This content was funded by Phillips and interactivity was curated and developed by EMJ.
Radiology
Introduction
A new approach to chest X-Ray Imaging
Using landmark identification in AI enabled 'Smart Collimation Thorax' to empower radiographers and to personalize patient care.
Unmet Need for Optimised Workflows
Although retakes increase dose to the patient, they frequently occur. In a study by Little et al.,2 positioning/collimation turned out to be the most common reject reason, accounting for 67.5% of rejected images (of overall rejection rates ranging from 13.0–24.5%). This was followed by incorrect technique (14.4%) and patient motion (8.1%).
67.5
Positioning/ collimation
Incorrect
technique
Patient motion
14.4
8.1
%
%
%
Repeating these exams results in additional radiation exposure for the patient. In addition, the repeat exams further strain radiographers’ time and resources, exacerbating their already high workload. This growing burden highlights the increasing demand for automation and workflow solutions in radiography.
AI Models in Smart Collimation Thorax
Philips Diagnostic X-ray is addressing these challenges by developing tools to support radiographers in their routine workflows. Smart Collimation Thorax (SCT) is one of these tools, based on AI, and offers assisted detector height alignment and collimation proposals for upright standing chest X-ray exams in adult patients.
Smart Collimation Thorax: an AI enabled feature to provide personalised detector height alignment and collimation proposals for upright standing chest exams
Smart Collimation Thorax includes two-features:
Assisted detector height alignment: offering a personalised proposal for the height of the wallstand detector, based on the patient’s anatomy (available for chest posteroanterior (PA) views).
Assisted Collimation: offering a personalised proposal for the collimated area, based on the patient’s anatomy (available for chest PA and lateral (LAT) views).
The SCT feature is based on multiple AI models, which detect patient-specific landmarks that are indicative of the extent of the lung and predict the parameters of interest: the detector height and the patient’s lung field.
Landmarks shown here are for illustrative purposes only.
Created with BioRender.
Learn more about landmarks
The final proposal provided by Smart Collimation Thorax, is based on the following steps:
As input, these models use a depth image from a 3D camera integrated into the collimator at the point of time of activating SCT and predict patient-specific landmarks.
When activating the assisted detector height alignment, a detector height proposal is directly derived from the patient-specific landmarks. For assisted collimation, the patient’s lung field (lung bounding box) is first predicted based on the detected landmarks.
Subsequently, margins around the lung field and geometrical system restrictions are considered in the final collimation proposal.
Finally, plausibility checks are performed. If these checks are passed, the detector height/collimation is proposed to the user by SCT. If needed, these proposals can be adapted manually by the system operator.
This approach is illustrated for assisted collimation in Figure 1, showing exemplary depth images (top: PA, bottom: LAT) as well as the landmark prediction and collimation proposal.
All four steps are visualised:
The left column shows the input: the depth image of the patient from the camera view.
The middle column shows the depth image of the patient from a focal point of view including detected landmarks.
The right column shows the depth image of the patient from a focal point view, including proposed collimation by SCT, which is predicted based on the detected landmarks.
Figure 1: Exemplary 3D images (PA and LAT view) showing landmark prediction and collimation proposal.
More detailed information on the models is provided in Medical Image Computing and Computer Assisted Intervention – MICCAI 2018.4
AI Models in Smart Collimation Thorax
Model development
Data used for development and testing
To achieve high data quality, multiple sites were selected with a broad geographical distribution. The data used for development and testing were collected at six clinical sites from five different countries in Europe and two sites in the USA. This way, a high variability of system settings and ways of working was considered.
The data was collected during clinical routine to ensure a wide variety of patient characteristics, for example in terms of sex, age and Body Mass Index (BMI). Since LAT images with different tube head angles were collected, the models can process data from a rotated tube head.
Model training was based on 2,274 PA views and 2,148 LAT views, while 279 PA views and 269 LAT views were used for validation. The test data set for final evaluation contained 755 PA views and 586 LAT projecttions from both inpatient and outpatient settings. Exemplary data sets are shown in Figure 2.
