RPS 1105 - Artificial intelligence and planet radiology: the green machine

Author Block: S. D. Jagadeesha¹, R. Botchu²; ¹Mysuru/IN, ²Birmingham/UK
Purpose: The purpose of this study is to investigate paper usage in the radiology department of a single hospital institution over the last three years to forecast paper usage up to 2050.
Methods or Background: This retrospective study was performed in the radiology department of our tertiary orthopedic hospital. The study included forms used for diagnostic and interventional procedures in various departmental modalities. Diagnostic procedures require one to three forms and interventional procedures require three forms each. Based on the established ratio that 1.2 trees are cut for every 10,000 papers used, the study calculated the number of trees cut annually over the past three years and projected paper usage and tree loss until 2050
Results or Findings: Paper usage was distributed between diagnostic and interventional procedures, with 67% used in diagnostics and 33% in interventions. The corresponding number of trees cut during this period amounted to 53.7 trees, with 47.4 trees for diagnostic procedures and 6.4 trees for interventional procedures. A total of 57.8 trees for diagnostic procedures and 11.7 trees for interventional procedures were forecasted to be cut annually from 2024 to 2050, cumulatively being 1227 trees by the year 2050.
Conclusion: Our individual department had a significant contribution from paper usage in the carbon footprint of the department. Adoption of digitalized appointment, prescribing and patient records is important in reducing this and achieving NHS net-zero targets.
Limitations: The use of paper for forms, there are other significant sources of paper consumption within the department. For example, extensive paper packaging used for interventional consumables, and tissue paper used for various applications, such as covering ultrasound and CT couches, are also contributing to the overall paper usage in the radiology department. This has been excluded in the study.
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This has obtained ethical committee clearance from the Hospital.

Author Block: P. Strouhal¹, N. Khan², A. Heathcote¹, M. Darwish³, S. Persichini³, B. Miles¹, M. Trumann⁴, I. Farid³; ¹Warwick/UK, ²Dubai/AE, ³Chalfont St Giles/UK, ⁴Freiburg/DE
Purpose: Alliance Medical Ltd (AML) provides diagnostic imaging for 800,000 NHS patients annually via networked facilities. Growing concerns over operational and energy efficiencies in 2022 prompted AML’s implementation of GE HealthCare’s Imaging360 solution. 18 months on, we showcase how such Artifical Intelligence (AI) driven solutions are pivotal in refining patient flows, scheduling, staffing, energy usage and logistics management within imaging services.
Methods or Background: Integration comprised 6 separate Imaging360 components utilising data from various sources, including HL7, DICOM, Business Intelligence software and CSV extracts; mobile and static CT and MRI scanners were incorporated from multiple sites across England (with PET-CT scanners now being onboarded).
Results or Findings: Using predictive analytics, AML reduced missed appointments from 17% to 3% per week, improving resource utilisation. Optimising protocols and schedules done on-cloud allowed reduced senior staff travel (approx. 380-480 km/month) and time (37.5hr/week) to manage scanner protocols; and significantly increased scanner efficiency:
- MRI: Throughput rose from 21 to 27 scans per day (+33%), with kWh/patient reduced from 15.5 to 11.8. This saved 3.7 kWh per exam —enough to power 45 average households annually. One MRI site increased throughput by 43%, achieving 410 exams per month increase.
- CT: scanner throughput improved by 256 scans per month average, cutting idle time; and reducing protocol variability for CT chest, abdomen, pelvis from 47 to 15 standardised protocols, with related radiation doses lowered from 500 to 350 mGy.cm.
Increased throughput was achieved with no extra staff or equipment.
Conclusion: Integrating AI into radiology workflows allows transformative changes not only of operational efficiencies and cost savings, but improved sustainable practice. Going forward, further eco-friendly innovations could enhance both performance and sustainability across the healthcare imaging sector.
Limitations: Imaging360 optimised for GE HealthCare scanners
Funding for this study: GE HealthCare supporting implementation of AI platform
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: N/A

