol. 33, pp 2547-2561. & Schoepf, U. J. 2021. RSNA CH03-A5. 35 Impact of Concurrent Use of Artificial Intelligence Tools on Radiologists Reading Time: A Prospective Feasibility Study. Müller, F. C., Raaschou, H., Akhtar, N., Brejnebøl, M., Collatz, L., & Andersen, M. B. 2021, Acad Radiol, pp. S1076-6332(21)00467-0. 36 The AI-Rad Companion Chest CT in clinical use at Diagnostic Linz. A field report in collaboration with Priv.- Doz. Dr. P. Brader 37 A Volumetric Analysis of Coronary Calcification on Non-Electrocardiogram Gated Chest Computed Tomography Using Commercially Available Deep Learning Artificial Intelligence. Watanabe, S., Yamaoka, T., Kurihara, K., Kishimoto, A. 28, 2022. Vol. 28, pp. 47-53 38 Thoracic Aorta Diameter Calculation by Artificial Intelligence Can Predict the Degree of Arterial Stiffness. Wang, Y., Yang, J., Lu, Y., Fan, W., Bai, L., Nie, Z., Wang, R., Yu, J., Liu, L., Liu, Y., He, L., Wen, K., Chen, L., Yang, F., Qi, B. 2021, Vol. 8 Siemens Healthineers AG, 2024 21 AI-Rad Companion Chest X-ray is not commercially available in all countries, and its future availability cannot be ensured. The information in this document contains general technical 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 Healthineers reserves the right to modify the design, packaging, specifications, and options described herein without prior notice. Please contact your local Siemens Healthineers 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 Healthineers AG Siemensstr. 3 91301 Forchheim, Germany Phone: +49 9191 18-0 siemens-healthineers.com Unrestricted · Published by Siemens Healthineers AG, Germany · © Siemens Healthineers AG, 2024