AI in Paediatric Radiology

Abstract

Use of AI has become popular in radiology for improving service efficiency and quality. Currently, there are 366 United States (US) Food and Drug Administration (FDA)-approved radiology AI products for potential use in routine clinical practice. Apparently, some of these products would be useful for its subspecialty, paediatric radiology to address the long-standing problem of paediatric radiologist shortage as well as other issues such as radiation risk, and sedation and anaesthesia use. However, great differences exist between adult and paediatric radiology because of age-dependant changes of anatomy and physiology leading to variations of image acquisition settings and diagnostic processes. Hence, AI products suitable for paediatric radiology have been developed. The purpose of this keynote presentation is to explore current opportunities, challenges and way forward for clinical use of AI in paediatric radiology.

As yet, 13 commercial AI products have been approved by FDA for paediatric radiology use, i.e. 3.6% of the total number of FDA-approved radiology AI products. Around a quarter of these (3 out of 13) are developed for computer aided detection / diagnosis of fractures and dental diseases such as caries and calculi based on general and dental x-ray images. Although FDA has determined the other 10 products as medical image management and processing systems (MIMPS), about half (6) can assist paediatric radiologists in making diagnoses faster through automatic segmentation and quantification of body structures such as urinary bladder, hip, cardiac chamber, brain, liver, and jaw bone in pelvis and hip ultrasound, cardiac, brain and liver magnetic resonance imaging (MRI), and dental x-ray, respectively. Hence, the majority of these products can address the paediatric radiologist shortage issue to some extent by increasing clinician efficiency and productivity while the others (except one designed for lumbar spine surgical outcome prediction) would be useful to alleviate the traditional burdens of paediatric radiology such as x-ray and computed tomography radiation dose, and sedation and anaesthesia use in MRI due to long examination time.

Nonetheless, the application areas of these existing FDA-approved products appear limited and only cover certain types of paediatric examinations and diseases. According to recent literature reviews, there are lots of potential AI applications for improving paediatric radiology service efficiency and quality such as image translation, quality assessment, acquisition setting selection and labelling for routing, billing and hanging protocol management, robotic assistance in guidewire insertion, computer assisted triage and notification, natural language processing for highlighting significant findings in reports, literature search / expert seeking for diagnosis support, trainee education and chatbot for responding to patient questions but none of them have become commercially available. Also, more than one-third (5) of the FDA-approved paediatric radiology products do not have any disclosures of their performance evaluation details, e.g. number of clinical cases and readers involved, patient demographics, etc., and only two have been involved in the American College of Radiology Transparent-AI program with their model training and evaluation arrangements disclosed in more detail. For those products which have their model evaluation information, it is concerning that some of their evaluation approaches seem not appropriate or robust. For example, the aforementioned bladder segmentation and quantification product has only been evaluated based on 122 ultrasound cases of adult patients (age range: 18-90 years) despite a lack of generalisability being a well-known problem of AI.

Although mixed results are noted in previous studies about applying adult radiology AI products to children, clinical centres are encouraged to evaluate feasibility of using adult radiology AI products in their own paediatric radiology practice when such AI applications unavailable for this subspeciality. This is because even for products intended for paediatric radiology use, the clinical centres should conduct their own evaluation to ensure these products being safe and effective for their patients to avoid any potential medico-legal issues. Given that the paediatric radiology AI is an underdeveloped area which has potentials to address its long-standing issues including the clinician shortage, radiation, and sedation and anaesthesia burdens, more financial incentives and research grants from governments and other funding bodies should be provided to manufacturers, researchers and clinicians for the paediatric radiology AI product development, evaluation and adoption. Otherwise, potential advantages of AI cannot be widely realised in paediatric clinical practice, subsequently, affecting this vulnerable patient group and resulting in health equity issue.