CARJ 2025: Year in Review

Artificial Intelligence (AI)

The applications of AI in radiology are rapidly evolving with algorithms capable of segmentation, localisation, diagnosis, report generation and triage in a wide range of disease processes. This past year has seen a number of publications on AI in the CARJ with 4 of the most interesting and novel selected for review.

AI models which can predict tumour classification, molecular subtype or behaviour have significant potential in radiology with the possibility of reducing the need for invasive biopsy procedures. Machine learning in the form of radiomics and convolutional neural networks (CNNs) have been proposed for the prediction of molecular subtypes in various tumours. Namdar et al describe the advantage of MRI-based CNNs for the prediction of paediatric low grade glioma molecular subtypes which, unlike radiomic-based models, are not limited to predefined pattern recognition formulas. ¹ With the incorporation of tumour location, their CNN-based model achieved a sensitivity of 85.1% (95% confidence interval [CI] 81.7%-88.5%), specificity of 84.7% (95% CI 81.2%-88.3%) and accuracy of 85.0% (95% CI 82.6%-87.3%). ¹ These findings encourage the use of tumour location probability maps in CNNs for molecular subtype identification in this setting, highlighting how manual segmentation, where location information is lost, may by suboptimal. ¹ As tailored treatment options for patients with specific molecular markers continue to evolve and expand, we expect this type of machine learning to continue to be a focus of AI research in the future.

Beyond diagnostic applications, AI models have been proposed as triage tools to streamline radiology workflow. Radiologists rely on brief clinical vignettes to determine the appropriate modality, study protocol and priority of imaging requests. The volume of studies ordered make this a time-consuming task which disrupts other duties including reporting. Protocolling has also been reported as a source of burnout among radiology trainees. ² Yao et al propose a bidirectional encoder representations for transformers (BERT) natural language processing (NLP) model to automatically triage and protocol cross-sectional imaging requests, using naive Bayes (NB) support vector machine (SVM) comparators. ² The BERT NLP model achieved the highest performance parameters for protocol selection with an F1 score of 0.901. ² For triage prediction, the BERT NLP and SVM models performed similarly with an F1 score of 0.844 and 0.845 respectively. ² F1 scores were highest for both protocol selection (0.957) and triage (0.986) for patients located in the emergency department. ² High volumes of time sensitive requests are issued from the emergency department and implementation of this model has the potential to improve efficiency and workflow. There is a strong clinical desire for such automated triage and protocolling tools with 95% of clinicians surveyed across Canada expressing support for an AI tool that prioritises MRI requests. ³

Large language models (LLMs) have also been proposed as decision support tools to guide physicians on the appropriate management of imaging findings, including incidental findings which have a wide range of implications. Given the specialised nature of such findings, general-purpose LLMs may fall short if insufficiently trained on proprietary or region-specific guidelines. Dietrich et al highlight the potential for general-purpose models to produce outputs which fail to align with expert consensus and propose the use of retrieval-augmented generation (RAG) to overcome this challenge. ⁴ They focus on incidental hepatobiliary lesions, which occur in up to 30% of radiological studies. ⁴ Accurate management recommendations regarding these lesions, which range from benign cysts to potentially malignant masses, is essential. Guidelines from the Canadian Association of Radiologists (CAR) on how to handle incidental hepatobiliary findings are a valuable tool for radiologists. However, given their specialised nature, are unlikely to be included in general LLM training. ⁴ RAG systems integrate an external retrieval mechanism into the generation process thus allowing LLMs to access domain-specific content instantaneously. By reducing knowledge gaps in this way, RAG-enabled models are less likely to provide inaccurate outputs, hallucinate or deviate from the evidence base. Dietrich and Stubbert demonstrate the utility of RAG-enabled LLMs for improving adherence to the CAR guidelines for incidental hepatobiliary findings, reporting an improvement of approximately 15% when compared with non-RAG LLMs. ⁴ Advances such as this reiterate the potential for AI as a decision support tool, particularly to drive overall adherence to expert consensus and clinical guidance.

Despite worldwide emphasis on the potential AI has to revolutionise modern living, including modern healthcare, there remains a pervasive distrust of the so called AI “black-box.” Earlier this year, Hughes et al explored the origin of this distrust, the undeniable promise of the “revolution in robotics” and the ethical considerations of robotic autonomy. ⁵ While many sectors fear replacement by robots, the provision of health services is under constant pressure to meet demands, with a shortage of 6.4 million doctors and 30.6 million nurses and midwives estimated in 2019. ⁵ Should we still fear replacement by robots or should we be recognising the promise AI has to alleviate personnel shortages? Much of the emphasis on AI “replacing” radiologists has been focused on diagnostic applications. The idea of a robot performing a biopsy seems far from reality but is the interventionalist safe? Hughes et al report robotic-AI systems for image-guided percutaneous needle biopsy and tumour ablation. ⁵ While robotic-assistance could improve efficiency and accuracy, it is unlikely to replace the human operator. However, the fear of fully autonomous robots performing at an equal or superior level to that of humans is a major obstacle to embracing this type of progress for physicians and patients alike. The ethical and legal implications of robotic autonomy also warrant consideration. If we cannot understand how an algorithm directs a robot to make a decision, can we trust it? Two points raised by Hughes et al stand out: (1) Is informed consent truly possible and (2) Who is ultimately responsible for decision errors or negative outcomes. ⁵ The answers to these questions remain unknown, and while our jobs are safe for the time being, we must continue to ponder the future of AI for progress in this field is less of a possibility and more of a certainty.