Author Block: A. Gupta, R. Dheeka, R. Kumar, A. Rastogi, H. Malhotra, K. Rangarajan; New Delhi/IN
Purpose: To investigate tradeoffs between performance and energy use when using different locally-run large language models (LLMs) and prompts for patient-centric simplification of radiology reports.
Methods or Background: This study evaluated three different open-source LLMs (Meta’s Llama 3.1-8B, Microsoft’s Phi-3.5-Mini and Mistral-7B) using five different prompts to simplify 50 computed tomography report impressions, collected from our tertiary-care oncology centre. Models were run on a local workstation with graphic processing unit. Energy use (in watt-hours) and carbon emissions (in grams) for each inference, were measured using an open-source tool (CodeCarbon). Readability of original and generated simplified reports was quantitatively assessed using an average score of four readability indices. LLM performance for simplification task was measured as difference in readability scores between original reports and LLM-generated reports. Energy efficiency ratios (performance per watt-hour) and carbon footprint (performance per gram of emissions) were calculated for each model-prompt combination.
Results or Findings: Llama-prompt 5 (multi-shot learning) demonstrated the highest task performance (7.36), best energy efficiency ratio (31.89/Wh), and least carbon footprint (44.70/g). Phi-prompt 5 achieved high simplification (6.14) and energy efficiency (25.87/Wh). For Mistral, prompt 1 (no context) was optimal (2.15/Wh and 3.01/g), but performance (1.16) lagged behind Llama and Phi. Friedman test revealed significant differences among readability scores (p < 0.001), with post-hoc Wilcoxon tests showing significant improvements for Llama and Phi over the original and Mistral, and Llama outperforming Phi (adjusted p < 0.0033).
Conclusion: Different LLM-prompt combinations showed variability in energy use, carbon emissions, and simplification task performance. These results highlight the importance of LLM-prompt combination selection for medical applications, balancing sustainability and performance.
Limitations: Development of test prompts has inherent potential for subjectivity. Apart from prompt engineering, we did not use other accuracy improving techniques like retrieval augmented generation.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Study approved by the Institute Ethics Committee All India Institute of Medical Sciences, New Delhi (Ref. No. - IEC-343/15.06.2023)

Author Block: Y. Jia¹, M. Deng¹, R. Burger¹, S. L. Sheard¹, K. Hanneman², M. Drucker Iarovich², R. Illing¹, A. G. Rockall¹; ¹London/UK, ²Toronto, ON/CA
Purpose: Image data storage and associated greenhouse gas (GHG) emissions is accelerating, yet strategies to minimise this are limited. Reducing the average file size of CT studies by reducing the number of reformats stored could help reduce emissions.
This study aims to estimate GHG emissions associated with storage of CT reformats by modelling measurements from endometrial cancer baseline staging CT. Secondary aims were to model the findings comparing cloud storage emissions and assess the hypothetical GHG mitigation impact of a data retention policy
Methods or Background: Baseline staging CT chest, abdomen, and pelvis (CT-CAP) in 183 endometrial cancer patients in a UK cancer centre between 2013-2016 were analysed (Cohort A). The number of stored multiplanar reformats, maximum intensity projections images and lung reconstructions were recorded. The file size of each reformat was noted for 30 studies (Cohort B). Comparison was made with an external dataset of 100 baseline CT-CAP from Canada between 2018-2023 (Cohort C). Mitigation of GHG emissions was projected for different storagescenarios.
Results or Findings: Reformatted series were present in cohort A (97%, 179/183), cohort B (97%, 29/30) and cohort C (100% ,100/100). Of the total file size of cohort B (25590mb), 65% (16685mb) was reformats and/or duplicate series.
On-premise storage of all reformats for cumulative new UK endometrial cancer cases from 2020-2040 would produce 349 metric tonnes CO2 equivalent (MTCO2e). Over 20 years, projected reductions in MTCO2e were 69%(107/349) for storing only acquired axial slices, 80%(70/349) for switching to cloud storage, and 36%(222/349) for implemented a data retention policy.
Conclusion: A significant number of studies contained unnecessary reformats, increasing average file size. A strategy to revise CT data storage protocols can substantially lower radiology GHG emissions, without compromising patient care.
Limitations: Small selective patient dataset.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Institutional approval was obtained for a quality improvement project.