This figure also shows examples of a rotated tube head, as described above.
Figure 2: Exemplary data sets (PA and LAT views) including 3D image (depth information displayed via colour gradation) and corresponding X-ray image. Data sets collected from hospitals in Europe.
Methods for Testing the Performance of the Models
Retrospective data analysis on the collected test dataset was conducted to test the performance of the models. This testing included claim substantiation. The detector height proposal by SCT was compared to the detector height chosen by the system user during data collection. Correspondingly, the collimation proposal by SCT (Assisted Collimation) was mapped onto the X-ray image and compared to the collimation set by the system user for the respective data set.
Collimation parameters analysed
The aim of the evaluation was to show that the proposed collimations by SCT are at least non-inferior to the collimations set by the user. The following parameters were investigated regarding the collimation:
The area exposed beyond the lung bounding box is considered unnecessarily exposed area. To test this parameter, both the area beyond the lung bounding box of the collimation set by the user and the collimation proposed by SCT were calculated and compared.
A visualisation of all three performance parameters is provided in Figure 3.
The importance of First time right collimation
Figure 3: Exemplary X-ray images from collected data sets, including lung bounding box (in red) for visualisation purposes and proposed collimation by SCT (in green).
Test Results of the Performance Analysis
Results – detector height alignment
The result for non-inferiority testing is significant with p<0.0001. This means the proposed detector height alignment by SCT places the wallstand detector equally often in an appropriate imaging position as the user.
SCT proposed collimations show a significant decrease in unnecessary X-ray exposed areas for PA as well as LAT views. The test results are shown in Table 1 for the PA.
Results – unnecessary exposed area
The test results are shown in Table 1 for the PA projection and in Table 2 for the LAT projection.
Table 1: Statistical testing of the unnecessarily exposed area (area beyond the lung bounding box) for PA projections.
Table 2: Statistical testing of unnecessarily exposed area (area beyond the lung bounding box) for LAT projections.
SCT proposed collimations lead to statistically fewer cases of over collimation for PA views. Further, SCT proposed collimations do not lead to fewer cases
of over collimation for LAT views.
Results – over collimation
SCT proposed collimations show statistically more uniformity for PA projections. Further, SCT proposed collimation is more uniform for the LAT view as well.
Results – uniformity
Results – first time right collimation
SCT is statistically more likely than the user to have first time right collimations for the PA and LAT view (Tables 3 and 4).
When comparing the number of cases of first time right collimations, a 95% CI of the difference is given using a normal approximation method. Table 3 shows the results for PA views (N=634). SCT resulted in 484 (76.34%) first time right collimations, and the user performed 110 first time right collimations (17.35%). The difference is 58.99%, with a 95% CI of 54.56–63.42%, which is statistically meaningful (Table 3).
Table 4 shows the results for LAT views (N=457). SCT resulted in 329 (71.99%) first time right collimation proposals, while the user performed 84 first time right collimations (18.38%). The difference is 53.61%, with a CI of 48.17–59.05%, which is statistically meaningful.
Table 3: Statistical testing of first time right collimations for PA views.
Table 4: Statistical testing of first time right collimations for LAT views.
As part of a Philips internal user study, 12 radiographers were invited to a Philips lab environment to use the SCT functionality and provide their feedback. After a short introduction and training, all participants were requested to perform a simulated PA chest exam with and without SCT and to answer some short questions.
Smart Collimation Thorax: Impact on Workflow
The following experiences/feedback have been provided:
Smart Collimation Thorax: Results on Workflow
Improvements (simulated use with 12 radiographers)
SCT reduces exam time by up to 35 seconds, resulting in a time saving up to 20 minutes daily for your medical team.*
*Compared to not using SCT on the Radiography 7300 C, validated by 12 clinicians in a Philips’ development environment. Daily time-saving calculation based on 35 adult upright chest patients per day. Results may vary. The average time saving is 8.1 sec.
11 out of 12 users (91%) think SCT can reduce human error caused by manual and repetitive tasks.
