Sophie Winandy, Executive Director of Medical Imaging at ICON explores the benefits of AI and how it can be integrated throughout medical imaging within clinical trials.
Artificial intelligence (AI) is rapidly taking hold in the healthcare sector1, including in medical imaging, where AI often targets screening, patient eligibility, detection and disease response (Figure 1). As we develop specialised AI, appropriate adoption and integration of AI systems can potentially alleviate the overburdened and understaffed healthcare industry, while also enhancing the value of care provided to patients.
When integrated into workflows, AI systems can support human expertise to drive efficiencies that benefit clinical trials,1,2,3 including the participating clinical sites, pharmaceutical industry stakeholders – like contract research organisations (CROs) and ultimately, the patients. These systems can drive quality and cost-effectiveness by mitigating the potential for human error, accelerating processes for faster results and potentially increasing the imaging biomarkers available to assess response. With high-quality data in hand faster, sponsors can make more timely decisions to support clinical milestones.
Figure 1: Applications of AI in medical imagingStreamlining image acquisition and data transfer
The human element and intuitive scientific overlay within medical imaging are irreplaceable. However, integrating AI can streamline workflows to save time and mitigate the potential for human error. The first integration point for AI in a medical imaging workflow is image acquisition, where AI is already playing a role in decreasing scan time, reducing ionising radiation and increasing availability of imaging tools.
AI can also assist during image upload from clinical sites, where it can automatically scan images and aid in quality control (QC) to 1) Validate patient information and image type, 2) Verify against a specific imaging protocol, 3) Identify any incomplete or missing images, and 4) Prevent inadvertent transfer of protected health information (PHI). If AI flags any errors during QC, site and sponsor personnel know immediately and can promptly remedy the issues to avoid delays. Thereby, the QC burden and timeline are greatly reduced at this stage.
Independent central reads: AI-assisted image assessment
Image assessment is critically important as it provides data for evaluating the safety and efficacy of new treatments. Advancements in AI2,3,4 now allow radiologists and readers to save precious time during interpretation and post-processing while delivering more endpoint-relevant biomarkers. These advances are seen in many indications of imaging as it relates to drug trials – oncological, cardiovascular, metabolic, NASH, musculoskeletal and other.
CROs often operate with a read paradigm that includes multiple reviewers plus adjudication for consensus. CROs will pull from a pool of specialised readers depending on the specific criteria requirements and anatomy of the assessments. Partial automation of the assessments, supported by AI integration, reduces reader fatigue and decreases variability between readers. Compounded over the many individual images generated per patient per visit, these optimisations significantly expedite assessments and increase consensus while maintaining quality and integrity.
Oncological imaging applications of AI
Over the past decade, there has been great progress in using AI for segmentation of different tissues and tumours/lesions using U-Net architecture, a deep-learning semantic segmentation technique originally proposed for medical image segmentation. For instance, manual segmentation of different brain tumour tissues (i.e., oedema, enhancing tumour, necrosis) from MRI images is an arduous and time-consuming task. This is in part due to the tumour’s heterogeneity in appearance and its irregular shape. In addition, this complexity often leads to high inter- and intra- reader variability which leads to difficulty in evaluating drug response in clinical trials.
Using AI can significantly reduce manual annotation time (up to 20 times faster) and reduce reader variability. Leveraging “Active Learning”, where AI models can incorporate changes to their own preliminary output once it has been validated by human experts, further supports the significance of using AI in such tasks. AI can now aid in semi-automated assessments of tumour measurements across various imaging assessment criteria including RECIST 1.1, RANO, mRECIST and Lugano assessment criteria.
For example, a liver oncology study may receive hundreds of images of a patient’s liver over multiple visits. Where earlier radiologists would have had to measure the lesion in each image, a time-consuming process, now they can measure the lesion on the first instance of the image and allow AI to identify and measure at the remaining timepoints. This can be followed with a round of radiologist-led QC to confirm automated measurements. Similarly, lung nodules can be characterised from HRCT data using advanced AI-driven algorithms, leading to greater sensitivity and lower adjudication rates.
This AI-enabled process optimises the radiologist’s reading time. In turn, faster assessments and data analysis can reduce study timelines.
Looking at the immediate future, AI can aid in the automated assessment of tumour volume burden (both anatomic and metabolic), a resource-intense process without AI5. This can lead to generation of more objective endpoints that might lead to a better assessment of treatment efficacy. Similarly, more accurate modelling of tumour growth rate kinetics is now more feasible due to volumetric assessments of tumour burden6.
Other applications of AI
Another application of AI is in liver tissue segmentation of its lobes and lesions. Segmented liver masks could be used to quantify markers of various liver disease including liver fat percentage and liver stiffness in non-alcoholic steatohepatitis (NASH).
Cardiac function and flow assessments using echocardiography and MRI, complex 4D flow analysis, assessments of plaque burden, viability and tissue characterisation of myocardial tissue are all feasible6,7. Cardiometabolic applications of AI include estimation of visceral and subcutaneous fat assessment for obesity assessments. Automated bone age assessments, organ volume assessments, musculoskeletal assessments like scoliosis assessment, fracture risk and detection, etc., are all available currently.
Judicious use of AI in clinical trials
There is a plethora of AI tools available in the market – some of them are regulatorily approved, some are well validated and published upon, and some are at various stages of incomplete approval. The benefits of AI are many. However, a judicious use of AI is warranted in clinical trials. Successful integration of AI depends on a clear understanding of the benefits of AI, the inherent biases due to it, and extensive validation8 to ensure it appropriately supports clinical trials and the wider healthcare system. AI has the potential to reduce workload and fatigue for personnel, facilitate faster decision-making backed by higher-quality data, and enable the availability of better biomarkers and endpoints for clinical trials.
References
- John C. Gore. Preface Document. Magnetic Resonance Imaging 68 (2020) A1–A4.
- Currie, et al. ‘Machine Learning and Deep Learning in Medical Imaging: Intelligent Imaging’ Journal of Medical Imaging and Radiation Sciences 50 (2019) 477-487
- Fazal, et al. 2018. ‘The past, present and future role of artificial intelligence in imaging’. European Journal of Radiology 105 (2018) 246–250
- Jiang, et al. ‘Development and application of artificial intelligence in cardiac imaging’. Br J Radiol. 2020 Sep 1;93(1113):20190812
- Fornier, et al. ‘Twenty Years On: RECIST as Biomarker of Response in Solid Tumours an EORTC Imaging Group – ESOI Joint Paper’. Frontiers in Oncology. January 2022 | Volume 11 | Article 800547
- Yeh, et al. Tumor Growth Rate Informs Treatment Efficacy in Metastatic Pancreatic adenocarcinoma: Application of a Growth and Regression Model to Pivotal Trial and Real-World Data. The Oncologist, 2023, 28, 139–148
- Davis, et al. ‘Artificial Intelligence and Echocardiography: A Primer for Cardiac Sonographers’. J Am Soc Echocardiogr. 2020 Sep; 33(9): 1061–1066.
- Recht, M.P., et al. Integrating artificial intelligence into the clinical practice of radiology: challenges and recommendations. Eur Radiol 30, 3576–3584 (2020).
About the author
Sophie Winandy is Executive Director of Medical Imaging at ICON. She has over 15 years of experience in clinical trial imaging, on both the operational and sales sides. As a subject matter expert on medical imaging, she is responsible for ICON’s imaging sales and strategy globally.