Author Block: C. L. Thiel¹, J. Leschied², D. Carver², J. R. Scheel², R. Omary², M. Vigil-Garcia, Phd³, P. Gehrels³, C. Meijer³, O. Struk³; ¹Madison, WI/US, ²Nashville, TN/US, ³Amsterdam/NL
Purpose: To understand the sources of environmental impact of ultrasound imaging in a US-based, adult diagnostic radiology service.
Methods or Background: A life cycle assessment (LCA) was used to evaluate the environmental impact of 2 ultrasound machines and their surrounding resource needs, including production, use and disposal of other capital equipment, linens, disposable supplies, pharmaceuticals, and data storage. A sensitivity analysis was performed to assess the impact of low-carbon electricity sources.
Results or Findings: Contrary to expectations, linens and disposable supplies emerged as the major contributors to ultrasound's greenhouse gas (GHG) emissions, each accounting for approximately 30% of its total impact. Energy use from the ultrasound units themselves was comparatively lower, at 7%, along with the production of the ultrasound units (7%), and the production and use of workstations (11%). The study also noted that ultrasound equipment spent 30-45% of the time in non-scanning mode. The sensitivity analysis showed the use of photovoltaics as an electricity source would reduce US’s GHGs by 9%; however, it would not shift the major sources of GHGs away from linens and disposable supplies.
Conclusion: Reducing linen use, adopting reusable alternatives for disposable supplies, and encouraging textile and supply manufacturers and laundering facilities to develop lower carbon alternatives are essential for improving the sustainability of ultrasound practices.
Limitations: The study, while providing valuable insights into the environmental impact of US, has limitations due to its single-center focus; exclusion of mammography, nuclear medicine, and interventional radiology; a one-month data collection period; and some data and modeling limitations.
Funding for this study: No external funding was received for this study. Philips and VUMC independently contributed to this study through in-kind labor.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study was deemed non-human subjects research.

Author Block: S. Pruthi¹, C. L. Thiel², D. Carver¹, J. R. Scheel¹, R. Omary¹, M. Vigil-Garcia, Phd³, P. Gehrels³, C. Meijer³, O. Struk³; ¹Nashville, TN/US, ²Madison, WI/US, ³Amsterdam/NL
Purpose: To understand the sources of environmental impact of CT scanning within a US-based, adult diagnostic radiology setting.
Methods or Background: A life cycle assessment (LCA) was conducted, evaluating the production, use, and end-of-life of CT scanners, including energy consumption, production and use of other capital equipment, disposable supplies, linens, pharmaceuticals, and data storage. A sensitivity analysis was performed assessing the impact of a low-carbon electricity source.
Results or Findings: Energy use and disposable supplies were identified as major contributors to CT's greenhouse gas (GHG) emissions, accounting for 42% and 20%, respectively. The production of CTs contributed 17% to GHG emissions. Furthermore, the study revealed a 50% difference in GHG emissions between CT scanners of the same model, suggesting opportunities for optimization. CT scanners were also found to spend between 44-72% of the time in a non-scanning mode. Sensitivity analysis showed that using low-carbon electricity could significantly decrease CT's energy emissions, shifting the major sources of emissions to the production of CT imaging equipment and disposable supplies.
Conclusion: Optimizing energy use, minimizing disposable supplies, and ensuring efficient equipment utilization are crucial for reducing CT's environmental impact.
Limitations: The study, while providing valuable insights into the environmental impact of CT, has limitations due to its single-center focus; exclusion of mammography, nuclear medicine, and interventional radiology; a one-month data collection period; and some data and modeling limitations.
Funding for this study: No external funding was received for this study. Philips and VUMC independently contributed to this study through in-kind labor.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study was deemed non-human subjects research.

Author Block: D. Carver¹, C. L. Thiel², J. R. Scheel¹, R. Omary¹, M. Vigil-Garcia, Phd³, P. Gehrels³, S. Thornander³, C. Meijer³, O. Struk³; ¹Nashville, TN/US, ²Madison, WI/US, ³Amsterdam/NL
Purpose: To understand the sources of environmental impact of MRI within a US based diagnostic radiology department.
Methods or Background: A life cycle assessment (LCA) was conducted that evaluated the production, use, and end-of-life of 3 MRI scanners in an adult diagnostic radiology department. Other model inputs included the production and energy use of other capital equipment, disposable supplies, linens, pharmaceuticals, and data storage. A sensitivity analysis assessed the impact of using a low-carbon electricity source.
Results or Findings: Energy consumption emerged as the dominant source of MRI's greenhouse gas (GHG) emissions, representing 79% of its total impact. Notably, the 3T MRI demonstrated 1.4 times higher energy use and 1.9 times higher production emissions compared to the 1.5T. Additionally, MRI scanners were found to be in low-power or ready-to-scan mode for 72-75% of the time, indicating potential for energy optimization. Sensitivity analysis revealed that decarbonizing the electricity grid could lead to an 87% reduction in energy-related GHG emissions from MRI. In this scenario, the production of imaging equipment itself would become the largest contributor to MRI's GHG emissions.
Conclusion: Improving energy efficiency through measures such as optimizing scan protocols, developing automation of scanner efficiency modes, and transitioning to renewable energy sources are crucial steps in reducing MRI's environmental footprint. If changing the grid is not possible, other opportunities include reducing scan times via AI (e.g. Smart speed) or optimized scheduling.
Limitations: The study, while providing valuable insights into the environmental impact of MRI, has limitations due to its: single-center focus; exclusion of mammography, nuclear medicine, and interventional radiology; one-month data collection period; and some data and modeling limitations. Shifting to a low-carbon electricity grid highlights the additional need to address emissions associated with the production of MRI equipment itself.
Funding for this study: No external funding was received for this study. Philips and VUMC independently contributed to this study through in-kind labor.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study was deemed non-human subjects research.