11 out of 12 users (91%) like triggering SCT from the wall stand control panel as it enables them to stand with their patient during positioning and collimation.
10 out of 12 users completed a PA chest X-ray exam faster with the use of SCT, than without the use of SCT.
This article described the development and testing of the AI models used in SCT. Performance testing was done for both sub-features: assisted collimation and assisted detector height alignment.
Performance testing showed a superior performance by assisted collimation compared to the system user, for most of the performance parameters examined. The performance of assisted detector height alignment is non-inferior to the user with statistical significance. This means SCT places the wallstand detector equally often in an appropriate imaging position as the user.
Summary
Smart Collimation Thorax: Results on Performance Evaluation
SCT proposed collimations show a significant decrease in unnecessary X-ray exposed area for PA as well as LAT views
SCT proposed collimations lead to fewer cases of over collimation for PA views
SCT proposed collimations show more uniformity for PA views
With SCT, it is more likely to have first time right collimations, both for PA as well as LAT views
References
Acknowledgements: Medical Writing assistance was provided by Jessica Jinks, EMJ, London, UK.
Keywords: AI, imaging workflow, model validation, performance testing, radiographers, radiography, simulated study, Smart Collimation Thorax (SCT).
Summary
Summary
This article describes the development and evaluation of the AI models underpinning Smart Collimation Thorax (SCT). The first section outlines the unmet clinical need addressed by SCT. The second section details model development, including training and validation procedures, and describes the datasets used for both development and independent testing. The methodology used to assess model performance is then presented, followed by the corresponding performance results. Finally, the impact of SCT on clinical workflow is examined.
Insufficient training and experience can result in positioning or collimation errors, leading to rejected images.
Radiographers rely on their training and experience to position the patient’s internal anatomy correctly for posteroanterior (PA) and lateral (LAT) chest images.
Landmarks are specific points on the human body, extracted from the depth camera, that are indicative of the extent of the lung and therefore can be used to propose collimation and height adjustment. Since these landmarks are extracted from individual patients, these are patient specific, resulting in a personalised proposal.
In fact, 23% of technologists believe that automation could make certain tasks more efficient.3
Automation could help streamline repetitive tasks, reduce errors, and improve image quality, thus decreasing the need for repeat exams, offering a potential solution to ease the pressure on radiographers and enhance the overall efficiency of diagnostic imaging.
Smart Collimation Thorax is an AI enabled feature, available on Philips fixed X-ray system Radiography 7300 C (and as an upgrade of DigitalDiagnost C90), providing the radiographer a personalized proposal for collimation and height adjustment of the detector, when performing upright standing X-ray exams in adult patients.
Model development
In various development cycles, the models were trained and validated (development stage). Based on the validation performance results, the hyperparameters of the models were optimised and the generalisability of the models on unseen data sets was estimated prior to the final evaluation. The performance of the final models was tested using a completely unseen test data set (testing stage).
Strength of data collection - Robust data set:
Routine clinical data (no change in workflow requested) Large number of samples (>3000 datasets)
Data from Europe and USA (6 sites in total), wide ranges of geographies and patient characteristics
AI algorithm input was based on certified radiographers annotating X-ray images with landmarks
The detector height and the collimation set by the operator were recorded by geometrical sensors and are available in the meta data of the collected data sets. Minimum sample size for testing has been calculated and was met.
Detector height alignment performance testing
It was tested whether the proposed detector height alignment by SCT places the wallstand detector equally often in an appropriate imaging position as the user. The number of height settings where the upper or the lower border of the lung field (lung bounding box) are not on the detector, are counted as inappropriate positions. These are used for comparison between the user height alignment and the detector height proposal by SCT.
The area exposed beyond the lung bounding box is considered unnecessarily exposed area. To test this parameter, both the area beyond the lung bounding box of the collimation set by the user and the collimation proposed by SCT were calculated and compared.
Uniformity refers to the extent in which collimations are uniform. Uniform collimations mean that the distances between each collimation border and the respective lung bounding box border are of a similar magnitude from image to image. The standard deviations of the distances are calculated, considering the different image sides and the clinical sites of the test data set, and aggregated to an overall uniformity measure.
