In this issue, we will be summarising several new meta-reviews on the topic of artificial intelligence (AI) methods in paediatric radiology. These articles, selected based on their clinical significance and the citation indexes of the journals, cover several important topics. This blog will also be the first of many insights on the future of AI in radiology from the team here at Osimis. The topics covered here include developments in paediatric chest x-ray evaluations, paediatric fracture assessment, paediatric brain tumour imaging and radiation dose optimisation.
Highlights covered in this issue include the ongoing development of AI algorithms, particularly around radiation load reduction with special attention to severe cases (such as brain tumours) and acute conditions (fractures, pneumonia, etc). Evidence suggests that modern AI algorithms can achieve radiation load reductions by up to 70 per cent. There are also the challenges faced when imaging children instead of adults and the unique challenges they bring (such as the thymus mimicking a mediastinal tumour). Please read on to learn more about these key areas of research regarding AI in paediatric imaging.
A recent literature review by S. Padash et al identified all paediatric radiology datasets and determined their potential use (and limitations) for paediatric AI studies. The review revealed a variety of insights, including the fact that the classification of chest radiographs as pneumonia is the most common application of AI, amounting to 65% of the studies evaluated.
As you might expect, adult chest radiographs often differ from those taken of children. For instance, the thymus (a small gland in the lymphatic system) can mimic the appearance of a parenchymal lung infection or a mediastinal mass. For AI to effectively identify one from the other, it requires better data with age-appropriate datasets. The authors identified that a lack of quality open-source data is clearly an obstacle for this technology right now.
The datasets used were collected prospectively and retrospectively, which also threw up some challenges. For instance, some datasets used labelling images that leveraged natural language processing (NLP). It’s common knowledge that images based on NLP from radiology reports are intrinsically associated with uncertainty since radiologists may only be able to identify a disease as a possibility rather an as a definitive diagnosis.
Another example is the diagnosis of pneumonia; normally this would be quite clinical with key radiographic evidence proving the validity of that diagnosis. However, many of the images from these large-scale datasets are only labelled as ‘pneumonia’ and lack the key specific signs needed to further differentiate them.
Then there is the issue of image quality and integration with clinical data. Studies have revealed that AI models have the potential to show false positives when working with insufficient clinical information, improper positioning, irregular expose, clothing, etc. These problems are particularly problematic with child patients since it can be difficult to provide optimal conditions for the capture of high-quality images.
Thankfully, the authors of these studies do present a potential roadmap to address said challenges. Some of these solutions include:
- correctly orienting images in PACS
- prioritising abnormal images for interpretation (especially in the case of life-threatening conditions such as pneumothorax)
- localising abnormalities by lung lobe, automatically assessing diffuse lung abnormalities (such as cystic fibrosis)
- providing preliminary interpretations when referring to physicians in resource-limited conditions.
The research found here provided an excellent review that analysed all available radiological datasets in paediatrics. The authors found most studies of AI related to chest radiographs in paediatrics were limited by challenging conditions and that progress has often been hampered by a lack of large-scale external paediatric radiological datasets.
Most research articles in this area focussed mainly on the use of AI for fracture detection in adults, despite the far more serious clinical consequences of missed fractures in children. In this review, Shelmerdine et al evaluated the available literature on the diagnostic efficacy of AI tools for paediatric fracture assessment imaging.
The authors identified that the additive skeleton accounted for most paediatric fractures, as well as noting that most studies compared the efficacy of a specialised AI algorithm with a commercially available version. Unsurprisingly, the commercially available algorithm was deemed inferior to the specialised versions as it detects fractures throughout the appendicular skeleton (the bones of the upper and lower extremities) rather than allowing for effective imaging of a specific body part.
So, it is clearly important for us to better understand the epidemiology of fractures and the age range of patients across different population groups, and whether AI algorithms that have increased diagnostic accuracies for fracture detection. Additionally, the authors note the limitations of the samples used and that they did not have enough data from patients with fractures suffered from physical abuse or cases with healing fractures, the presence of casts and indwelling orthopaedic hardware.
Overall, this was an excellent review of the data available for the application of AI in paediatric fracture assessment. The authors have identified areas where further progress is needed, such as:
- the development of AI solutions for imaging modalities other than radiography
- the development of AI solutions for the effective assessment of certain risk groups
While advances in the molecular characterisation of brain tumours have improved the overall effectiveness of therapy methods, the prognosis for many tumour cases remains poor. There has been a recent increase in interest for the use of AI tools for imaging brain tumours in children, but there remain some significant barriers for integration into clinical workflows.
Huang at al performed the first systematic review of the application of AI in paediatric brain tumour diagnosis. The studies covered in this review covered the use of AI in diagnosis, tumour segmentation, outcome prediction, O-(2-18F-fluoroethyl)-L-tyrosine positron emission tomography attenuation correction, and creation of synthetic CTs for radiotherapy dose calculation.
The issue here is that none of these applications have been used in a routine clinical setting, so it warrants further investigation as whether these techniques could work as part of a clinical workflow. Despite some promising results, it’s clear the limitations of these studies must be addressed before such an application can be realistically considered. All diagnostic studies covered here focussed on specific types of posterior cranial fossa tumours, but to consider the effective use of AI it would need to be able to identify the presence of other tumour types (including unknown variants).
The first-ever review devoted to the application of AI tools in the diagnosis of paediatric brain tumours reveals that existing software is currently in an experimental stage. To proceed, their validity and efficacy will need to be scrutinised before deployment into clinical practice.
Children are far more susceptible to the effects of radiation exposure, so dose optimisation is vital in the field of paediatric radiology. Curtise et all focussed on this topic for their report to analyse the effectiveness of AI when used to mitigate radiation exposure.
According to the report, most studies suggest that AI can reduce a radiation dose by 36-70% without a significant loss in diagnostic quality. The problem comes in the sheer variation in radiation dose reduction rates. In part, this is down to the retrospective nature of the studies – after all, it’s difficult to control scanning parameters to accurately achieve a low-dose study when you’re interpreting data after the fact. Phantom studies also provide a set of problems since their results cannot be directly translated into clinical practice.
A pioneering review of AI in paediatric radiation exposure reduction studies, this report covers the impact of algorithms on radiography, CT and PET/MRI and the practicality of protecting children through optimised dosage. However, the authors rightly state that the studies available on this topic are too few and that they do not cover all modalities or the implementation of phantom study data.
List of publications
1. Padash, S., Mohebbian, M.R., Adams, S.J. et al. Paediatric chest radiograph interpretation: how far has artificial intelligence come? A systematic literature review. Pediatr Radiol 52, 1568–1580 (2022). https://doi.org/10.1007/s00247-022-05368-w
2. Shelmerdine, S.C., White, R.D., Liu, H. et al. Artificial intelligence for radiological paediatric fracture assessment: a systematic review. Insights Imaging 13, 94 (2022). https://doi.org/10.1186/s13244-022-01234-3
3. Huang J, Shlobin NA, Lam SK, DeCuypere M. Artificial Intelligence Applications in Paediatric Brain Tumour Imaging: A Systematic Review. World Neurosurg. 2022 Jan;157:99-105. doi: 10.1016/j.wneu.2021.10.068. Epub 2021 Oct 11. PMID: 34648981.
4. Curtise K. C. 2022. "Artificial Intelligence for Radiation Dose Optimization in Paediatric Radiology: A Systematic Review" Children 9, no. 7: 1044. https://doi.org/10.3390/children9071044