Author Block: T. Polidori, M. Zerunian, D. De Santis, F. Pucciarelli, B. Masci, A. Del Gaudio, F. Fanelli, D. Caruso, A. Laghi; Rome/IT
Purpose: The study aimed to assess the energy and greenhouse-gas (GHG) emission savings feasible using artificial intelligence (AI) in multi-district MRI-protocols, including lumbar-spine-MRI, cardiac-MRI, and upper abdomen-MRI. We evaluated the impact of AI-algorithms applied on MRI acquisition on scan time reduction, energy consumption, and CO2 emissions per patient, providing insights into the potential benefits of AI in routine clinical practice.
Methods or Background: This retrospective study analyzed 148 patients, including 45 upper abdomen-MRI, 53 cardiac-MRI, and 50 lumbar spine-MRI.MRI scans were acquired both without and with AI assistance applied to specific 2D and 3D sequences.The Air Recon-DL (GE Healthcare) was used for T2 and DWI sequences in upper abdomen-MRI, as well as T1, T2, and STIR sequences in lumbar spine-MRI. The Sonic-DL (GE Healthcare) was applied to SSFP sequences specifically for cardiac-MRI.The outcomes measured were time savings per patient, reduced energy consumption (kW/h), and the corresponding reduction in CO2-equivalent emissions.
Results or Findings: The application of AI across all three districts studied resulted in significant time savings per patient compared to non-AI protocol (p<0.01): 5’11’’ (58%) for upper abdomen-MRI, 1’30’’ (52%) for cardiac-MRI, and 6’ (50%) for lumbar spine-MRI.These time reductions corresponded to significant energy savings of 1.39kW/h, 0.40kW/h, and 1.68kW/h per patient (p<0.05), respectively. The equivalent reduction in CO2 emissions was 0.57kg for upper abdomen-MRI, 0.16kg for cardiac-MRI, and0.69 kg for lumbar spine-MRI (p<0.05).
Conclusion: The implementation of AI in MRI protocols significantly reduces scan time, energy consumption, and GHG emissions, highlighting its potential for enhancing the sustainability of medical imaging practices.Integrating AI into routine clinical protocols could offer considerable environmental benefits, contributing to the reduction of the healthcare sector’s carbon footprint.
Limitations: Limitations include a small patient cohort and the use of a single vendor for MRI-protocols.
Funding for this study: No-funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: No

Author Block: R. T. Schirrmeister, P. S. Friemel, M. Reisert, F. Bamberg, J. Weiß, A. Rau; Freiburg/DE
Purpose: To develop and test a deep learning algorithm for approximation of contrast agent dosage based on CT scout images.
Methods or Background: We prospectively enrolled 817 patients undergoing clinically indicated CT imaging, predominantly of the chest and/or abdomen. Patient weight was collected 1) manually and 2) self-reported prior to the examination by study staff. Based on the scout images, we developed an EfficientNet convolutional neural network pipeline to estimate the optimal contrast agent dose based on patient weight and provide a browser-based user interface as a versatile open-source tool to account for different contrast agent compounds We additionally analyzed the body-weight-informative CT features using a weight-conditional variational autoencoder.
Results or Findings: The training cohort consisted of 218 chest, 51 abdominal, 511 whole-body, and 37 CT scans of various other anatomical regions. Self-reported patient weight was statistically significantly lower than manual measurements (75.02 kg vs.76.92 kg; p < 10⁻⁵, Wilcoxon signed-rank test). Our pipeline predicted patient weight with a mean absolute error of 4.74 ± 0.14 kg in 5-fold cross-validation and is publicly available at https://nora-imaging.org/ct-scout-weight/. Interpretability analysis revealed that both larger anatomical shape and higher overall Hounsfield units were predictive of body weight.
Conclusion: Our open-source deep learning pipeline allows for automatic estimation of accurate contrast agent dosing based on scout images in routine CT imaging studies. This approach has the potential to streamline contrast agent dosing workflows, improve efficiency, and enhance patient safety by providing quick and accurate weight estimates without additional measurements or reliance on potentially outdated records.
Limitations: The model's performance may vary depending on patient positioning and scout image quality and the approach requires validation on larger patient cohorts and other clinical centers.
Funding for this study: Funded by an unrestricted research grant from Siemens Healthineers.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by an ethics committee. Written informed consent was obtained from each participant.