Sustainability

The CAR released a statement on environmental sustainability in medical imaging earlier this year, calling for immediate action. ⁶ Human activities in our personal and professional lives impact the environment, with downstream implications including climate change. In radiology, the delivery of medical imaging generates greenhouse gas emissions and waste, contributing to the climate crisis. In turn, medical imaging equipment is vulnerable to the effects of climate change through flooding, extreme temperatures and power shortages. ⁶ Improving our environmental impact is therefore essential for the future of radiology, but is it an achievable feat?

Hanneman et al delve into the environmentally sustainable radiology programmes in Canada, highlighting their successes and reminding us that it is not too late for real, impactful change. ⁶ Powering down a single computed tomography (CT) scanner in Vancouver overnight and at the weekend translated to energy savings of over 14 000 kWh and a reduction of CO₂ emissions of 5 tons per year. ⁷ Portable, low-field MRI achieved diagnostic image quality with lower energy and potential financial savings in the region of $CA 854 841, allowing high quality diagnostic imaging to be brought to remote communities. ⁶ Bringing diagnostics closer to patients reduces wait times, travel costs and greenhouse gas emissions; however, operator dependent modalities such as ultrasound are not always feasible in these settings. Telerobotic ultrasound overcomes this challenge with 70% of remotely controlled robotic ultrasounds performed in Northern Saskatchewan of sufficient diagnostic quality. ⁶ Abbreviated MRI protocols can also achieve significant energy savings without compromising diagnostic quality. An abbreviated cardiac MRI protocol reduced energy use by 19% in Toronto, conferring a potential to avoid 5600 kg of CO₂ emissions per year. ⁶ In addition, provincial-wide electronic health and imaging records have the potential to positively impact greenhouse gas emissions by reducing unnecessary repeat imaging. ⁶ While these are positive examples of the potential for change, further work is needed to achieve net-zero targets. A major component of this work requires collaboration with industry to improve the energy efficiency of equipment and move toward circular economy models.

Planetary health education must be at the forefront of the sustainable radiology movement to strive for long-term change. Brown et al released a statement providing guidelines on how to deliver education on climate change and environmental sustainability among radiologists. ⁸ Trainee education is central in this statement, highlighting the importance of providing future radiologists with an understanding of the origin of the climate crisis, the impact on human health and how healthcare contributes to the problem. ⁸ Brown et al suggest an “environmental sustainability in radiology” module on the CAR online teaching platform, RAD academy and the development of a Canadian radiology fellowship in sustainable imaging as ways to formalise this education. Within residency programmes, they suggest formal identification of a sustainability lead, establishment of green teams and sustainability research days to drive efforts. ⁸ 2025 has seen radiologists recognised as well positioned healthcare professionals to lead transformative change in low emission healthcare that is resilient to the effects of climate change and this is an expectation we must now live up to.

System Efficiency

Managing the ever-increasing demand for radiological investigations and interventional procedures in a timely fashion has been a challenge facing radiology departments for some time. While high volumes and low wait times were historically the parameters emphasised to determine system efficiency, the concept of value-based radiology (VBR) has emerged which shifts focus to patient outcomes including alterations in patient management or clinical course. A critical aspect of VBR is assuring appropriateness of the radiological investigation and reducing “low-value” imaging. Radiological studies that would fall into the “low-value” category include imaging for atraumatic pain or following minor head trauma and unsurprisingly, such studies contribute significantly to global healthcare costs. ⁹ The CAR have issued 5 top recommendations to reduce this type of “low-value” imaging in Canada including using CT first line in paediatric appendicitis and ankle series in adults following minor injuries. ⁹ Furthermore, the CAR have extensive referral guidelines developed with interdisciplinary input from a wide range of subspecialist experts to provide referrers with the best possible information to guide the ordering of radiological investigations. ⁹ While it is an ultimate goal to have this information integrated into the radiology ordering systems, we remind our readers to promote the use of referral guidelines within their hospitals until this goal becomes a reality. Focusing on VBR and reducing inappropriate requests for CT head for example, could save your department over $204 000. ⁹ Beyond the economic implications, such changes also have the potential to improve wait times and reduce our environmental impact. At an individual level, radiologists also have a role in improving system efficiency. Interventions which lead to improvements in radiologist workflow include the integration of imaging archives with pertinent clinical information from the patient’s electronic health record within the radiologists workstation to limit time taken to chase up relevant clinical details. ⁹ Centralisation of imaging archiving systems is also impactful in this way, allowing the radiologist to review relevant external imaging, not only improving accuracy in reporting but also in improving individual radiologist workflow. ⁹ As the demand for provision of radiological services continues to exponentially grow, we expect to see further literature emerging in the years to come on how to further our efficiency efforts.