Unnecessarily exposed area: This image shows an example of less unnecessary X-ray exposed area: the area that was unnecessarily exposed by the user in comparison with the proposal by SCT leads to less unnecessary exposed area.
Over collimation: This image shows an example of a cropped lung due to over collimation by the system user.
Uniformity: This image illustrates the concept of uniformity: blue arrows indicate the distances between the lung bounding box to the respective collimation proposed by SCT on all four image sides, which are used to assess the parameter uniformity.
First time right collimation is defined as a collimation which neither shows over collimation nor a larger unnecessarily exposed area comparing user collimation and the collimation proposal by SCT. According to American College of Radiology (ACR) guidelines, ‘the radiographic beam should be appropriately collimated to include the structures listed while limiting exposure of the remainder of the patient’.5
To enable development and testing of the AI models, ground truth was generated. According to clinical guidelines (American College of Radiologists- ACR) the collimation should include both lung apices and the costophrenic sulci5. Ground truth is defined as the lung bounding box which is in line with the clinical guidelines and is equivalent to the minimum collimation to cover the lung field. For this purpose, anatomical landmarks were annotated on the collected X-ray images by international clinical experts, radiographers who are certified to acquire X-ray images in clinical settings. For model training, the annotated landmarks are mapped onto a 3D image and used as input parameters. (Shown in Figure 3)
The comparison between SCT and the user regarding detector height alignment was performed using a chi-square method.
An adjusted one-sided alpha of 0.0083 was used in this analysis.
Non-inferiority testing was performed with a non-inferiority margin of 0.015 (N=645). SCT proposed three positions in which the entire lung field cannot be imaged on the detector (0.47% of the total cases) and the user eight (1.24% of the total cases).
To investigate the parameter unnecessary X-ray exposed area, a paired t-test was performed. An adjusted one-sided alpha of 0.0083 was used in these analyses. The mean values of the unnecessarily exposed area were calculated for both the collimation proposal by SCT and the user collimation. The difference between these means is statistically significant with p<0.0001, for both PA views (N=644) and LAT views (N=474).
The number of cases with over collimation were compared using a chi-squared method. No adjustment of alpha was made in this comparison. Superiority testing for PA views (N=634) shows a p value of 0.03 for the number of cases with over collimation for the algorithm (n=17) versus the user (n=36) with a 95% two-sided CI (-0.0519– -0.0080).
Superiority testing for LAT views (N=457) shows a nonsignificant difference between the numbers of cases; the number of cases with over collimation was higher for the algorithm (n=43) versus the user (n=5; 95% two-sided CI: 0.0547– -0.1116).
The overall uniformity measure for the PA view (N=644) is 0.8239. As this value and the upper bound of the 95% CI (0.7562–0.8837) are below one, it is demonstrated that SCT proposed collimations show statistically more uniformity for
PA projections.
For the LAT view (N=474), the overall uniformity measure is 0.9544, which is below one, indicating that SCT proposed collimation is more uniform for the LAT view as well. However, the result is not statistically meaningful with a 95% CI of 0.8818–1.0384, as the upper bound is slightly above one.
The results of over collimation should always be considered in the context of a decrease in unnecessary X-ray exposed area. ‘First time right’ is the measure combining over collimation with unnecessary X-ray exposed area. The number of cases with ‘first time right’ collimation is significantly higher for SCT, compared to users collimation, for both PA (76,34% SCT versus 17,35% user) and LAT (71,99% algorithm versus 18.38% user) views.
Key messages
Fully patient adaptive collimation suggestion, up to the level of being unique for that individual at that moment, while at the same time, proven to be consistent between examinations.
Consistent exam performance irrespective of radiographer experience.
Flexible workflow where the functionality isn't always on, but always available. All of this while providing fast and accurate support, without forcing the user to adopt a certain way of working.