Author Block: E. Almar Munoz, C. G. Colintenorio, C. Kremser, M. Haltmeier, A. Mayr; Innsbruck/AT
Purpose: Transcatheter Aortic Valve Implantation (TAVI) is the preferred treatment for patients with severe aortic stenosis at high to intermediate surgical risk. The gold-standard preoperative imaging modality is contrast-enhanced CTA; however, non-contrast CMR is an alternative for patients with contraindications to contrast agents despite its limitations in detecting calcifications. We propose diffusion models to improve CMR-to-CTA conversion, facilitating comprehensive TAVI planning and predicting valve calcifications without contrast.
Methods or Background: Our pipeline integrates Denoising Diffusion Probabilistic Models (DDPMs) and Stochastic Differential Equation (SDE) models. This pipeline was evaluated using an in-house dataset consisting of 39 paired CTA and CMR scans. The image pairs were aligned using rigid registration techniques. To improve the registration process, we utilized aorta segmentation masks obtained using nnUNet for CMR scans and TotalSegmentator for CTA scans.
Results or Findings: Regarding the aorta segmentation, we obtained Dice values of 0.987±0.006 for CMR and 0.980±0.005 for CTA. The Dice Score obtained in the rigid registration was above 0.87. Regarding the image conversion, our results demonstrate that the overall synthesized CTA images exhibit high fidelity to their real counterparts, validated by metrics including the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR), both exceeding 0.80 and 22, respectively. Focusing on the valve's calcifications, some are accurately converted into CTA-calcified regions but are not always consistent or repeatable.
Conclusion: This study highlights the potential of diffusion models in medical imaging, offering a promising solution for patients unable to receive contrast agents, thereby improving the safety and efficacy of TAVI planning.
Limitations: Firstly, the model encounters difficulties in replicating small details in the CTA, including calcifications. Secondly, the diffusion models applied are very sensitive to image training; both data modalities must present low noise levels or artifacts.
Funding for this study: Fund provided by FWF-DOC-110
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Nothing to declare

Author Block: S. Dong, Z. Shen, F. Yan, R. Li; Shanghai/CN
Purpose: To develop a diffusion model for generating virtual dynamic contrast-enhanced MRI (DCE-MRI) images from non-contrast T1-weighted scans and assess its efficacy in FLL diagnosis.
Methods or Background: Gadolinium-based contrast agents (GBCAs) in DCE-MRI are crucial for characterizing focal liver lesions (FLLs), but their use increases risks for patients with renal impairment and adds to imaging costs. Virtual contrast-enhanced images from non-contrast T1-weighted scans could reduce these risks and streamline diagnostics.
FLLs ≥1 cm, identified through DCE-MRI, were included, with lesion types such as HCC, ICC, liver metastases, cysts, hemangiomas, and FNH. A diffusion model was trained on non-contrast T1-weighted and corresponding multiphase DCE-MRI images (arterial, portal venous, and delayed phases). Training occurred at Center 1 (2018–2023) with a 3:1 split for training and internal testing. External validation used data from three other centers (2018–2024). A diagnostic model for FLLs was also trained on synthetic DCE-MRI images. Normalized mean absolute error (NMAE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) were used for evaluation. Three radiologists scored image quality on a three-point scale. The human machine comparison was conducted with six radiologists in different experience.
Results or Findings: The study included 1187 patients in the training set (mean age, 51 ±12), with 395 internal (52 ±16) and 347, 271, and 219 external patients (57 ±11, 56 ±12, 58 ±11). The model showed strong similarity between virtual and real DCE-MRI images, with NMAE 0.021–0.038, PSNR 28.9–31.8 dB, and SSIM 0.881–0.927. Diagnostic accuracy was 93% for the internal and 91% for external sets, outperforming three junior radiologists (P < .001) and matching three senior radiologists (P = .19).
Conclusion: The diffusion model provides a safe, cost-effective alternative to traditional DCE-MRI, maintaining high diagnostic accuracy for FLLs.
Limitations: None reported.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: None