While ensuring the appropriate test is performed for the correct clinical indication is the first step in achieving VBR and improving system productivity, we must also ensure that the test we are performing is optimised, limiting or removing aspects that do not bring true diagnostic value. Wagner and Samji evaluated the utility of dynamic contrast enhanced (DCE) imaging sequences in patients undergoing multiparametric prostate magnetic resonance imaging (MRI) at 2 large teaching hospitals in Alberta, seeking to characterise the proportion of lesions within the peripheral zone which were assigned a higher score within the PI-RADS v2.1 protocol based on DCE sequences. ¹⁰ A total of 2742 MRI examinations were included in the analysis with 87 (3.2%) patients upgraded to PI-RADS 4 on the basis of the DCE sequence, accounting for 7.4% of all PI-RADS 4 lesions detected. ¹⁰ Sixty-five of the 85 patients upgraded by DCE underwent a biopsy of the lesion within a year. However, only 18 had clinically significant prostate cancer at this site. ¹⁰ It follows that in the total study cohort, DCE only influenced patient management in 0.66% of patients. ¹⁰ Therefore, a mean of 152.3 administrations of a gadolinium-based contrast agent was required for the detection of only one clinically significant prostate cancer at DCE MRI. ¹⁰ Not only is the administration of gadolinium costly, it is liable to supply-chain failure and increases the overall scan time 3 fold. These findings suggest exclusion of DCE imaging in this setting may be justified to improve cost effectiveness with little to no impact on patient care. ¹⁰ They also remind us to continually question the value of every sequence in a study protocol.

Similarly, it is not only our responsibility to ensure the aspects of a radiological study are useful, but also to ensure we are reporting on all useful information gleamed from the study, including opportunistic screening. For example, reporting of breast arterial calcifications (BAC) on mammography is not routine despite their links to coronary artery calcification (CAC) and cardiovascular disease. McKee et al investigated the association between BAC detected at mammography and the presence of CAC on CT as well as the impact of BAC/CAC notification on primary care provider follow-up. ¹¹ Two hundred and eighty-six participants without known cardiovascular disease were prospectively enrolled and underwent a coronary calcium score CT within 6 months of mammography. BAC was noted in 13% of participants and CAC in 37%. ¹¹ BAC had improved accuracy for detecting moderate or severe CAC with a negative predictive value of 89% for CAC >100 and 98% for CAC >400. ¹¹ The majority (92%) of patients with BAC and CAC had follow-up appointments with their primary care provider and 42% implemented lifestyle changes. ¹¹ While the long term impact of this intervention is unknown and further cost-effectiveness analysis is warranted to determine if screening for CAC on CT following identification of BAC on mammography is a worthwhile endeavour, it reminds us to look beyond answering the specific clinical question to maximise the amount of information obtained from each radiological study.

While maximising system and equipment efficiency, we must also consider what to do in the face of failure. Although functional and efficient equipment is central to modern healthcare, function may be sacrificed for efficiency in the modern day when limiting costs and reducing wait times becomes of more importance than machine servicing or upgrades. However, the risk of equipment malfunction or breakdown further jeopardizes timely patient care and must be considered in the design of a resilient healthcare system. The extreme example of multi-system angiography unit failure as described by O’Leary et al provides a framework on how to build a healthcare sector that although reliant on technology, is resilient to its limitations. ¹² The first stage, anticipation, is to expect the expected. All good things (healthcare equipment) come to an end and while maximum life expectancy cannot be determined by age and case volume alone, radiology departments must engage in close communication with physicians, engineers and departmental leadership to pre-emptively replace ageing equipment prior to failure. ¹² However, this is not always possible and when system failures occur despite our best anticipatory efforts, enter step 2, coping. First you must accept you have a problem. Response to critical events may be delayed where organisations are initially in denial of the issue. However, once acknowledged, the departmental efforts can focus on solutions. In the setting of angiography unit failure, some solutions include the use of mobile C-arms, using alternate imaging modalities such as ultrasound or CT and adjusting staff scheduling to maximise the use of available resources. ¹² Adaptation is the final step to this framework and is directed toward planning to hopefully avert future crisis. ¹² Its components include contingency, renewal and financial planning. We remind our readers of this article to reinforce the concept that adaptation should not require a crisis nor does a crisis have to occur in your own unit or organisation in the hope they will start to consider how this framework could be applied to their own hospitals to limit system failures throughout Canada in the future.