References
World Health Organization (WHO). To X-ray or not to X-ray. 2016. Available at: https://www.who.int/news-room/feature-stories/detail/to-x-ray-or-not-to-x-ray-. Last accessed: 10 January 2025.
Little KJ et al. Unified database for rejected image analysis across multiple vendors in radiography. J Am Coll Radiology. 2017;14:208-16.
Philips. Radiology staff in focus - A radiology services impact and satisfaction survey of technologists and imaging directors. 2019. Available at: https://www.philips.com/c-dam/b2bhc/master/Specialties/radiology/radiology-staff-in-focus/radiology-staff-in-focus.pdf. Last accessed: 9th February 2026.
Sénégas J et al., “Evaluation of collimation prediction based on depth images and automated landmark detection for routine clinical chest X-ray exams,” Frangi A et al. (eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2018. (2018), Springer Cham, pp.571-9.
ACR–SPR–STR Practice Parameter For The Performance Of Chest Radiography. Available at: https://www.acr.org/-/media/ACR/Files/Practice-Parameters/ChestRad.pdf. Last accessed: 9th February 2026.
Hayre C et al. Collimation and cropping in diagnostic radiography: how concerned are we? J Med Radiat Sci. 2023;70(1):8-10.
References
Every year, 3.1 billion diagnostic exams, like radiography exams, are performed globally.1
Many X-ray exams are upright standing chest images in radiography rooms. Radiographers handle these exams repetitively, manually setting the system height and adjusting collimation for each patient individually, following the As Low As Reasonably Achievable (ALARA) principle.
Upright standing chest exams were collected during routine clinical practice from adult patients performed on the Philips DigitalDiagnost C90 X-ray system with an integrated 3D depth camera. The acquired data contains depth camera sequences and X-ray images of PA and LAT views from over 3,000 patients.
Discussion
In the final data set for the performance testing of the algorithm:
51
%
of patients were women
37
%
of patients were aged between 45 and 64
37
%
of patients aged 65 and older
Patients in this data set had a BMI ranging from below 18.5 kg/m2 to over 30.5 kg/m2.
In this article, the following performance parameters have been investigated for SCT, unnecessary X-ray exposed area, over collimation, uniformity and first time right collimations. It could be shown with statistical significance that compared to the user, SCT proposed collimations show a decrease in unnecessary X-ray exposed area for both PA and LAT projections.
Underweight
Normal
Overweight
Obese
Morbidly Obese
18.5 kg/m2
30.5 kg/m2
Regarding over collimation, a statistically meaningful decrease was observed for the PA view. This could not be shown for the LAT view, indicating a way of working by the system users that focusses on avoiding over collimation and considering that the field of view can be reduced after exposure. This phenomenon has been described in literature as ‘collimation creep’6. Collimations by SCT statistically display more uniformity for PA projections as compared to the user. For LAT views, results indicated that SCT proposed collimations are more uniform, but the result is not statistically meaningful. It is to be expected that improved uniformity of X-ray images will facilitate reading of images.
The comparison between SCT and the user in terms of first time right collimations showed for both projections that the probability of first time right collimations is statistically higher using SCT than applying a collimation set by the user. This result underlines the potential for quality improvement by integrating SCT into the standard clinical workflow.
Exemplary X-ray images from collected data sets including lung bounding box (in red) for visualisation purposes and proposed collimation by SCT (in green).
In future studies, real-world evidence might provide valuable insights into the actual use and effectiveness of SCT. By examining data from clinical settings, it is possible to assess the impact of these tools on reject rates, offering a clearer picture of how they contribute to reducing errors and improving image quality. Additionally, real-world evidence can help quantify the health economic benefits of SCT, particularly in terms of time savings for radiographers, which can translate into more efficient resource utilisation, lower operational costs, and improved patient throughput across healthcare facilities.
As described in the Model Development section, to achieve a high data quality, multiple sites were selected with a broad geographical distribution. This way, a high variability of system settings and ways of working was considered. Data have been collected during clinical routine to ensure a wide variety of patient characteristics, for example in terms of sex, age and BMI.